Authors: Marshall Thomas, Edward Fish, Richard Bowden
Abstract: Despite progress in gloss-free Sign Language Translation (SLT), monolithic end-to-end models consistently fail on two critical components of natural signing: the precise recognition of high-speed fingerspelling and the integration of asynchronous non-manual cues from the face. Recent progress in Automated Sign Language Translation with Large Language Models has side stepped this challenge, forcing a single network to learn these simultaneously resulting in poor performance when tasked with translating crucial information such as names,places, and technical terms. We introduce MultiStream-LLM, a modular framework designed to overcome these limitations. Our approach employs separate, specialized predictors for continuous signing, fingerspelling, and lipreading. Each expert network first decodes its specific modality into a sequence of tokens. These parallel streams are then fused by a lightweight transformer that resolves temporal misalignments before passing the combined representation to a Large Language Model (LLM) for final sentence generation. Our method establishes a new state-of-the-art on the How2Sign benchmark with a BLEU-4 score of 23.5 and achieves 73.2% letter accuracy on the challenging ChicagoFSWildPlus fingerspelling dataset. These results validate our core hypothesis: by isolating and solving distinct recogni tion tasks before fusion, our multi-expert approach provides a more powerful and effective pathway to robust, high-fidelity sign language translation.
Authors: Teo Susnjak
Abstract: Large language models (LLMs) offer significant potential to accelerate systematic literature reviews (SLRs), yet current approaches often rely on brittle, manually crafted prompts that compromise reliability and reproducibility. This fragility undermines scientific confidence in LLM-assisted evidence synthesis. In response, this work adapts recent advances in declarative prompt optimisation, developed for general-purpose LLM applications, and demonstrates their applicability to the domain of SLR automation. This research proposes a structured, domain-specific framework that embeds task declarations, test suites, and automated prompt tuning into a reproducible SLR workflow. These emerging methods are translated into a concrete blueprint with working code examples, enabling researchers to construct verifiable LLM pipelines that align with established principles of transparency and rigour in evidence synthesis. This is a novel application of such approaches to SLR pipelines.
Authors: Jian Wu, Sarah Rajtmajer
Abstract: Over the past decades, alongside advancements in natural language processing, significant attention has been paid to training models to automatically extract, understand, test, and generate hypotheses in open and scientific domains. However, interpretations of the term \emph{hypothesis} for various natural language understanding (NLU) tasks have migrated from traditional definitions in the natural, social, and formal sciences. Even within NLU, we observe differences defining hypotheses across literature. In this paper, we overview and delineate various definitions of hypothesis. Especially, we discern the nuances of definitions across recently published NLU tasks. We highlight the importance of well-structured and well-defined hypotheses, particularly as we move toward a machine-interpretable scholarly record.
Authors: Sheldon Yu, Yuxin Xiong, Junda Wu, Xintong Li, Tong Yu, Xiang Chen, Ritwik Sinha, Jingbo Shang, Julian McAuley
Abstract: Recent advances in chain-of-thought (CoT) prompting have enabled large language models (LLMs) to perform multi-step reasoning. However, the explainability of such reasoning remains limited, with prior work primarily focusing on local token-level attribution, such that the high-level semantic roles of reasoning steps and their transitions remain underexplored. In this paper, we introduce a state-aware transition framework that abstracts CoT trajectories into structured latent dynamics. Specifically, to capture the evolving semantics of CoT reasoning, each reasoning step is represented via spectral analysis of token-level embeddings and clustered into semantically coherent latent states. To characterize the global structure of reasoning, we model their progression as a Markov chain, yielding a structured and interpretable view of the reasoning process. This abstraction supports a range of analyses, including semantic role identification, temporal pattern visualization, and consistency evaluation.
Authors: Seiji Maekawa, Hayate Iso, Nikita Bhutani
Abstract: Effective decision-making often relies on identifying what makes each candidate distinctive. While existing benchmarks for LLMs emphasize retrieving or summarizing information relevant to a given query, they do not evaluate a model's ability to identify globally distinctive features across a set of documents. We introduce Distinctive Feature Mining (DFM), a new task that challenges models to analyze a small-to-medium collection (10-40 documents) and surface features that are rare in the global context (e.g., appearing in less than 10% of documents). This setting mirrors real-world scenarios such as candidate selection or product differentiation, where statistical reasoning, not retrieval, is key. To enable systematic evaluation of this capability, we present DiFBench, a configurable benchmark creation framework with controllable parameters such as document set size and distinctiveness thresholds. Using DiFBench, we perform a large-scale assessment of distinctive feature mining across ten state-of-the-art LLMs. Our findings reveal a significant performance gap between general-purpose and reasoning-enhanced models. All models, however, substantially degrade as the task complexity and document count increase. We also find that a common failure mode is misidentifying frequent features as distinctive. These insights reveal core limitations in contemporary LLMs' abilities to perform fine-grained, statistical reasoning and rarity detection.
Authors: Zachary K. Stine, James E. Deitrick
Abstract: The proliferation of methods for modeling of human meaning-making constitutes a powerful class of instruments for the analysis of complex semiotic systems. However, the field lacks a general theoretical framework for describing these modeling practices across various model types in an apples-to-apples way. In this paper, we propose such a framework grounded in the semiotic theory of C. S. Peirce. We argue that such models measure latent symbol geometries, which can be understood as hypotheses about the complex of semiotic agencies underlying a symbolic dataset. Further, we argue that in contexts where a model's value cannot be straightforwardly captured by proxy measures of performance, models can instead be understood relationally, so that the particular interpretive lens of a model becomes visible through its contrast with other models. This forms the basis of a theory of model semantics in which models, and the modeling decisions that constitute them, are themselves treated as signs. In addition to proposing the framework, we illustrate its empirical use with a few brief examples and consider foundational questions and future directions enabled by the framework.
Authors: Hugo Sousa, Ricardo Campos, Al\'ipio Jorge
Abstract: In this paper we demo the Temporal Game, a novel approach to temporal relation extraction that casts the task as an interactive game. Instead of directly annotating interval-level relations, our approach decomposes them into point-wise comparisons between the start and end points of temporal entities. At each step, players classify a single point relation, and the system applies temporal closure to infer additional relations and enforce consistency. This point-based strategy naturally supports both interval and instant entities, enabling more fine-grained and flexible annotation than any previous approach. The Temporal Game also lays the groundwork for training reinforcement learning agents, by treating temporal annotation as a sequential decision-making task. To showcase this potential, the demo presented in this paper includes a Game mode, in which users annotate texts from the TempEval-3 dataset and receive feedback based on a scoring system, and an Annotation mode, that allows custom documents to be annotated and resulting timeline to be exported. Therefore, this demo serves both as a research tool and an annotation interface. The demo is publicly available at https://temporal-game.inesctec.pt, and the source code is open-sourced to foster further research and community-driven development in temporal reasoning and annotation.
Authors: Yuxiang Liu, Tian Wang, Gourab Kundu, Tianyu Cao, Guang Cheng, Zhen Ge, Jianshu Chen, Qingjun Cui, Trishul Chilimbi
Abstract: Transformer-based models such as BERT and E5 have significantly advanced text embedding by capturing rich contextual representations. However, many complex real-world queries require sophisticated reasoning to retrieve relevant documents beyond surface-level lexical matching, where encoder-only retrievers often fall short. Decoder-only large language models (LLMs), known for their strong reasoning capabilities, offer a promising alternative. Despite this potential, existing LLM-based embedding methods primarily focus on contextual representation and do not fully exploit the reasoning strength of LLMs. To bridge this gap, we propose Reasoning-Infused Text Embedding (RITE), a simple but effective approach that integrates logical reasoning into the text embedding process using generative LLMs. RITE builds upon existing language model embedding techniques by generating intermediate reasoning texts in the token space before computing embeddings, thereby enriching representations with inferential depth. Experimental results on BRIGHT, a reasoning-intensive retrieval benchmark, demonstrate that RITE significantly enhances zero-shot retrieval performance across diverse domains, underscoring the effectiveness of incorporating reasoning into the embedding process.
Authors: Mir Tafseer Nayeem, Davood Rafiei
Abstract: We study the problem of opinion highlights generation from large volumes of user reviews, often exceeding thousands per entity, where existing methods either fail to scale or produce generic, one-size-fits-all summaries that overlook personalized needs. To tackle this, we introduce OpinioRAG, a scalable, training-free framework that combines RAG-based evidence retrieval with LLMs to efficiently produce tailored summaries. Additionally, we propose novel reference-free verification metrics designed for sentiment-rich domains, where accurately capturing opinions and sentiment alignment is essential. These metrics offer a fine-grained, context-sensitive assessment of factual consistency. To facilitate evaluation, we contribute the first large-scale dataset of long-form user reviews, comprising entities with over a thousand reviews each, paired with unbiased expert summaries and manually annotated queries. Through extensive experiments, we identify key challenges, provide actionable insights into improving systems, pave the way for future research, and position OpinioRAG as a robust framework for generating accurate, relevant, and structured summaries at scale.
Authors: Taihei Sone
Abstract: The emergence of generative Artificial Intelligence (AI) has created new opportunities for economic text analysis. This study proposes a Wage Sentiment Index (WSI) constructed with Large Language Models (LLMs) to forecast wage dynamics in Japan. The analysis is based on the Economy Watchers Survey (EWS), a monthly survey conducted by the Cabinet Office of Japan that captures real-time economic assessments from workers in industries highly sensitive to business conditions. The WSI extends the framework of the Price Sentiment Index (PSI) used in prior studies, adapting it specifically to wage related sentiment. To ensure scalability and adaptability, a data architecture is also developed that enables integration of additional sources such as newspapers and social media. Experimental results demonstrate that WSI models based on LLMs significantly outperform both baseline approaches and pretrained models. These findings highlight the potential of LLM-driven sentiment indices to enhance the timeliness and effectiveness of economic policy design by governments and central banks.
Authors: Chen Zheng, Yiyuan Ma, Yuan Yang, Deyi Liu, Jing Liu, Zuquan Song, Yuxin Song, Cheng Ren, Hang Zhu, Xin Liu, Yiyuan Ma, Siyuan Qiao, Xun Zhou, Liang Xiang, Yonghui Wu
Abstract: The development of alignment and reasoning capabilities in large language models has seen remarkable progress through two paradigms: instruction tuning and reinforcement learning from human feedback (RLHF) alignment paradigm, and distillation-based reasoning fine-tuning paradigm. While both approaches prove effective independently, the third paradigm of applying RLHF to distillation-trained models presents significant challenges. Our investigation reveals two critical phenomena that emerge in this paradigm: Sequence Length Collapse, where language generation dramatically reduces during early RLHF training, and the Reward Hockey Stick Curve, featuring severe reward score drops followed by gradual recovery. These instabilities fundamentally compromise the model's alignment and reasoning capabilities. To address these challenges, we propose Balanced Actor Initialization (BAI), a two-stage weighted model merging approach. BAI first merges instruction-following and distillation-based reasoning fine-tuned models, then further combines this intermediate model with the pretrained model to preserve foundational knowledge. Through comprehensive experiments across diverse benchmarks and detailed analysis of training experiments, we demonstrate that BAI resolves Sequence Length Collapse, mitigates the Reward Hockey Stick Curve, and enables continuous sequence length improvement during training. Additionally, our analysis reveals that balanced merging ratios achieve optimal trade-offs between training stability and reasoning capability preservation. Our work provides the effective solution for stable training in this third paradigm, enabling more capable reasoning models that combine distillation efficiency with RLHF alignment.
Authors: Rinku Dewri
Abstract: We introduce GIER (Gap-driven Iterative Enhancement of Responses), a general framework for improving large language model (LLM) outputs through self-reflection and revision based on conceptual quality criteria. Unlike prompting strategies that rely on demonstrations, examples, or chain-of-thought templates, GIER utilizes natural language descriptions of reasoning gaps, and prompts a model to iteratively critique and refine its own outputs to better satisfy these criteria. Across three reasoning-intensive tasks (SciFact, PrivacyQA, and e-SNLI) and four LLMs (GPT-4.1, GPT-4o Mini, Gemini 1.5 Pro, and Llama 3.3 70B), GIER improves rationale quality, grounding, and reasoning alignment without degrading task accuracy. Our analysis demonstrates that models can not only interpret abstract conceptual gaps but also translate them into concrete reasoning improvements.
Authors: Ziyi Xia, Kun Luo, Hongjin Qian, Zheng Liu
Abstract: Large language models (LLMs) are increasingly expected to go beyond simple factual queries toward Deep Research-tasks that require decomposing questions into sub-problems, coordinating multi-step reasoning, and synthesizing evidence from diverse sources. We formalize Deep Research tasks with verifiable answers as Hierarchical Constraint Satisfaction Problems (HCSPs), which are fundamentally different from single-constraint, multi-hop, or flat CSP formulations. However, existing benchmarks (e.g., Natural Questions, HotpotQA) fail to capture this complexity, while recent synthetic datasets often introduce shortcut reasoning, knowledge leakage, or lack sufficient structural depth. To address this gap, we introduce InfoSeek, a scalable framework for synthesizing complex Deep Research tasks. InfoSeek uses a dual-agent system to recursively build a Research Tree from large-scale webpages, blurring intermediate nodes into valid sub-problems, and converting these trees into natural language questions that require traversing the full hierarchy. It also enables rapid scaling, yielding over 50K training examples, a curated test set, and reasoning trajectories generated via reject sampling. Experiments show that models trained on InfoSeek consistently outperform strong baselines. On a challenging benchmark BrowseComp-Plus, 3B LLMs optimized with InfoSeek surpass much larger 32B models and lightweight commercial APIs (e.g., Gemini2.5-Flash), while achieving performance comparable to stronger APIs (e.g., Gemini2.5-Pro). By preserving meta-information such as intermediate steps and retrieval labels, InfoSeek further supports advanced optimization strategies, including compound reward design and trajectory-level exploration. We provide our codes and datasets in \href{https://github.com/VectorSpaceLab/InfoSeek}{this repository}.
Authors: Xuelin Li, Xiangqi Jin, Linfeng Zhang
Abstract: Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes will be released on Github.
Authors: Yuting Tan, Xuying Li, Zhuo Li, Huizhen Shu, Peikang Hu
Abstract: Gradient-based adversarial prompting, such as the Greedy Coordinate Gradient (GCG) algorithm, has emerged as a powerful method for jailbreaking large language models (LLMs). In this paper, we present a systematic appraisal of GCG and its annealing-augmented variant, T-GCG, across open-source LLMs of varying scales. Using Qwen2.5-0.5B, LLaMA-3.2-1B, and GPT-OSS-20B, we evaluate attack effectiveness on both safety-oriented prompts (AdvBench) and reasoning-intensive coding prompts. Our study reveals three key findings: (1) attack success rates (ASR) decrease with model size, reflecting the increasing complexity and non-convexity of larger models' loss landscapes; (2) prefix-based heuristics substantially overestimate attack effectiveness compared to GPT-4o semantic judgments, which provide a stricter and more realistic evaluation; and (3) coding-related prompts are significantly more vulnerable than adversarial safety prompts, suggesting that reasoning itself can be exploited as an attack vector. In addition, preliminary results with T-GCG show that simulated annealing can diversify adversarial search and achieve competitive ASR under prefix evaluation, though its benefits under semantic judgment remain limited. Together, these findings highlight the scalability limits of GCG, expose overlooked vulnerabilities in reasoning tasks, and motivate further development of annealing-inspired strategies for more robust adversarial evaluation.
Authors: Juraj Vladika, Florian Matthes
Abstract: In the digital age, people often turn to the Internet in search of medical advice and recommendations. With the increasing volume of online content, it has become difficult to distinguish reliable sources from misleading information. Similarly, millions of medical studies are published every year, making it challenging for researchers to keep track of the latest scientific findings. These evolving studies can reach differing conclusions, which is not reflected in traditional search tools. To address these challenges, we introduce MedSEBA, an interactive AI-powered system for synthesizing evidence-based answers to medical questions. It utilizes the power of Large Language Models to generate coherent and expressive answers, but grounds them in trustworthy medical studies dynamically retrieved from the research database PubMed. The answers consist of key points and arguments, which can be traced back to respective studies. Notably, the platform also provides an overview of the extent to which the most relevant studies support or refute the given medical claim, and a visualization of how the research consensus evolved through time. Our user study revealed that medical experts and lay users find the system usable and helpful, and the provided answers trustworthy and informative. This makes the system well-suited for both everyday health questions and advanced research insights.
Authors: Fenghua Liu, Yulong Chen, Yixuan Liu, Zhujun Jin, Solomon Tsai, Ming Zhong
Abstract: Large Language Models (LLMs) achieve gold-medal performance across many benchmarks, yet it remains unclear whether such success reflects genuine reasoning or pattern matching. From a cognitive science perspective, an informative test is whether models can master an unfamiliar language through explicit metalinguistic deductive learning, a paradigm where human learners can reliably internalise grammatical systems through metalinguistic reasoning. We address this question with Camlang, a novel constructed language that exhibits naturalistic yet unattested feature combinations. Camlang consists of two explicit resources, a grammar book and a bilingual dictionary, which mirror adult second-language learning via explicit grammar rules and lexical lookup, and enable us to disentangle errors in morpho-syntax, lexical semantics, and sentence-level reasoning. Human experiments show that these resources are sufficient for participants to acquire Camlang and successfully solve Camlang tasks. To operationalise evaluation, we adapt CommonsenseQA into Camlang, creating Camlang-CSQA-v0, the first task in a broader suite where solving questions requires applying grammar rules and lexical mappings. Experimental results show that GPT-5 achieves 98\% EM accuracy in English but only 47\% in Camlang, far below human performance at 87\%, while other state-of-the-art reasoning LLMs perform even worse. Human verification further reveals that most model successes stem from shallow lexical alignment while GPT-5 shows emerging metalinguistic awareness to a limited extent but not systematic grammatical mastery as humans. Camlang establishes a cognitively grounded evaluation paradigm that exposes fundamental gaps between current models and human metalinguistic competence.
Authors: Xuecheng Zou, Ke Liu, Bingbing Wang, Huafei Deng, Li Zhang, Yu Tang
Abstract: Building upon the standard graph-based Retrieval-Augmented Generation (RAG), the introduction of heterogeneous graphs and hypergraphs aims to enrich retrieval and generation by leveraging the relationships between multiple entities through the concept of semantic units (SUs). But this also raises a key issue: The extraction of high-level SUs limited to local text chunks is prone to ambiguity, complex coupling, and increased retrieval overhead due to the lack of global knowledge or the neglect of fine-grained relationships. To address these issues, we propose GOSU, a semantic unit-centric RAG framework that efficiently performs global disambiguation and utilizes SUs to capture interconnections between different nodes across the global context. In the graph construction phase, GOSU performs global merging on the pre-extracted SUs from local text chunks and guides entity and relationship extraction, reducing the difficulty of coreference resolution while uncovering global semantic objects across text chunks. In the retrieval and generation phase, we introduce hierarchical keyword extraction and semantic unit completion. The former uncovers the fine-grained binary relationships overlooked by the latter, while the latter compensates for the coarse-grained n-ary relationships missing from the former. Evaluation across multiple tasks demonstrates that GOSU outperforms the baseline RAG methods in terms of generation quality.
Authors: Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Hichem Telli, Cosimo Distante, Abdenour Hadid
Abstract: Islamic inheritance law (Ilm al-Mawarith) requires precise identification of heirs and calculation of shares, which poses a challenge for AI. In this paper, we present a lightweight framework for solving multiple-choice inheritance questions using a specialised Arabic text encoder and Attentive Relevance Scoring (ARS). The system ranks answer options according to semantic relevance, and enables fast, on-device inference without generative reasoning. We evaluate Arabic encoders (MARBERT, ArabicBERT, AraBERT) and compare them with API-based LLMs (Gemini, DeepSeek) on the QIAS 2025 dataset. While large models achieve an accuracy of up to 87.6%, they require more resources and are context-dependent. Our MARBERT-based approach achieves 69.87% accuracy, presenting a compelling case for efficiency, on-device deployability, and privacy. While this is lower than the 87.6% achieved by the best-performing LLM, our work quantifies a critical trade-off between the peak performance of large models and the practical advantages of smaller, specialized systems in high-stakes domains.
Authors: Beining Xu
Abstract: Uncertainty quantification (UQ) for open-ended language generation remains a critical yet underexplored challenge, especially under black-box constraints where internal model signals are inaccessible. In this paper, we introduce Token-Entropy Conformal Prediction (TECP), a novel framework that leverages token-level entropy as a logit-free, reference-free uncertainty measure and integrates it into a split conformal prediction (CP) pipeline to construct prediction sets with formal coverage guarantees. Unlike existing approaches that rely on semantic consistency heuristics or white-box features, TECP directly estimates epistemic uncertainty from the token entropy structure of sampled generations and calibrates uncertainty thresholds via CP quantiles to ensure provable error control. Empirical evaluations across six large language models and two benchmarks (CoQA and TriviaQA) demonstrate that TECP consistently achieves reliable coverage and compact prediction sets, outperforming prior self-consistency-based UQ methods. Our method provides a principled and efficient solution for trustworthy generation in black-box LLM settings.
Authors: Saksorn Ruangtanusak, Pittawat Taveekitworachai, Kunat Pipatanakul
Abstract: This report investigates approaches for prompting a tool-augmented large language model (LLM) to act as a role-playing dialogue agent in the API track of the Commonsense Persona-grounded Dialogue Challenge (CPDC) 2025. In this setting, dialogue agents often produce overly long in-character responses (over-speaking) while failing to use tools effectively according to the persona (under-acting), such as generating function calls that do not exist or making unnecessary tool calls before answering. We explore four prompting approaches to address these issues: 1) basic role prompting, 2) human-crafted role prompting, 3) automatic prompt optimization (APO), and 4) rule-based role prompting. The rule-based role prompting (RRP) approach achieved the best performance through two novel techniques--character-card/scene-contract design and strict enforcement of function calling--which led to an overall score of 0.571, improving on the zero-shot baseline score of 0.519. These findings demonstrate that RRP design can substantially improve the effectiveness and reliability of role-playing dialogue agents compared with more elaborate methods such as APO. To support future efforts in developing persona prompts, we are open-sourcing all of our best-performing prompts and the APO tool. Source code is available at https://github.com/scb-10x/apo.
Authors: Li S. Yifei, Allen Chang, Chaitanya Malaviya, Mark Yatskar
Abstract: Evaluating long-form responses to research queries heavily relies on expert annotators, restricting attention to areas like AI where researchers can conveniently enlist colleagues. Yet, research expertise is widespread: survey articles synthesize knowledge distributed across the literature. We introduce ResearchQA, a resource for evaluating LLM systems by distilling survey articles from 75 research fields into 21K queries and 160K rubric items. Each rubric, derived jointly with queries from survey sections, lists query-specific answer evaluation criteria, i.e., citing papers, making explanations, and describing limitations. Assessments by 31 Ph.D. annotators in 8 fields indicate 96% of queries support Ph.D. information needs and 87% of rubric items should be addressed in system responses by a sentence or more. Using our rubrics, we are able to construct an automatic pairwise judge obtaining 74% agreement with expert judgments. We leverage ResearchQA to analyze competency gaps in 18 systems in over 7.6K pairwise evaluations. No parametric or retrieval-augmented system we evaluate exceeds 70% on covering rubric items, and the highest-ranking agentic system shows 75% coverage. Error analysis reveals that the highest-ranking system fully addresses less than 11% of citation rubric items, 48% of limitation items, and 49% of comparison items. We release our data to facilitate more comprehensive multi-field evaluations.
Authors: Jialong Zuo, Guangyan Zhang, Minghui Fang, Shengpeng Ji, Xiaoqi Jiao, Jingyu Li, Yiwen Guo, Zhou Zhao
Abstract: Discrete speech representation learning has recently attracted increasing interest in both acoustic and semantic modeling. Existing approaches typically encode 16 kHz waveforms into discrete tokens at a rate of 25 or 50 tokens per second. However, given that speech generally conveys only 2 to 5 words per second, such fine-grained tokenization introduces redundancy and hinders efficiency in downstream training and inference. Moreover, semantic speech representations at this frequency primarily capture phonetic-level information, while semantic understanding may not require such detailed token-level resolution. To address these limitations, we propose an entropy-based dynamic aggregation framework for learning compressed semantic speech representations. A speech language model is first pre-trained via next-token prediction on large-scale unlabeled data to capture frequent token patterns. Predictive entropy is then used to adaptively determine aggregation boundaries, followed by a cross-attention module that fuses information within each segment. By adjusting the entropy threshold, the granularity and compression ratio of the representations can be flexibly controlled. Experiments on ASR, speech-to-text translation, and voice conversion tasks demonstrate that the compressed representations perform on par with or better than dense token sequences, demonstrating the effectiveness of the proposed approach.
Authors: Eunjung Cho, Alexander Hoyle, Yoan Hermstr\"uwer
Abstract: Large Language Models (LLMs) are increasingly used to generate user-tailored summaries, adapting outputs to specific stakeholders. In legal contexts, this raises important questions about motivated reasoning -- how models strategically frame information to align with a stakeholder's position within the legal system. Building on theories of legal realism and recent trends in legal practice, we investigate how LLMs respond to prompts conditioned on different legal roles (e.g., judges, prosecutors, attorneys) when summarizing judicial decisions. We introduce an evaluation framework grounded in legal fact and reasoning inclusion, also considering favorability towards stakeholders. Our results show that even when prompts include balancing instructions, models exhibit selective inclusion patterns that reflect role-consistent perspectives. These findings raise broader concerns about how similar alignment may emerge as LLMs begin to infer user roles from prior interactions or context, even without explicit role instructions. Our results underscore the need for role-aware evaluation of LLM summarization behavior in high-stakes legal settings.
Authors: Hanqi Yan, Hainiu Xu, Yulan He
Abstract: With Large Language Models (LLMs) becoming increasingly widely adopted, concerns regarding their safety and alignment with human values have intensified. Previous studies have shown that fine-tuning LLMs on narrow and malicious datasets induce misaligned behaviors. In this work, we report a more concerning phenomenon, Reasoning-Induced Misalignment. Specifically, we observe that LLMs become more responsive to malicious requests when reasoning is strengthened, via switching to "think-mode" or fine-tuning on benign math datasets, with dense models particularly vulnerable. Moreover, we analyze internal model states and find that both attention shifts and specialized experts in mixture-of-experts models help redirect excessive reasoning towards safety guardrails. These findings provide new insights into the emerging reasoning-safety trade-off and underscore the urgency of advancing alignment for advanced reasoning models.
Authors: Lang Xiong, Nishant Bhargava, Wesley Chang, Jianhang Hong, Haihao Liu, Kevin Zhu
Abstract: Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting, a phenomenon known as "evaluation awareness." This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty. In this work, we systematically quantify these behavioral changes by manipulating the perceived context of prompts. We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like" and leverage an LLM rewriting strategy to shift these prompts towards a more natural, deployment-style context while preserving the original task. Using this method, we achieved a 30% increase in the average probe score across a strategic role-playing dataset after rewriting. Evaluating a suite of state-of-the-art models on these original and rewritten prompts, we find that rewritten "deploy-like" prompts induce a significant and consistent shift in behavior. Across all models, we observed an average increase in honest responses of 5.26% and a corresponding average decrease in deceptive responses of 12.40%. Furthermore, refusal rates increased by an average of 6.38%, indicating heightened safety compliance. Our findings demonstrate that evaluation awareness is a quantifiable and manipulable factor that directly influences LLM behavior, revealing that models are more prone to unsafe or deceptive outputs in perceived test environments. This underscores the urgent need for more realistic evaluation frameworks to accurately gauge true model alignment before deployment.
Authors: Rishiraj Acharya
Abstract: The Transformer architecture, underpinned by the self-attention mechanism, has become the de facto standard for sequence modeling tasks. However, its core computational primitive scales quadratically with sequence length (O(N^2)), creating a significant bottleneck for processing long contexts. In this paper, we propose the Gated Associative Memory (GAM) network, a novel, fully parallel architecture for sequence modeling that exhibits linear complexity (O(N)) with respect to sequence length. The GAM block replaces the self-attention layer with two parallel pathways: a causal convolution to efficiently capture local, position-dependent context, and a parallel associative memory retrieval mechanism to model global, content-based patterns. These pathways are dynamically fused using a gating mechanism, allowing the model to flexibly combine local and global information for each token. We implement GAM from scratch and conduct a rigorous comparative analysis against a standard Transformer model and a modern linear-time baseline (Mamba) on the WikiText-2 benchmark, as well as against the Transformer on the TinyStories dataset. Our experiments demonstrate that GAM is consistently faster, outperforming both baselines on training speed, and achieves a superior or competitive final validation perplexity across all datasets, establishing it as a promising and efficient alternative for sequence modeling.
Authors: Ali Zain, Sareem Farooqui, Muhammad Rafi
Abstract: This paper presents presents three distinct systems developed for the M-DAIGT shared task on detecting AI generated content in news articles and academic abstracts. The systems includes: (1) A fine-tuned RoBERTa-base classifier, (2) A classical TF-IDF + Support Vector Machine (SVM) classifier , and (3) An Innovative ensemble model named Candace, leveraging probabilistic features extracted from multiple Llama-3.2 models processed by a customTransformer encoder.The RoBERTa-based system emerged as the most performant, achieving near-perfect results on both development and test sets.
Authors: Md Tanzib Hosain, Md Kishor Morol
Abstract: Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample, high-quality unit, and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1\% pass@1 solve rate. With our best inference technique, which combines multi-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2\%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open-source our code and data at https://github.com/kraritt/zolve.
Authors: Sanjeeevan Selvaganapathy, Mehwish Nasim
Abstract: We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining whether models with minimal safety alignment (uncensored) might provide more objective classification capabilities compared to their heavily-aligned (censored) counterparts. While uncensored models theoretically offer a less constrained perspective free from moral guardrails that could bias classification decisions, our results reveal a surprising trade-off: censored models significantly outperform their uncensored counterparts in both accuracy and robustness, achieving 78.7% versus 64.1% strict accuracy. However, this enhanced performance comes with its own limitation -- the safety alignment acts as a strong ideological anchor, making censored models resistant to persona-based influence, while uncensored models prove highly malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency.
Authors: Junfeng Ran, Guangxiang Zhao, Yuhan Wu, Dawei Zhu, Longyun Wu, Yikai Zhao, Tong Yang, Lin Sun, Xiangzheng Zhang, Sujian Li
Abstract: The Mixture-of-Experts (MoE) models have gained significant attention in deep learning due to their dynamic resource allocation and superior performance across diverse tasks. However, efficiently training these models remains challenging. The MoE upcycling technique has been proposed to reuse and improve existing model components, thereby minimizing training overhead. Despite this, simple routers, such as linear routers, often struggle with complex routing tasks within MoE upcycling. In response, we propose a novel routing technique called Router Upcycling to enhance the performance of MoE upcycling models. Our approach initializes multiple routers from the attention heads of preceding attention layers during upcycling. These routers collaboratively assign tokens to specialized experts in an attention-like manner. Each token is processed into diverse queries and aligned with the experts' features (serving as keys). Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance, outperforming other upcycling baselines.
Authors: Austin McCutcheon, Chris Brogly
Abstract: Small language models (SLMs) have the capability for text generation and may potentially be used to generate falsified texts online. This study evaluates 14 SLMs (1.7B-14B parameters) including LLaMA, Gemma, Phi, SmolLM, Mistral, and Granite families in generating perceived low and high quality fake news headlines when explicitly prompted, and whether they appear to be similar to real-world news headlines. Using controlled prompt engineering, 24,000 headlines were generated across low-quality and high-quality deceptive categories. Existing machine learning and deep learning-based news headline quality detectors were then applied against these SLM-generated fake news headlines. SLMs demonstrated high compliance rates with minimal ethical resistance, though there were some occasional exceptions. Headline quality detection using established DistilBERT and bagging classifier models showed that quality misclassification was common, with detection accuracies only ranging from 35.2% to 63.5%. These findings suggest the following: tested SLMs generally are compliant in generating falsified headlines, although there are slight variations in ethical restraints, and the generated headlines did not closely resemble existing primarily human-written content on the web, given the low quality classification accuracy.
Authors: Chen Su, Yuanhe Tian, Yan Song, Yongdong Zhang
Abstract: Recent studies in time series forecasting (TSF) use multimodal inputs, such as text and historical time series data, to predict future values. These studies mainly focus on developing advanced techniques to integrate textual information with time series data to perform the task and achieve promising results. Meanwhile, these approaches rely on high-quality text and time series inputs, whereas in some cases, the text does not accurately or fully capture the information carried by the historical time series, which leads to unstable performance in multimodal TSF. Therefore, it is necessary to enhance the textual content to improve the performance of multimodal TSF. In this paper, we propose improving multimodal TSF by reinforcing the text modalities. We propose a text reinforcement model (TeR) to generate reinforced text that addresses potential weaknesses in the original text, then apply this reinforced text to support the multimodal TSF model's understanding of the time series, improving TSF performance. To guide the TeR toward producing higher-quality reinforced text, we design a reinforcement learning approach that assigns rewards based on the impact of each reinforced text on the performance of the multimodal TSF model and its relevance to the TSF task. We optimize the TeR accordingly, so as to improve the quality of the generated reinforced text and enhance TSF performance. Extensive experiments on a real-world benchmark dataset covering various domains demonstrate the effectiveness of our approach, which outperforms strong baselines and existing studies on the dataset.
Authors: Alex Gulko, Yusen Peng, Sachin Kumar
Abstract: Probing with sparse autoencoders is a promising approach for uncovering interpretable features in large language models (LLMs). However, the lack of automated evaluation methods has hindered their broader adoption and development. In this work, we introduce CE-Bench, a novel and lightweight contrastive evaluation benchmark for sparse autoencoders, built on a curated dataset of contrastive story pairs. We conduct comprehensive ablation studies to validate the effectiveness of our approach. Our results show that CE-Bench reliably measures the interpretability of sparse autoencoders and aligns well with existing benchmarks, all without requiring an external LLM. The official implementation and evaluation dataset are open-sourced under the MIT License.
Authors: Kaiwen Wei, Jinpeng Gao, Jiang Zhong, Yuming Yang, Fengmao Lv, Zhenyang Li
Abstract: Large language models (LLMs) have shown strong potential in recommendation tasks due to their strengths in language understanding, reasoning and knowledge integration. These capabilities are especially beneficial for review-based recommendation, which relies on semantically rich user-generated texts to reveal fine-grained user preferences and item attributes. However, effectively incorporating reviews into LLM-based recommendation remains challenging due to (1) inefficient to dynamically utilize user reviews under LLMs' constrained context windows, and (2) lacking effective mechanisms to prioritize reviews most relevant to the user's current decision context. To address these challenges, we propose RevBrowse, a review-driven recommendation framework inspired by the "browse-then-decide" decision process commonly observed in online user behavior. RevBrowse integrates user reviews into the LLM-based reranking process to enhance its ability to distinguish between candidate items. To improve the relevance and efficiency of review usage, we introduce PrefRAG, a retrieval-augmented module that disentangles user and item representations into structured forms and adaptively retrieves preference-relevant content conditioned on the target item. Extensive experiments on four Amazon review datasets demonstrate that RevBrowse achieves consistent and significant improvements over strong baselines, highlighting its generalizability and effectiveness in modeling dynamic user preferences. Furthermore, since the retrieval-augmented process is transparent, RevBrowse offers a certain level of interpretability by making visible which reviews influence the final recommendation.
Authors: Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park, Junha Hyung, Jaegul Choo
Abstract: Masked diffusion models (MDMs) offer a promising non-autoregressive alternative for large language modeling. Standard decoding methods for MDMs, such as confidence-based sampling, select tokens independently based on individual token confidences at each diffusion step. However, we observe that this independent token selection often results in generation orders resembling sequential autoregressive processes, limiting the advantages of non-autoregressive modeling. To mitigate this pheonomenon, we propose Reward-Weighted Sampling (RWS), a novel decoding strategy that leverages an external reward model to provide a principled global signal during the iterative diffusion process. Specifically, at each diffusion step, RWS evaluates the quality of the entire intermediate sequence and scales token logits accordingly, guiding token selection by integrating global sequence-level coherence. This method selectively increases the confidence of tokens that initially have lower scores, thereby promoting a more non-autoregressive generation order. Furthermore, we provide theoretical justification showing that reward-weighted logit scaling induces beneficial rank reversals in token selection and consistently improves expected reward. Experiments demonstrate that RWS significantly promotes non-autoregressive generation orders, leading to improvements across multiple evaluation metrics. These results highlight the effectiveness of integrating global signals in enhancing both the non-autoregressive properties and overall performance of MDMs.
Authors: Elias Ra, Seung Je Kim, Eui-Yeong Seo, Geunju So
Abstract: Higher education faces growing challenges in delivering personalized, scalable, and pedagogically coherent learning experiences. This study introduces a structured framework for designing an AI-powered Learning Management System (AI-LMS) that integrates generative and conversational AI to support adaptive, interactive, and learner-centered instruction. Using a design-based research (DBR) methodology, the framework unfolds through five phases: literature review, SWOT analysis, development of ethical-pedagogical principles, system design, and instructional strategy formulation. The resulting AI-LMS features modular components -- including configurable prompts, adaptive feedback loops, and multi-agent conversation flows -- aligned with pedagogical paradigms such as behaviorist, constructivist, and connectivist learning theories. By combining AI capabilities with human-centered design and ethical safeguards, this study advances a practical model for AI integration in education. Future research will validate and refine the system through real-world implementation.
Authors: Houji Jin, Negin Ashrafi, Armin Abdollahi, Wei Liu, Jian Wang, Ganyu Gui, Maryam Pishgar, Huanghao Feng
Abstract: The rapid growth of large language models (LLMs) has heightened the demand for accurate detection of AI-generated text, particularly in languages like Chinese, where subtle linguistic nuances pose significant challenges to current methods. In this study, we conduct a systematic comparison of encoder-based Transformers (Chinese BERT-large and RoBERTa-wwm-ext-large), a decoder-only LLM (Alibaba's Qwen2.5-7B/DeepSeek-R1-Distill-Qwen-7B fine-tuned via Low-Rank Adaptation, LoRA), and a FastText baseline using the publicly available dataset from the NLPCC 2025 Chinese AI-Generated Text Detection Task. Encoder models were fine-tuned using a novel prompt-based masked language modeling approach, while Qwen2.5-7B was adapted for classification with an instruction-format input and a lightweight classification head trained via LoRA. Experiments reveal that although encoder models nearly memorize training data, they suffer significant performance degradation under distribution shifts (RoBERTa: 76.3% test accuracy; BERT: 79.3%). FastText demonstrates surprising lexical robustness (83.5% accuracy) yet lacks deeper semantic understanding. In contrast, the LoRA-adapted Qwen2.5-7B achieves 95.94% test accuracy with balanced precision-recall metrics, indicating superior generalization and resilience to dataset-specific artifacts. These findings underscore the efficacy of decoder-based LLMs with parameter-efficient fine-tuning for robust Chinese AI-generated text detection. Future work will explore next-generation Qwen3 models, distilled variants, and ensemble strategies to enhance cross-domain robustness further.
Authors: Zhichao Yan, Jiaoyan Chen, Jiapu Wang, Xiaoli Li, Ru Li, Jeff Z. Pan
Abstract: Attribution is crucial in question answering (QA) with Large Language Models (LLMs).SOTA question decomposition-based approaches use long form answers to generate questions for retrieving related documents. However, the generated questions are often irrelevant and incomplete, resulting in a loss of facts in retrieval.These approaches also fail to aggregate evidence snippets from different documents and paragraphs. To tackle these problems, we propose a new fact decomposition-based framework called FIDES (\textit{faithful context enhanced fact decomposition and evidence aggregation}) for attributed QA. FIDES uses a contextually enhanced two-stage faithful decomposition method to decompose long form answers into sub-facts, which are then used by a retriever to retrieve related evidence snippets. If the retrieved evidence snippets conflict with the related sub-facts, such sub-facts will be revised accordingly. Finally, the evidence snippets are aggregated according to the original sentences.Extensive evaluation has been conducted with six datasets, with an additionally proposed new metric called $Attr_{auto-P}$ for evaluating the evidence precision. FIDES outperforms the SOTA methods by over 14\% in average with GPT-3.5-turbo, Gemini and Llama 70B series.
Authors: Weizhe Shi, Qiqi Wang, Yihong Pan, Qian Liu, Kaiqi Zhao
Abstract: A criminal judicial opinion represents the judge's disposition of a case, including the decision rationale and sentencing. Automatically generating such opinions can assist in analyzing sentencing consistency and provide judges with references to similar past cases. However, current research typically approaches this task by dividing it into two isolated subtasks: legal reasoning and sentencing prediction. This separation often leads to inconsistency between the reasoning and predictions, failing to meet real-world judicial requirements. Furthermore, prior studies rely on manually curated knowledge to enhance applicability, yet such methods remain limited in practical deployment. To address these limitations and better align with legal practice, we propose a new LegalAI task: Judicial Opinion Generation, which simultaneously produces both legal reasoning and sentencing decisions. To achieve this, we introduce LegalChainReasoner, a framework that applies structured legal chains to guide the model through comprehensive case assessments. By integrating factual premises, composite legal conditions, and sentencing conclusions, our approach ensures flexible knowledge injection and end-to-end opinion generation. Experiments on two real-world and open-source Chinese legal case datasets demonstrate that our method outperforms baseline models.
Authors: Reem Abdel-Salam, Mary Adewunmi, Modinat A. Abayomi
Abstract: Large language models (LLMs) are increasingly evident for accurate question answering across various domains. However, rigorous evaluation of their performance on complex question-answering (QA) capabilities is essential before deployment in real-world biomedical and healthcare applications. This paper presents our approach to the MedHopQA track of the BioCreative IX shared task, which focuses on multi-hop biomedical question answering involving diseases, genes, and chemicals. We adopt a supervised fine-tuning strategy leveraging LLaMA 3 8B, enhanced with a curated biomedical question-answer dataset compiled from external sources including BioASQ, MedQuAD, and TREC. Three experimental setups are explored: fine-tuning on combined short and long answers, short answers only, and long answers only. While our models demonstrate strong domain understanding, achieving concept-level accuracy scores of up to 0.8, their Exact Match (EM) scores remain significantly lower, particularly in the test phase. We introduce a two-stage inference pipeline for precise short-answer extraction to mitigate verbosity and improve alignment with evaluation metrics. Despite partial improvements, challenges persist in generating strictly formatted outputs. Our findings highlight the gap between semantic understanding and exact answer evaluation in biomedical LLM applications, motivating further research in output control and post-processing strategies.
Authors: Felix Engl, Andreas Henrich
Abstract: The training of topic models for a multilingual environment is a challenging task, requiring the use of sophisticated algorithms, topic-aligned corpora, and manual evaluation. These difficulties are further exacerbated when the developer lacks knowledge of the target language or is working in an environment with limited data, where only small or unusable multilingual corpora are available. Considering these challenges, we introduce Topic Model Translation (TMT), a novel, robust and transparent technique designed to transfer topic models (e.g., Latent Dirichlet Allocation (LDA) based topic models) from one language to another, without the need for metadata, embeddings, or aligned corpora. TMT enables the reuse of topic models across languages, making it especially suitable for scenarios where large corpora in the target language are unavailable or manual translation is infeasible. Furthermore, we evaluate TMT extensively using both quantitative and qualitative methods, demonstrating that it produces semantically coherent and consistent topic translations.
Authors: Michelle Elizabeth, Alicja Kasicka, Natalia Krawczyk, Magalie Ochs, Gw\'enol\'e Lecorv\'e, Justyna Gromada, Lina M. Rojas-Barahona
Abstract: The growing number of generative AI-based dialogue systems has made their evaluation a crucial challenge. This paper presents our contribution to this important problem through the Dialogue System Technology Challenge (DSTC-12, Track 1), where we developed models to predict dialogue-level, dimension-specific scores. Given the constraint of using relatively small models (i.e. fewer than 13 billion parameters) our work follows two main strategies: employing Language Models (LMs) as evaluators through prompting, and training encoder-based classification and regression models. Our results show that while LM prompting achieves only modest correlations with human judgments, it still ranks second on the test set, outperformed only by the baseline. The regression and classification models, with significantly fewer parameters, demonstrate high correlation for some dimensions on the validation set. Although their performance decreases on the test set, it is important to note that the test set contains annotations with significantly different score ranges for some of the dimensions with respect to the train and validation sets.
Authors: Tengyu Pan, Zhichao Duan, Zhenyu Li, Bowen Dong, Ning Liu, Xiuxing Li, Jianyong Wang
Abstract: Text embedding models are essential for various natural language processing tasks, enabling the effective encoding of semantic information into dense vector representations. These models are typically optimized using triplets of (query, positive, negative) data pairs for contrastive learning, where the negative samples play a critical role in enhancing the model's ability to discern subtle semantic distinctions. In this work, we introduce a Multi-Granularity Hard-negative (MGH) synthesis framework that leverages large language models (LLMs) to generate diverse negative samples with varying levels of similarity with the query. This approach facilitates a coarse-to-fine curriculum learning strategy during supervised training, allowing the embedding model to progressively learn more nuanced semantic representations. Meanwhile, we propose an Anchor Token Aware (ATA) pooling method that assigns higher weights to anchor tokens based on aggregation patterns observed in LLMs, improving text embedding accuracy without increasing model complexity. Comprehensive experiments on the MTEB benchmark demonstrate that our methods achieve state-of-the-art performance, surpassing existing synthesis strategies both with synthetic data and when combined with public retrieval datasets.
Authors: Shaina Raza, Maximus Powers, Partha Pratim Saha, Mahveen Raza, Rizwan Qureshi
Abstract: Text-to-Image (TTI) models are powerful creative tools but risk amplifying harmful social biases. We frame representational societal bias assessment as an image curation and evaluation task and introduce a pilot benchmark of occupational portrayals spanning five socially salient roles (CEO, Nurse, Software Engineer, Teacher, Athlete). Using five state-of-the-art models: closed-source (DALLE 3, Gemini Imagen 4.0) and open-source (FLUX.1-dev, Stable Diffusion XL Turbo, Grok-2 Image), we compare neutral baseline prompts against fairness-aware controlled prompts designed to encourage demographic diversity. All outputs are annotated for gender (male, female) and race (Asian, Black, White), enabling structured distributional analysis. Results show that prompting can substantially shift demographic representations, but with highly model-specific effects: some systems diversify effectively, others overcorrect into unrealistic uniformity, and some show little responsiveness. These findings highlight both the promise and the limitations of prompting as a fairness intervention, underscoring the need for complementary model-level strategies. We release all code and data for transparency and reproducibility https://github.com/maximus-powers/img-gen-bias-analysis.
URLs: https://github.com/maximus-powers/img-gen-bias-analysis.
Authors: Zixuan Shangguan, Yanjie Dong, Lanjun Wang, Xiaoyi Fan, Victor C. M. Leung, Xiping Hu
Abstract: Large language models (LLMs) have demonstrated exceptional proficiency in language understanding. However, when LLMs align their outputs with deceptive and/or misleading prompts, the generated responses could deviate from the de facto information. Such observations are known as fawning hallucinations, where the model prioritizes alignment with the input's implied perspective over accuracy and truthfulness. In this work, we analyze fawning hallucinations in various natural language processing tasks and tailor the so-termed contrastive decoding method for fawning-hallucination mitigation. Specifically, we design two paradigms to generate corresponding deceptive and/or misleading inputs for the consistent fawning hallucinations induction. Then, we propose the collaborative contrastive decoding (CCD) to handle the fawning hallucinations across different tasks in LLMs. By contrasting the deviation in output distribution between induced and transformed neutral inputs, the proposed CCD can reduce reliance on deceptive and/or misleading information without requiring additional training. Extensive experiments demonstrate that the proposed CCD can effectively mitigate fawning hallucinations and improve the factuality of the generated responses over various tasks.
Authors: Yuqin Dai, Guoqing Wang, Yuan Wang, Kairan Dou, Kaichen Zhou, Zhanwei Zhang, Shuo Yang, Fei Tang, Jun Yin, Pengyu Zeng, Zhenzhe Ying, Can Yi, Changhua Meng, Yuchen Zhou, Yongliang Shen, Shuai Lu
Abstract: Large Language Models (LLMs) empowered with retrieval mechanisms have achieved strong progress in open-domain question answering (QA). Yet, the conventional retrieve--then--answer paradigm often suffers from two key limitations: (1) low signal-to-noise ratio in retrieved evidence, where useful information is buried under irrelevant content, and (2) error accumulation in multi-hop reasoning when incomplete or noisy passages are involved. To address these challenges, we present EviNote-RAG, an agentic RAG framework that introduces a structured retrieve--note--answer pipeline. Instead of directly reasoning over raw retrievals, the model is trained to compose Supportive-Evidence Notes (SENs), concise, human-like notes that preserve only answer-relevant information, highlight uncertainty, and explicitly state when no useful evidence exists. This distillation process is further reinforced by the Evidence Quality Reward (EQR), an entailment-based signal that evaluates whether SENs logically support the final answer. Together, SENs and EQR guide the model toward faithful and robust reasoning, while reducing the impact of noise. Experiments on in-domain and out-of-domain QA benchmarks show that EviNote-RAG consistently outperforms strong baselines in accuracy, generalization, and training stability. In particular, it achieves state-of-the-art results while enhancing robustness and efficiency, yielding relative F1 gains of 20\% on HotpotQA (+0.093), 40\% on Bamboogle (+0.151), and 91\% on 2Wiki (+0.256) via denser rewards and reduced verbosity.
Authors: R\u{a}zvan-Alexandru Sm\u{a}du, Andreea Iuga, Dumitru-Clementin Cercel, Florin Pop
Abstract: Satire, irony, and sarcasm are techniques typically used to express humor and critique, rather than deceive; however, they can occasionally be mistaken for factual reporting, akin to fake news. These techniques can be applied at a more granular level, allowing satirical information to be incorporated into news articles. In this paper, we introduce the first sentence-level dataset for Romanian satire detection for news articles, called SeLeRoSa. The dataset comprises 13,873 manually annotated sentences spanning various domains, including social issues, IT, science, and movies. With the rise and recent progress of large language models (LLMs) in the natural language processing literature, LLMs have demonstrated enhanced capabilities to tackle various tasks in zero-shot settings. We evaluate multiple baseline models based on LLMs in both zero-shot and fine-tuning settings, as well as baseline transformer-based models. Our findings reveal the current limitations of these models in the sentence-level satire detection task, paving the way for new research directions.
Authors: David Duki\'c, Goran Glava\v{s}, Jan \v{S}najder
Abstract: Sequence labeling (SL) tasks, where labels are assigned to tokens, are abundant in NLP (e.g., named entity recognition and aspect-based sentiment analysis). Owing to the intuition that they require bidirectional context, SL tasks are commonly tackled with encoder-only models. Recent work also shows that removing the causal mask in fine-tuning enables decoder-based LLMs to become effective token classifiers. Less work, however, focused on (supervised) generative SL, a more natural setting for causal LLMs. Due to their rapid scaling, causal LLMs applied to SL are expected to outperform encoders, whose own development has stagnated. In this work, we propose supervised in-context fine-tuning (SIFT) for generative SL. SIFT casts SL tasks as constrained response generation, natural to LLMs, combining (1) in-context learning (ICL) from demonstrations with (2) supervised fine-tuning. SIFT considerably outperforms both ICL and decoder-as-encoder fine-tuning baselines on a range of standard SL tasks. We further find that although long context hinders the performance of generative SL in both ICL and SIFT, this deficiency can be mitigated by removing the instruction, as instructions are shown to be largely unnecessary for achieving strong SL performance with SIFT. Our findings highlight strengths and limitations of SL with LLMs, underscoring the importance of a response-based generative task formulation for effective SL performance.
Authors: Md Shahidul Salim, Lian Fu, Arav Adikesh Ramakrishnan, Zonghai Yao, Hong Yu
Abstract: We present MedCOD (Medical Chain-of-Dictionary), a hybrid framework designed to improve English-to-Spanish medical translation by integrating domain-specific structured knowledge into large language models (LLMs). MedCOD integrates domain-specific knowledge from both the Unified Medical Language System (UMLS) and the LLM-as-Knowledge-Base (LLM-KB) paradigm to enhance structured prompting and fine-tuning. We constructed a parallel corpus of 2,999 English-Spanish MedlinePlus articles and a 100-sentence test set annotated with structured medical contexts. Four open-source LLMs (Phi-4, Qwen2.5-14B, Qwen2.5-7B, and LLaMA-3.1-8B) were evaluated using structured prompts that incorporated multilingual variants, medical synonyms, and UMLS-derived definitions, combined with LoRA-based fine-tuning. Experimental results demonstrate that MedCOD significantly improves translation quality across all models. For example, Phi-4 with MedCOD and fine-tuning achieved BLEU 44.23, chrF++ 28.91, and COMET 0.863, surpassing strong baseline models like GPT-4o and GPT-4o-mini. Ablation studies confirm that both MedCOD prompting and model adaptation independently contribute to performance gains, with their combination yielding the highest improvements. These findings highlight the potential of structured knowledge integration to enhance LLMs for medical translation tasks.
Authors: Yihong Chen
Abstract: The making of knowledge engines in natural language processing has been shaped by two seemingly distinct paradigms: one grounded in structure, the other driven by massively available unstructured data. The structured paradigm leverages predefined symbolic interactions, such as knowledge graphs, as priors and designs models to capture them. In contrast, the unstructured paradigm centers on scaling transformer architectures with increasingly vast data and model sizes, as seen in modern large language models. Despite their divergence, this thesis seeks to establish conceptual connections bridging these paradigms. Two complementary forces, structure and destructure, emerge across both paradigms: structure organizes seen symbolic interactions, while destructure, through periodic embedding resets, improves model plasticity and generalization to unseen scenarios. These connections form a new recipe for developing general knowledge engines that can support transparent, controllable, and adaptable intelligent systems.
Authors: Chia-Hsuan Hsu, Jun-En Ding, Hsin-Ling Hsu, Feng Liu, Fang-Ming Hung
Abstract: Medical question answering requires advanced reasoning that integrates domain knowledge with logical inference. However, existing large language models (LLMs) often generate reasoning chains that lack factual accuracy and clinical reliability. We propose Ranked Preference Reinforcement Optimization (RPRO), a novel framework that uniquely combines reinforcement learning with preference-driven reasoning refinement to enhance clinical chain-of-thought (CoT) performance. RPRO differentiates itself from prior approaches by employing task-adaptive reasoning templates and a probabilistic evaluation mechanism that aligns outputs with established clinical workflows, while automatically identifying and correcting low-quality reasoning chains. Unlike traditional pairwise preference methods, RPRO introduces a groupwise ranking optimization based on the Bradley-Terry model and incorporates KL-divergence regularization for stable training. Experiments on PubMedQA and MedQA-USMLE show consistent improvements over strong baselines. Remarkably, our 1.1B parameter model outperforms much larger 7B-13B models, including medical-specialized variants. These findings demonstrate that combining preference optimization with quality-driven refinement offers a scalable and effective approach to building more reliable, clinically grounded medical LLMs.
Authors: Sadia Zaman Mishu, S M Rafiuddin
Abstract: The demand for text classification is growing significantly in web searching, data mining, web ranking, recommendation systems, and so many other fields of information and technology. This paper illustrates the text classification process on different datasets using some standard supervised machine learning techniques. Text documents can be classified through various kinds of classifiers. Labeled text documents are used to classify the text in supervised classifications. This paper applies these classifiers on different kinds of labeled documents and measures the accuracy of the classifiers. An Artificial Neural Network (ANN) model using Back Propagation Network (BPN) is used with several other models to create an independent platform for labeled and supervised text classification process. An existing benchmark approach is used to analyze the performance of classification using labeled documents. Experimental analysis on real data reveals which model works well in terms of classification accuracy.
Authors: S M Rafiuddin
Abstract: Ranking words is an important way to summarize a text or to retrieve information. A word graph is a way to represent the words of a sentence or a text as the vertices of a graph and to show the relationship among the words. It is also useful to determine the relative importance of a word among the words in the word-graph. In this research, the ranking of Bangla words are calculated, representing Bangla words from a text in a word graph using various graph based ranking algorithms. There is a lack of a standard Bangla word database. In this research, the Indian Language POS-tag Corpora is used, which has a rich collection of Bangla words in the form of sentences with their parts of speech tags. For applying a word graph to various graph based ranking algorithms, several standard procedures are applied. The preprocessing steps are done in every word graph and then applied to graph based ranking algorithms to make a comparison among these algorithms. This paper illustrate the entire procedure of calculating the ranking of Bangla words, including the construction of the word graph from text. Experimental result analysis on real data reveals the accuracy of each ranking algorithm in terms of F1 measure.
Authors: Nikta Gohari Sadr, Sahar Heidariasl, Karine Megerdoomian, Laleh Seyyed-Kalantari, Ali Emami
Abstract: Large language models (LLMs) struggle to navigate culturally specific communication norms, limiting their effectiveness in global contexts. We focus on Persian taarof, a social norm in Iranian interactions, which is a sophisticated system of ritual politeness that emphasizes deference, modesty, and indirectness, yet remains absent from existing cultural benchmarks. We introduce TaarofBench, the first benchmark for evaluating LLM understanding of taarof, comprising 450 role-play scenarios covering 12 common social interaction topics, validated by native speakers. Our evaluation of five frontier LLMs reveals substantial gaps in cultural competence, with accuracy rates 40-48% below native speakers when taarof is culturally appropriate. Performance varies between interaction topics, improves with Persian-language prompts, and exhibits gender-based asymmetries. We also show that responses rated "polite" by standard metrics often violate taarof norms, indicating the limitations of Western politeness frameworks. Through supervised fine-tuning and Direct Preference Optimization, we achieve 21.8% and 42.3% improvement in model alignment with cultural expectations. Our human study with 33 participants (11 native Persian, 11 heritage, and 11 non-Iranian speakers) forms baselines in varying degrees of familiarity with Persian norms. This work lays the foundation for developing diverse and culturally aware LLMs, enabling applications that better navigate complex social interactions.
Authors: Xiaoying Song, Anirban Saha Anik, Eduardo Blanco, Vanessa Frias-Martinez, Lingzi Hong
Abstract: In response to the urgent need for effective communication with crisis-affected populations, automated responses driven by language models have been proposed to assist in crisis communications. A critical yet often overlooked factor is the consistency of response style, which could affect the trust of affected individuals in responders. Despite its importance, few studies have explored methods for maintaining stylistic consistency across generated responses. To address this gap, we propose a novel metric for evaluating style consistency and introduce a fusion-based generation approach grounded in this metric. Our method employs a two-stage process: it first assesses the style of candidate responses and then optimizes and integrates them at the instance level through a fusion process. This enables the generation of high-quality responses while significantly reducing stylistic variation between instances. Experimental results across multiple datasets demonstrate that our approach consistently outperforms baselines in both response quality and stylistic uniformity.
Authors: Xiaoying Song, Anirban Saha Anik, Dibakar Barua, Pengcheng Luo, Junhua Ding, Lingzi Hong
Abstract: Health misinformation spreading online poses a significant threat to public health. Researchers have explored methods for automatically generating counterspeech to health misinformation as a mitigation strategy. Existing approaches often produce uniform responses, ignoring that the health literacy level of the audience could affect the accessibility and effectiveness of counterspeech. We propose a Controlled-Literacy framework using retrieval-augmented generation (RAG) with reinforcement learning (RL) to generate tailored counterspeech adapted to different health literacy levels. In particular, we retrieve knowledge aligned with specific health literacy levels, enabling accessible and factual information to support generation. We design a reward function incorporating subjective user preferences and objective readability-based rewards to optimize counterspeech to the target health literacy level. Experiment results show that Controlled-Literacy outperforms baselines by generating more accessible and user-preferred counterspeech. This research contributes to more equitable and impactful public health communication by improving the accessibility and comprehension of counterspeech to health misinformation.
Authors: Abdessalam Bouchekif, Samer Rashwani, Heba Sbahi, Shahd Gaben, Mutez Al-Khatib, Mohammed Ghaly
Abstract: This paper evaluates the knowledge and reasoning capabilities of Large Language Models in Islamic inheritance law, known as 'ilm al-mawarith. We assess the performance of seven LLMs using a benchmark of 1,000 multiple-choice questions covering diverse inheritance scenarios, designed to test models' ability to understand the inheritance context and compute the distribution of shares prescribed by Islamic jurisprudence. The results reveal a significant performance gap: o3 and Gemini 2.5 achieved accuracies above 90%, whereas ALLaM, Fanar, LLaMA, and Mistral scored below 50%. These disparities reflect important differences in reasoning ability and domain adaptation. We conduct a detailed error analysis to identify recurring failure patterns across models, including misunderstandings of inheritance scenarios, incorrect application of legal rules, and insufficient domain knowledge. Our findings highlight limitations in handling structured legal reasoning and suggest directions for improving performance in Islamic legal reasoning. Code: https://github.com/bouchekif/inheritance_evaluation
Authors: Farah Adeeba, Rajesh Bhatt
Abstract: In this paper, we document a paradigm gap in the combinatorial possibilities of verbs and aspect in Urdu: the perfective form of the -ya: kar construction (e.g. ro-ya: ki: cry-Pfv do.Pfv) is sharply ungrammatical in modern Urdu and Hindi, despite being freely attested in 19th century literature. We investigate this diachronic shift through historical text analysis, a large-scale corpus study which confirms the stark absence of perfective forms and subjective evaluation tasks with native speakers, who judge perfective examples as highly unnatural. We argue that this gap arose from a fundamental morphosyntactic conflict: the construction's requirement for a nominative subject and an invariant participle clashes with the core grammatical rule that transitive perfective assign ergative case. This conflict rendered the perfective form unstable, and its functional replacement by other constructions allowed the gap to become entrenched in the modern grammar.
Authors: Jinwen Chen, Hainan Zhang, Liang Pang, Yongxin Tong, Haibo Zhou, Yuan Zhan, Wei Lin, Zhiming Zheng
Abstract: The current RAG system requires uploading plaintext documents to the cloud, risking private data leakage. Parametric RAG (PRAG) addresses this by encoding documents as LoRA within LLMs, enabling reasoning without exposing raw content. However, it still faces two issues: (1) PRAG demands synthesizing QA pairs and fine-tuning LLM for each individual document to create its corresponding LoRA, leading to unacceptable inference latency. (2) The performance of PRAG relies solely on synthetic QA data, lacking internal alignment with standard RAG, resulting in poor generalization on out-of-distribution(OOD) inputs. Therefore, achieving high-efficiency parameterization while maintaining RAG-level performance remains a critical challenge for privacy-preserving reasoning. In this paper, we propose DistilledPRAG, a generalizable knowledge-distilled parametric RAG model aligned with standard RAG in document structure and parameter activation. We first synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. Then, we mask the plaintext documents with a special token and translate them to LoRA via a parameter generator, maintaining the standard RAG document structure. Finally, guided by synthetic QA data, we train the parameter generator to match standard RAG's hidden states and output logits, enabling RAG-style reasoning without original documents. Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on OOD data.
Authors: Xiaoqiang Lin, Aritra Ghosh, Bryan Kian Hsiang Low, Anshumali Shrivastava, Vijai Mohan
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.
Authors: Yulong Wu, Viktor Schlegel, Riza Batista-Navarro
Abstract: How does the natural evolution of context paragraphs affect question answering in generative Large Language Models (LLMs)? To investigate this, we propose a framework for curating naturally evolved, human-edited variants of reading passages from contemporary QA benchmarks and for analyzing LLM performance across a range of semantic similarity scores, which quantify how closely each variant aligns with content seen during pretraining. Using this framework, we evaluate six QA datasets and eight LLMs with publicly available training data. Our experiments reveal that LLM performance declines as reading passages naturally diverge from the versions encountered during pretraining-even when the question and all necessary information remains present at inference time. For instance, average model accuracy on BoolQ drops by over 30% from the highest to lowest similarity bins, with slopes exceeding 70 across several LLMs. These findings suggest that natural text evolution poses a significant challenge to the language understanding capabilities of LLMs.
Authors: Zhihui Xie, Jiacheng Ye, Lin Zheng, Jiahui Gao, Jingwei Dong, Zirui Wu, Xueliang Zhao, Shansan Gong, Xin Jiang, Zhenguo Li, Lingpeng Kong
Abstract: We present Dream-Coder 7B, an open-source discrete diffusion language model for code generation that exhibits emergent any-order generation capabilities. Unlike traditional autoregressive (AR) models that decode strictly left-to-right, Dream-Coder 7B adaptively determines its decoding strategy based on the coding task: sketch-first generation for complex algorithms, left-to-right generation for straightforward completions, and interleaved reasoning generation for code understanding tasks. We adapt a pretrained AR checkpoint to a discrete diffusion frameworks with a continuous-time weighted cross-entropy objective. Our post-training recipe comprises (i) supervised fine-tuning, where we mitigate padding pathologies via random truncation and a padding penalty to improve sample efficiency and stabilize generation; and (ii) reinforcement learning with verifiable rewards over a curated high-quality prompt set drawn from open-source datasets, using a tailored reinforcement learning recipe for diffusion language models. The resulting Dream-Coder 7B Instruct attains 21.4\% pass@1 on LiveCodeBench (2410--2505) and demonstrates competitive performance on HumanEval, MBPP, BigCodeBench, and CRUXEval. We release Dream-Coder-7B and Dream-Coder-7B-Instruct checkpoints, training recipes, preprocessing pipelines, and inference code to facilitate reproducibility and further research.
Authors: Zhihao Zhang, Sophia Yat Mei Lee, Dong Zhang, Shoushan Li, Guodong Zhou
Abstract: Cross-lingual Named Entity Recognition (CL-NER) aims to transfer knowledge from high-resource languages to low-resource languages. However, existing zero-shot CL-NER (ZCL-NER) approaches primarily focus on Latin script language (LSL), where shared linguistic features facilitate effective knowledge transfer. In contrast, for non-Latin script language (NSL), such as Chinese and Japanese, performance often degrades due to deep structural differences. To address these challenges, we propose an entity-aligned translation (EAT) approach. Leveraging large language models (LLMs), EAT employs a dual-translation strategy to align entities between NSL and English. In addition, we fine-tune LLMs using multilingual Wikipedia data to enhance the entity alignment from source to target languages.
Authors: Xuemei Tang, Chengxi Yan, Jinghang Gu, Chu-Ren Huang
Abstract: Chinese information extraction (IE) involves multiple tasks across diverse temporal domains, including Classical and Modern documents. Fine-tuning a single model on heterogeneous tasks and across different eras may lead to interference and reduced performance. Therefore, in this paper, we propose Tea-MOELoRA, a parameter-efficient multi-task framework that combines LoRA with a Mixture-of-Experts (MoE) design. Multiple low-rank LoRA experts specialize in different IE tasks and eras, while a task-era-aware router mechanism dynamically allocates expert contributions. Experiments show that Tea-MOELoRA outperforms both single-task and joint LoRA baselines, demonstrating its ability to leverage task and temporal knowledge effectively.
Authors: Yu Liu, Yanan Cao, Xixun Lin, Yanmin Shang, Shi Wang, Shirui Pan
Abstract: Knowledge graph completion (KGC) aims to infer new knowledge and make predictions from knowledge graphs. Recently, large language models (LLMs) have exhibited remarkable reasoning capabilities. LLM-enhanced KGC methods primarily focus on designing task-specific instructions, achieving promising advancements. However, there are still two critical challenges. First, existing methods often ignore the inconsistent representation spaces between natural language and graph structures. Second, most approaches design separate instructions for different KGC tasks, leading to duplicate works and time-consuming processes. To address these challenges, we propose SAT, a novel framework that enhances LLMs for KGC via structure-aware alignment-tuning. Specifically, we first introduce hierarchical knowledge alignment to align graph embeddings with the natural language space through multi-task contrastive learning. Then, we propose structural instruction tuning to guide LLMs in performing structure-aware reasoning over KGs, using a unified graph instruction combined with a lightweight knowledge adapter. Experimental results on two KGC tasks across four benchmark datasets demonstrate that SAT significantly outperforms state-of-the-art methods, especially in the link prediction task with improvements ranging from 8.7% to 29.8%.
Authors: Seganrasan Subramanian, Abhigya Verma
Abstract: The ability of large language models (LLMs) to process and reason over long textual inputs is critical for a wide range of real-world applications. However, progress in this area is significantly constrained by the absence of high-quality, diverse, and verifiable long-context datasets suitable for both training and evaluation. This work introduces a modular, extensible framework for synthetic long-context data generation via prompt-based interaction with LLMs. The framework supports multiple training and alignment objectives, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO). It encompasses four core generation paradigms: multi-turn conversational dialogues, document-grounded input-output pairs, verifiable instruction-response tasks, and long-context reasoning examples. Through templated prompting, a model-agnostic architecture, and metadata-enriched outputs, the proposed approach facilitates scalable, controllable, and purpose-aligned dataset creation for advancing long-context capabilities in LLMs.
Authors: Luxi He, Nimra Nadeem, Michel Liao, Howard Chen, Danqi Chen, Mariano-Florentino Cu\'ellar, Peter Henderson
Abstract: AI systems are increasingly governed by natural language principles, yet a key challenge arising from reliance on language remains underexplored: interpretive ambiguity. As in legal systems, ambiguity arises both from how these principles are written and how they are applied. But while legal systems use institutional safeguards to manage such ambiguity, such as transparent appellate review policing interpretive constraints, AI alignment pipelines offer no comparable protections. Different interpretations of the same rule can lead to inconsistent or unstable model behavior. Drawing on legal theory, we identify key gaps in current alignment pipelines by examining how legal systems constrain ambiguity at both the rule creation and rule application steps. We then propose a computational framework that mirrors two legal mechanisms: (1) a rule refinement pipeline that minimizes interpretive disagreement by revising ambiguous rules (analogous to agency rulemaking or iterative legislative action), and (2) prompt-based interpretive constraints that reduce inconsistency in rule application (analogous to legal canons that guide judicial discretion). We evaluate our framework on a 5,000-scenario subset of the WildChat dataset and show that both interventions significantly improve judgment consistency across a panel of reasonable interpreters. Our approach offers a first step toward systematically managing interpretive ambiguity, an essential step for building more robust, law-following AI systems.
Authors: Sajjad Kachuee, Mohammad Sharifkhani
Abstract: Using Large Language Models (LLMs) in real-world applications presents significant challenges, particularly in balancing computational efficiency and performance. Optimizing acceleration after the fine-tuning phase and during inference is crucial for building an efficient architecture. This paper introduces Zero-Shot Adjustable Acceleration, a novel training and inference method that dynamically adjusts hardware usage during inference without requiring additional fine-tuning. The proposed approach is applied to newly developed models and evaluated across multiple classification and text generation tasks. Experimental results demonstrate that the method enables a wide range of acceleration in a zero-shot manner and achieves up to a 11x speedup compared to the baseline.
Authors: Chenyang Le, Bing Han, Jinshun Li, Songyong Chen, Yanmin Qian
Abstract: Simultaneous Speech Translation (SimulST) enables real-time cross-lingual communication by jointly optimizing speech recognition and machine translation under strict latency constraints. Existing systems struggle to balance translation quality, latency, and semantic coherence, particularly in multilingual many-to-many scenarios where divergent read and write policies hinder unified strategy learning. In this paper, we present SimulMEGA (Simultaneous Generation by Mixture-of-Experts Gating), an unsupervised policy learning framework that combines prefix-based training with a Mixture-of-Experts refiner to learn effective read and write decisions in an implicit manner, without adding inference-time overhead. Our design requires only minimal modifications to standard transformer architectures and generalizes across both speech-to-text and text-to-speech streaming tasks. Through comprehensive evaluation on six language pairs, our 500M parameter speech-to-text model outperforms the Seamless baseline, achieving under 7 percent BLEU degradation at 1.5 seconds average lag and under 3 percent at 3 seconds. We further demonstrate the versatility of SimulMEGA by extending it to streaming TTS with a unidirectional backbone, yielding superior latency quality tradeoffs.
Authors: Ege S\"ualp, Mina Rezaei
Abstract: Catastrophic forgetting is a significant challenge in continual learning, in which a model loses prior knowledge when it is fine-tuned on new tasks. This problem is particularly critical for large language models (LLMs) undergoing continual learning, as retaining performance across diverse domains is important for their general utility. In this paper, we explore model growth, a promising strategy that leverages smaller models to expedite and structure the training of larger ones for mitigating the catastrophic forgetting problem. Although growth-based pretraining, particularly via transformer stacking, has shown promise in accelerating convergence, its impact on forgetting remains under-explored. Therefore, we evaluate whether growth-based models can retain previously learned capabilities more effectively across a sequence of fine-tuning tasks involving domain knowledge, reasoning, reading comprehension, and bias. Our findings show that both models -- one trained with growth (Stack LLM) and one without (LLM) -- exhibit improvements in domain knowledge. However, reasoning and reading comprehension degrade over time, indicating signs of catastrophic forgetting. Stack LLM consistently shows less degradation, especially in reading comprehension, suggesting enhanced retention capabilities. Interestingly, in bias evaluation, the baseline LLM becomes progressively more neutral with continued fine-tuning, while Stack LLM maintains a steady bias ratio around 60--61\%. These results indicate that growth-based pretraining may deliver modest improvements in resisting catastrophic forgetting, though trade-offs remain in handling social biases.
Authors: Wei Huang, Huang Wei, Yinggui Wang
Abstract: Large language models (LLMs) excel in general tasks but struggle with domain-specific ones, requiring fine-tuning with specific data. With many open-source LLMs available, selecting the best model for fine-tuning downstream tasks is challenging, primarily focusing on how to quickly identify the optimal LLM. We introduce a Data and Model Compression Framework (DaMoC) that addresses this challenge by: 1) Data Level: A systematic categorization of data filtering methodologies for LLMs is first established, classifying them into three distinct paradigms: (1) distribution-aware methods, (2) quality-aware methods, and (3) hybrid approaches considering both dimensions. Further, we enhance the density of key tokens in the text achieving token compression. Subsequently, we use an LLM to iterative rewrite the text to optimize its expression. 2) Model Level: We use layer similarity scores to assess each layer's importance and remove those with lower importance. Then, we introduce a sparse merging paradigm to preserve as much of the original model's capability as possible. Extensive experiments on four datasets, medical Q&A, financial Q&A, general Q&A, and reading comprehension, show that we can select the optimal LLM while saving approximately 20-fold in training time.
Authors: Hao Yang, Zhiyu Yang, Yunjie Zhang, Shanyi Zhu, Lin Yang
Abstract: Chain-of-Thought reasoning has emerged as a pivotal methodology for enhancing model inference capabilities. Despite growing interest in Chain-of-Thought reasoning, its underlying mechanisms remain unclear. This paper explores the working mechanisms of Chain-of-Thought reasoning from the perspective of the dual relationship between in-context learning and pretrained priors. We first conduct a fine-grained lexical-level analysis of rationales to examine the model's reasoning behavior. Then, by incrementally introducing noisy exemplars, we examine how the model balances pretrained priors against erroneous in-context information. Finally, we investigate whether prompt engineering can induce slow thinking in large language models. Our extensive experiments reveal three key findings: (1) The model not only quickly learns the reasoning structure at the lexical level but also grasps deeper logical reasoning patterns, yet it heavily relies on pretrained priors. (2) Providing sufficient exemplars shifts the model's decision-making from pretrained priors to in-context signals, while misleading prompts introduce instability. (3) Long Chain-of-Thought prompting can induce the model to generate longer reasoning chains, thereby improving its performance on downstream tasks.
Authors: Jonas Noblet (LIDILEM)
Abstract: Emotion is a crucial phenomenon in the functioning of human beings in society. However, it remains a widely open subject, particularly in its textual manifestations. This paper examines an industrial corpus manually annotated following an evaluative approach to emotion. This theoretical framework, which is currently underutilized, offers a different perspective that complements traditional approaches. Noting that the annotations we collected exhibit significant disagreement, we hypothesized that they nonetheless follow stable statistical trends. Using language models trained on these annotations, we demonstrate that it is possible to model the labeling process and that variability is driven by underlying linguistic features. Conversely, our results indicate that language models seem capable of distinguishing emotional situations based on evaluative criteria.
Authors: Juhyun Oh, Inha Cha, Michael Saxon, Hyunseung Lim, Shaily Bhatt, Alice Oh
Abstract: The prevailing ``trivia-centered paradigm'' for evaluating the cultural alignment of large language models (LLMs) is increasingly inadequate as these models become more advanced and widely deployed. Existing approaches typically reduce culture to static facts or values, testing models via multiple-choice or short-answer questions that treat culture as isolated trivia. Such methods neglect the pluralistic and interactive realities of culture, and overlook how cultural assumptions permeate even ostensibly ``neutral'' evaluation settings. In this position paper, we argue for \textbf{intentionally cultural evaluation}: an approach that systematically examines the cultural assumptions embedded in all aspects of evaluation, not just in explicitly cultural tasks. We systematically characterize the what, how, and circumstances by which culturally contingent considerations arise in evaluation, and emphasize the importance of researcher positionality for fostering inclusive, culturally aligned NLP research. Finally, we discuss implications and future directions for moving beyond current benchmarking practices, discovering important applications that we don't know exist, and involving communities in evaluation design through HCI-inspired participatory methodologies.
Authors: Sishi Xiong, Ziyang He, Zhongjiang He, Yu Zhao, Changzai Pan, Jie Zhang, Zhenhe Wu, Shuangyong Song, Yongxiang Li
Abstract: While large language models (LLMs) have shown promise in the table question answering (TQA) task through prompt engineering, they face challenges in industrial applications, including structural heterogeneity, difficulties in target data localization, and bottlenecks in complex reasoning. To address these limitations, this paper presents TableZoomer, a novel LLM-powered, programming-based agent framework. It introduces three key innovations: (1) replacing the original fully verbalized table with structured table schema to bridge the semantic gap and reduce computational complexity; (2) a query-aware table zooming mechanism that dynamically generates sub-table schema through column selection and entity linking, significantly improving target localization efficiency; and (3) a Program-of-Thoughts (PoT) strategy that transforms queries into executable code to mitigate numerical hallucination. Additionally, we integrate the reasoning workflow with the ReAct paradigm to enable iterative reasoning. Extensive experiments demonstrate that our framework maintains the usability advantages while substantially enhancing performance and scalability across tables of varying scales. When implemented with the Qwen3-8B-Instruct LLM, TableZoomer achieves accuracy improvements of 19.34% and 25% over conventional PoT methods on the large-scale DataBench dataset and the small-scale Fact Checking task of TableBench dataset, respectively.
Authors: Anum Afzal, Mehul Kumawat, Florian Matthes
Abstract: Large Language Models (LLMs), being generic task solvers, are versatile. However, despite the vast amount of data they are trained on, there are speculations about their adaptation capabilities to a new domain. Additionally, the simple fine-tuning of the model to incorporate knowledge of a new domain is computationally expensive and time-consuming. This becomes more challenging when the domain in question is also low-resource, and labeled data is unavailable. We leverage parameter-efficient fine-tuning techniques (PEFTs) on high-resource datasets to address these challenges to improve performance on unseen low-resource domains. Throughout our experiments, we evaluate whether intrinsic linguistic commonalities between datasets can be leveraged for efficient domain adaptation. We benchmark six PEFTs with \texttt{Llama-3-8B-Instruct} on 14 training datasets from the Scientific, Medical, Legal, and News domains for a Text Summarization task. Our experiments show that for low-resource domains, inference using Within-Domain Adapters can achieve better performance than Few-Shot as well as a much larger \texttt{Llama-3-70B-Instruct}. Lastly, in the absence of Within-Domain Adapters, we explore the concept of using Cross-Domain Adapters as well as the strategic combinations of adapters to leverage intrinsic language similarities across domains, facilitating better adaptability and performance in low-resource settings.
Authors: Meituan LongCat Team, Bayan, Bei Li, Bingye Lei, Bo Wang, Bolin Rong, Chao Wang, Chao Zhang, Chen Gao, Chen Zhang, Cheng Sun, Chengcheng Han, Chenguang Xi, Chi Zhang, Chong Peng, Chuan Qin, Chuyu Zhang, Cong Chen, Congkui Wang, Dan Ma, Daoru Pan, Defei Bu, Dengchang Zhao, Deyang Kong, Dishan Liu, Feiye Huo, Fengcun Li, Fubao Zhang, Gan Dong, Gang Liu, Gang Xu, Ge Li, Guoqiang Tan, Guoyuan Lin, Haihang Jing, Haomin Fu, Haonan Yan, Haoxing Wen, Haozhe Zhao, Hong Liu, Hongmei Shi, Hongyan Hao, Hongyin Tang, Huantian Lv, Hui Su, Jiacheng Li, Jiahao Liu, Jiahuan Li, Jiajun Yang, Jiaming Wang, Jian Yang, Jianchao Tan, Jiaqi Sun, Jiaqi Zhang, Jiawei Fu, Jiawei Yang, Jiaxi Hu, Jiayu Qin, Jingang Wang, Jiyuan He, Jun Kuang, Junhui Mei, Kai Liang, Ke He, Kefeng Zhang, Keheng Wang, Keqing He, Liang Gao, Liang Shi, Lianhui Ma, Lin Qiu, Lingbin Kong, Lingtong Si, Linkun Lyu, Linsen Guo, Liqi Yang, Lizhi Yan, Mai Xia, Man Gao, Manyuan Zhang, Meng Zhou, Mengxia Shen, Mingxiang Tuo, Mingyang Zhu, Peiguang Li, Peng Pei, Peng Zhao, Pengcheng Jia, Pingwei Sun, Qi Gu, Qianyun Li, Qingyuan Li, Qiong Huang, Qiyuan Duan, Ran Meng, Rongxiang Weng, Ruichen Shao, Rumei Li, Shizhe Wu, Shuai Liang, Shuo Wang, Suogui Dang, Tao Fang, Tao Li, Tefeng Chen, Tianhao Bai, Tianhao Zhou, Tingwen Xie, Wei He, Wei Huang, Wei Liu, Wei Shi, Wei Wang, Wei Wu, Weikang Zhao, Wen Zan, Wenjie Shi, Xi Nan, Xi Su, Xiang Li, Xiang Mei, Xiangyang Ji, Xiangyu Xi, Xiangzhou Huang, Xianpeng Li, Xiao Fu, Xiao Liu, Xiao Wei, Xiaodong Cai, Xiaolong Chen, Xiaoqing Liu, Xiaotong Li, Xiaowei Shi, Xiaoyu Li, Xili Wang, Xin Chen, Xing Hu, Xingyu Miao, Xinyan He, Xuemiao Zhang, Xueyuan Hao, Xuezhi Cao, Xunliang Cai, Xurui Yang, Yan Feng, Yang Bai, Yang Chen, Yang Yang, Yaqi Huo, Yerui Sun, Yifan Lu, Yifan Zhang, Yipeng Zang, Yitao Zhai, Yiyang Li, Yongjing Yin, Yongkang Lv, Yongwei Zhou, Yu Yang, Yuchen Xie, Yueqing Sun, Yuewen Zheng, Yuhua Wei, Yulei Qian, Yunfan Liang, Yunfang Tai, Yunke Zhao, Zeyang Yu, Zhao Zhang, Zhaohua Yang, Zhenchao Zhang, Zhikang Xia, Zhiye Zou, Zhizhao Zeng, Zhongda Su, Zhuofan Chen, Zijian Zhang, Ziwen Wang, Zixu Jiang, Zizhe Zhao, Zongyu Wang, Zunhai Su
Abstract: We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat
URLs: https://longcat.ai, https://huggingface.co/meituan-longcat, https://github.com/meituan-longcat
Authors: Jihyung Lee, Daehui Kim, Seonjeong Hwang, Hyounghun Kim, Gary Lee
Abstract: Large Language Models (LLM) have achieved remarkable performances in general domains and are now extending into the expert domain of law. Several benchmarks have been proposed to evaluate LLMs' legal capabilities. However, these benchmarks fail to evaluate open-ended and provision-grounded Question Answering (QA). To address this, we introduce a Korean Benchmark for Legal EXplainable QA (KoBLEX), designed to evaluate provision-grounded, multi-hop legal reasoning. KoBLEX includes 226 scenario-based QA instances and their supporting provisions, created using a hybrid LLM-human expert pipeline. We also propose a method called Parametric provision-guided Selection Retrieval (ParSeR), which uses LLM-generated parametric provisions to guide legally grounded and reliable answers. ParSeR facilitates multi-hop reasoning on complex legal questions by generating parametric provisions and employing a three-stage sequential retrieval process. Furthermore, to better evaluate the legal fidelity of the generated answers, we propose Legal Fidelity Evaluation (LF-Eval). LF-Eval is an automatic metric that jointly considers the question, answer, and supporting provisions and shows a high correlation with human judgments. Experimental results show that ParSeR consistently outperforms strong baselines, achieving the best results across multiple LLMs. Notably, compared to standard retrieval with GPT-4o, ParSeR achieves +37.91 higher F1 and +30.81 higher LF-Eval. Further analyses reveal that ParSeR efficiently delivers consistent performance across reasoning depths, with ablations confirming the effectiveness of ParSeR.
Authors: Wei Wang, Fuqing Bie, Junzhe Chen, Dan Zhang, Shiyu Huang, Evgeny Kharlamov, Jie Tang
Abstract: Complex games have long been an important benchmark for testing the progress of artificial intelligence algorithms. AlphaGo, AlphaZero, and MuZero have defeated top human players in Go and Chess, garnering widespread societal attention towards artificial intelligence. Concurrently, large language models (LLMs) have exhibited remarkable capabilities across various tasks, raising the question of whether LLMs can achieve similar success in complex games. In this paper, we explore the potential of LLMs in mastering complex card games. We systematically assess the learning capabilities of LLMs across eight diverse card games, evaluating the impact of fine-tuning on high-quality gameplay data, and examining the models' ability to retain general capabilities while mastering these games. Our findings indicate that: (1) LLMs can approach the performance of strong game AIs through supervised fine-tuning on high-quality data, (2) LLMs can master multiple complex card games simultaneously, with performance augmentation for games with similar rules and conflicts for dissimilar ones, and (3) LLMs experience a decline in general capabilities when mastering complex games, but this decline can be mitigated by integrating a certain amount of general instruction data. The evaluation results demonstrate strong learning ability and versatility of LLMs.
Authors: Mohammad Zbeeb, Hasan Abed Al Kader Hammoud, Bernard Ghanem
Abstract: Large language models often require costly optimization, such as reinforcement learning, to master complex reasoning tasks. This work demonstrates that reasoning ability, once learned, can be extracted and transferred between models as a compact task vector. We source two publicly available, identically initialized Qwen2.5 models, one fine-tuned with supervised fine-tuning (SFT) and the other with group relative policy optimization (GRPO) on the same dataset. From these, we extract a reasoning vector: $v_{\text{reason}} = \theta_{\text{GRPO}} - \theta_{\text{SFT}}$. We hypothesize that this vector captures the reasoning capability instilled by reinforcement learning while factoring out shared knowledge from the SFT process. When added to compatible instruction-tuned models through simple arithmetic, this vector consistently improves performance across diverse reasoning benchmarks: GSM8K (+4.9%), HumanEval (+4.3%), SciQ (+1.7%), and BigBenchHard (+12.3% for the 1.5B model). The performance improvements persist under adversarial conditions. Conversely, subtracting the vector causes significant performance degradation (-11.8% on GSM8K), demonstrating the vector's strong contribution to the model's reasoning abilities. This work shows how reasoning capabilities, typically developed through expensive training, can be extracted from existing open-source models and reused through simple tensor arithmetic, offering a practical way to enhance models by recycling prior computational investments.
Authors: Paloma Piot, Diego S\'anchez, Javier Parapar
Abstract: Online harms are a growing problem in digital spaces, putting user safety at risk and reducing trust in social media platforms. One of the most persistent forms of harm is hate speech. To address this, we need tools that combine the speed and scale of automated systems with the judgment and insight of human moderators. These tools should not only find harmful content but also explain their decisions clearly, helping to build trust and understanding. In this paper, we present WATCHED, a chatbot designed to support content moderators in tackling hate speech. The chatbot is built as an Artificial Intelligence Agent system that uses Large Language Models along with several specialised tools. It compares new posts with real examples of hate speech and neutral content, uses a BERT-based classifier to help flag harmful messages, looks up slang and informal language using sources like Urban Dictionary, generates chain-of-thought reasoning, and checks platform guidelines to explain and support its decisions. This combination allows the chatbot not only to detect hate speech but to explain why content is considered harmful, grounded in both precedent and policy. Experimental results show that our proposed method surpasses existing state-of-the-art methods, reaching a macro F1 score of 0.91. Designed for moderators, safety teams, and researchers, the tool helps reduce online harms by supporting collaboration between AI and human oversight.
Authors: Serwar Basch, Ilia Kuznetsov, Tom Hope, Iryna Gurevych
Abstract: Understanding fine-grained relations between documents is crucial for many application domains. However, the study of automated assistance is limited by the lack of efficient methods to create training and evaluation datasets of cross-document links. To address this, we introduce a new domain-agnostic framework for selecting a best-performing approach and annotating cross-document links in a new domain from scratch. We first generate and validate semi-synthetic datasets of interconnected documents. This data is used to perform automatic evaluation, producing a shortlist of best-performing linking approaches. These approaches are then used in an extensive human evaluation study, yielding performance estimates on natural text pairs. We apply our framework in two distinct domains -- peer review and news -- and show that combining retrieval models with LLMs achieves 78\% link approval from human raters, more than doubling the precision of strong retrievers alone. Our framework enables systematic study of cross-document understanding across application scenarios, and the resulting novel datasets lay foundation for numerous cross-document tasks like media framing and peer review. We make the code, data, and annotation protocols openly available.
Authors: Joonyong Park, Shinnosuke Takamichi, David M. Chan, Shunsuke Kando, Yuki Saito, Hiroshi Saruwatari
Abstract: This study presents a comparative analysis of the statistical and linguistic properties of neural audio codecs (NACs). We investigate discrete speech tokens produced by various NAC models, examining their adherence to linguistic statistical laws such as Zipf's law and Heaps' law, as well as their entropy and redundancy. To assess how these token-level properties relate to semantic and acoustic preservation in synthesized speech, we evaluate intelligibility using error rates of automatic speech recognition, and quality using the UTMOS score. Our results reveal that NAC tokens, particularly 3-grams, exhibit language-like statistical patterns. Moreover, these properties, together with measures of information content, are found to correlate with improved performances in speech recognition and resynthesis tasks. These findings offer insights into the structure of NAC token sequences and inform the design of more effective generative speech models.
Authors: KV Aditya Srivatsa, Kaushal Kumar Maurya, Ekaterina Kochmar
Abstract: Large language models (LLMs) demonstrate remarkable performance on math word problems, yet they have been shown to struggle with meta-reasoning tasks such as identifying errors in student solutions. In this work, we investigate the challenge of locating the first error step in stepwise solutions using two error reasoning datasets: VtG and PRM800K. Our experiments show that state-of-the-art LLMs struggle to locate the first error step in student solutions even when given access to the reference solution. To that end, we propose an approach that generates an intermediate corrected student solution, aligning more closely with the original student's solution, which helps improve performance.
Authors: Kaviraj Pather, Elena Hadjigeorgiou, Arben Krasniqi, Claire Schmit, Irina Rusu, Marc Pons, Kabir Khan
Abstract: Large language models (LLMs) show strong reasoning via chain-of-thought (CoT) prompting, but the process is opaque, which makes verification, debugging, and control difficult in high-stakes settings. We present Vis-CoT, a human-in-the-loop framework that converts linear CoT text into an interactive reasoning graph. Users can visualize the logical flow, identify flawed steps, and intervene by pruning incorrect paths and grafting new, user-defined premises. This shifts interaction from passive observation to active collaboration, steering models toward more accurate and trustworthy conclusions. Across GSM8K and StrategyQA, Vis-CoT improves final-answer accuracy by up to 24 percentage points over non-interactive baselines. A user study also shows large gains in perceived usability and trust. Vis-CoT points to a practical path for more reliable, understandable, and collaborative reasoning by combining LLMs with targeted human oversight.
Authors: Yang Liu, Masahiro Kaneko, Chenhui Chu
Abstract: Today's large language models (LLMs) are capable of supporting multilingual scenarios, allowing users to interact with LLMs in their native languages. When LLMs respond to subjective questions posed by users, they are expected to align with the views of specific demographic groups or historical periods, shaped by the language in which the user interacts with the model. Existing studies mainly focus on researching the opinions represented by LLMs among demographic groups in the United States or a few countries, lacking worldwide country samples and studies on human opinions in different historical periods, as well as lacking discussion on using language to steer LLMs. Moreover, they also overlook the potential influence of prompt language on the alignment of LLMs' opinions. In this study, our goal is to fill these gaps. To this end, we create an evaluation framework based on the World Values Survey (WVS) to systematically assess the alignment of LLMs with human opinions across different countries, languages, and historical periods around the world. We find that LLMs appropriately or over-align the opinions with only a few countries while under-aligning the opinions with most countries. Furthermore, changing the language of the prompt to match the language used in the questionnaire can effectively steer LLMs to align with the opinions of the corresponding country more effectively than existing steering methods. At the same time, LLMs are more aligned with the opinions of the contemporary population. To our knowledge, our study is the first comprehensive investigation of the topic of opinion alignment in LLMs across global, language, and temporal dimensions. Our code and data are publicly available at https://github.com/nlply/global-opinion-alignment.
Authors: Markus Oehri, Giulia Conti, Kaviraj Pather, Alexandre Rossi, Laia Serra, Adrian Parody, Rogvi Johannesen, Aviaja Petersen, Arben Krasniqi
Abstract: Deployed language models must decide not only what to answer but also when not to answer. We present UniCR, a unified framework that turns heterogeneous uncertainty evidence including sequence likelihoods, self-consistency dispersion, retrieval compatibility, and tool or verifier feedback into a calibrated probability of correctness and then enforces a user-specified error budget via principled refusal. UniCR learns a lightweight calibration head with temperature scaling and proper scoring, supports API-only models through black-box features, and offers distribution-free guarantees using conformal risk control. For long-form generation, we align confidence with semantic fidelity by supervising on atomic factuality scores derived from retrieved evidence, reducing confident hallucinations while preserving coverage. Experiments on short-form QA, code generation with execution tests, and retrieval-augmented long-form QA show consistent improvements in calibration metrics, lower area under the risk-coverage curve, and higher coverage at fixed risk compared to entropy or logit thresholds, post-hoc calibrators, and end-to-end selective baselines. Analyses reveal that evidence contradiction, semantic dispersion, and tool inconsistency are the dominant drivers of abstention, yielding informative user-facing refusal messages. The result is a portable recipe of evidence fusion to calibrated probability to risk-controlled decision that improves trustworthiness without fine-tuning the base model and remains valid under distribution shift.
Authors: Yuchen Wu, Liang Ding, Li Shen, Dacheng Tao
Abstract: Large language models (LLMs) encode vast amounts of world knowledge but remain static once trained, making the timely integration of emerging facts prohibitively expensive via full retraining. Knowledge-editing techniques have thus emerged to inject or overwrite specific facts into LLMs, yet they either over-rely on superficial cues or incur complex, iterative pipelines that collapse under noisy, multi-hop conditions. We introduce Reason-KE, an end-to-end reasoning-chain-based editing framework that steers a pretrained LLM through four structured stages-fact acknowledgment, relevance determination, selective application, and final reasoning-to filter distractors in a single pass. Trained on MQuAKE-CF with up to four irrelevant facts, Reason-KE elevates Qwen2.5-7B's multi-hop QA accuracy to 90.2% while suffering merely a 6.3% drop under heavy distraction and <1% when answers are leaked. Our quantitative analysis confirms Reason-KE's resilience and efficiency, establishing a new state-of-the-art for reliable LLM knowledge updates.
Authors: Youchao Zhou, Heyan Huang, Yicheng Liu, Rui Dai, Xinglin Wang, Xingchen Zhang, Shumin Shi, Yang Deng
Abstract: Existing Large Language Models (LLMs) occasionally generate plausible yet factually incorrect responses, known as hallucinations. Researchers are primarily using two approaches to mitigate hallucinations, namely Retrieval Augmented Language Models (RALMs) and refusal post-training. However, current research predominantly emphasizes their individual effectiveness while overlooking the evaluation of the refusal capability of RALMs. In this study, we ask the fundamental question: Do RALMs know when they don't know? Specifically, we ask three questions. First, are RALMs well-calibrated regarding different internal and external knowledge states? We examine the influence of various factors. Contrary to expectations, we find that LLMs exhibit significant \textbf{over-refusal} behavior. Then, how does refusal post-training affect the over-refusal issue? We investigate the Refusal-aware Instruction Tuning and In-Context Fine-tuning methods. Our results show that the over-refusal problem is mitigated by In-context fine-tuning. but magnified by R-tuning. However, we also find that the refusal ability may conflict with the quality of the answer. Finally, we develop a simple yet effective refusal method for refusal post-trained models to improve their overall answer quality in terms of refusal and correct answers. Our study provides a more comprehensive understanding of the influence of important factors on RALM systems.
Authors: Andreas Ottem
Abstract: Retrieval-Augmented Generation (RAG) systems typically face constraints because of their inherent mechanism: a simple top-k semantic search [1]. The approach often leads to the incorporation of irrelevant or redundant information in the context, degrading performance and efficiency [10][11]. This paper presents MeVe, a novel modular architecture intended for Memory Verification and smart context composition. MeVe rethinks the RAG paradigm by proposing a five-phase modular design that distinctly breaks down the retrieval and context composition process into distinct, auditable, and independently tunable phases: initial retrieval, relevance verification, fallback retrieval, context prioritization, and token budgeting. This architecture enables fine-grained control of what knowledge is made available to an LLM, enabling task-dependent filtering and adaptation. We release a reference implementation of MeVe as a proof of concept and evaluate its performance on knowledge-heavy QA tasks over a subset of English Wikipedia [22]. Our results demonstrate that by actively verifying information before composition, MeVe significantly improves context efficiency, achieving a 57% reduction on the Wikipedia dataset and a 75% reduction on the more complex HotpotQA dataset compared to standard RAG implementations [25]. This work provides a framework for more scalable and reliable LLM applications. By refining and distilling contextual information, MeVe offers a path toward better grounding and more accurate factual support [16].
Authors: Thomas Compton
Abstract: This paper presents a comparative analysis of community unionism (CU) in two distinct historical and organizational contexts: the National Boot and Shoe Union (B\&S) in the 1920s and Unite Community in the 2010s--2020s. Using BERTopic for thematic modeling and cTF-IDF weighting, alongside word frequency analysis, the study examines the extent to which each union's discourse aligns with key features of CU -- such as coalition-building, grassroots engagement, and action beyond the workplace. The results reveal significant differences in thematic focus and discursive coherence. While Unite Community demonstrates stronger alignment with outward-facing, social justice-oriented themes, the B\&S corpus emphasizes internal administration, industrial relations, and member services -- reflecting a more traditional, servicing-oriented union model. The analysis also highlights methodological insights, demonstrating how modern NLP techniques can enhance the study of historical labor archives. Ultimately, the findings suggest that while both unions engage with community-related themes, their underlying models of engagement diverge significantly, challenging assumptions about the continuity and universality of community unionism across time and sector.
Authors: Kairong Han, Wenshuo Zhao, Ziyu Zhao, JunJian Ye, Lujia Pan, Kun Kuang
Abstract: Large Language Models (LLMs) have achieved remarkable success across various domains. However, a fundamental question remains: Can LLMs effectively utilize causal knowledge for prediction and generation? Through empirical studies, we find that LLMs trained directly on large-scale data often capture spurious correlations rather than true causal relationships, leading to suboptimal performance, especially in out-of-distribution (OOD) scenarios. To address this challenge, we propose Causal Attention Tuning (CAT), a novel approach that injects fine-grained causal knowledge into the attention mechanism. We propose an automated pipeline that leverages human priors to automatically generate token-level causal signals and introduce the Re-Attention mechanism to guide training, helping the model focus on causal structures while mitigating noise and biases in attention scores. Experimental results on our proposed Spurious Token Game (STG) benchmark and multiple downstream tasks demonstrate that our approach effectively leverages causal knowledge for prediction and remains robust in OOD scenarios. Implementation details can be found at https://github.com/Kairong-Han/CAT.
Authors: Seungkyu Lee, Nalim Kim, Yohan Jo
Abstract: Tool agents -- LLM-based systems that interact with external APIs -- offer a way to execute real-world tasks. However, as tasks become increasingly complex, these agents struggle to identify and call the correct APIs in the proper order. To tackle this problem, we investigate converting API documentation into a structured API graph that captures API dependencies and leveraging it for multi-tool queries that require compositional API calls. To support this, we introduce In-N-Out, the first expert-annotated dataset of API graphs built from two real-world API benchmarks and their documentation. Using In-N-Out significantly improves performance on both tool retrieval and multi-tool query generation, nearly doubling that of LLMs using documentation alone. Moreover, graphs generated by models fine-tuned on In-N-Out close 90% of this gap, showing that our dataset helps models learn to comprehend API documentation and parameter relationships. Our findings highlight the promise of using explicit API graphs for tool agents and the utility of In-N-Out as a valuable resource. We will release the dataset and code publicly.
Authors: Zeguan Xiao, Diyang Dou, Boya Xiong, Yun Chen, Guanhua Chen
Abstract: Large Language Models (LLMs) have achieved remarkable success across a wide range of natural language tasks, but often exhibit overconfidence and generate plausible yet incorrect answers. This overconfidence, especially in models undergone Reinforcement Learning from Human Feedback (RLHF), poses significant challenges for reliable uncertainty estimation and safe deployment. In this paper, we propose EAGLE (Expectation of AGgregated internaL bEief), a novel self-evaluation-based calibration method that leverages the internal hidden states of LLMs to derive more accurate confidence scores. Instead of relying on the model's final output, our approach extracts internal beliefs from multiple intermediate layers during self-evaluation. By aggregating these layer-wise beliefs and calculating the expectation over the resulting confidence score distribution, EAGLE produces a refined confidence score that more faithfully reflects the model's internal certainty. Extensive experiments on diverse datasets and LLMs demonstrate that EAGLE significantly improves calibration performance over existing baselines. We also provide an in-depth analysis of EAGLE, including a layer-wise examination of uncertainty patterns, a study of the impact of self-evaluation prompts, and an analysis of the effect of self-evaluation score range.
Authors: Vivi Nastase, Paola Merlo
Abstract: Transformer models learn to encode and decode an input text, and produce contextual token embeddings as a side-effect. The mapping from language into the embedding space maps words expressing similar concepts onto points that are close in the space. In practice, the reverse implication is also assumed: words corresponding to close points in this space are similar or related, those that are further are not. Does closeness in the embedding space extend to shared properties for sentence embeddings? We present an investigation of sentence embeddings and show that the geometry of their embedding space is not predictive of their relative performances on a variety of tasks. We compute sentence embeddings in three ways: as averaged token embeddings, as the embedding of the special [CLS] token, and as the embedding of a random token from the sentence. We explore whether there is a correlation between the distance between sentence embedding variations and their performance on linguistic tasks, and whether despite their distances, they do encode the same information in the same manner. The results show that the cosine similarity -- which treats dimensions shallowly -- captures (shallow) commonalities or differences between sentence embeddings, which are not predictive of their performance on specific tasks. Linguistic information is rather encoded in weighted combinations of different dimensions, which are not reflected in the geometry of the sentence embedding space.
Authors: Shanshan Wang, Junchao Wu, Fengying Ye, Jingming Yao, Lidia S. Chao, Derek F. Wong
Abstract: The rapid development of advanced large language models (LLMs) has made AI-generated text indistinguishable from human-written text. Previous work on detecting AI-generated text has made effective progress, but has not involved modern Chinese poetry. Due to the distinctive characteristics of modern Chinese poetry, it is difficult to identify whether a poem originated from humans or AI. The proliferation of AI-generated modern Chinese poetry has significantly disrupted the poetry ecosystem. Based on the urgency of identifying AI-generated poetry in the real Chinese world, this paper proposes a novel benchmark for detecting LLMs-generated modern Chinese poetry. We first construct a high-quality dataset, which includes both 800 poems written by six professional poets and 41,600 poems generated by four mainstream LLMs. Subsequently, we conduct systematic performance assessments of six detectors on this dataset. Experimental results demonstrate that current detectors cannot be used as reliable tools to detect modern Chinese poems generated by LLMs. The most difficult poetic features to detect are intrinsic qualities, especially style. The detection results verify the effectiveness and necessity of our proposed benchmark. Our work lays a foundation for future detection of AI-generated poetry.
Authors: Hind Aljuaid, Areej Alhothali, Ohoud Al-Zamzami, Hussein Assalahi
Abstract: Essay writing is a critical component of student assessment, yet manual scoring is labor-intensive and inconsistent. Automated Essay Scoring (AES) offers a promising alternative, but current approaches face limitations. Recent studies have incorporated Graph Neural Networks (GNNs) into AES using static word embeddings that fail to capture contextual meaning, especially for polysemous words. Additionally, many methods rely on holistic scoring, overlooking specific writing aspects such as grammar, vocabulary, and cohesion. To address these challenges, this study proposes TransGAT, a novel approach that integrates fine-tuned Transformer models with GNNs for analytic scoring. TransGAT combines the contextual understanding of Transformers with the relational modeling strength of Graph Attention Networks (GAT). It performs two-stream predictions by pairing each fine-tuned Transformer (BERT, RoBERTa, and DeBERTaV3) with a separate GAT. In each pair, the first stream generates essay-level predictions, while the second applies GAT to Transformer token embeddings, with edges constructed from syntactic dependencies. The model then fuses predictions from both streams to produce the final analytic score. Experiments on the ELLIPSE dataset show that TransGAT outperforms baseline models, achieving an average Quadratic Weighted Kappa (QWK) of 0.854 across all analytic scoring dimensions. These findings highlight the potential of TransGAT to advance AES systems.
Authors: Dominic Plein
Abstract: We present a method to calculate the similarity between words based on their phonetic transcription (their pronunciation) using the Needleman-Wunsch algorithm. We implement this algorithm in Rust and parallelize it on both CPU and GPU to handle large datasets efficiently. The GPU implementation leverages CUDA and the cudarc Rust library to achieve significant performance improvements. We validate our approach by constructing a fully-connected graph where nodes represent words and edges have weights according to the similarity between the words. This graph is then analyzed using clustering algorithms to identify groups of phonetically similar words. Our results demonstrate the feasibility and effectiveness of the proposed method in analyzing the phonetic structure of languages. It might be easily expanded to other languages.
Authors: Zhengjia Wang, Qiang Sheng, Danding Wang, Beizhe Hu, Juan Cao
Abstract: Fake news detection is an important and challenging task for defending online information integrity. Existing state-of-the-art approaches typically extract news semantic clues, such as writing patterns that include emotional words, stylistic features, etc. However, detectors tuned solely to such semantic clues can easily fall into surface detection patterns, which can shift rapidly in dynamic environments, leading to limited performance in the evolving news landscape. To address this issue, this paper investigates a novel perspective by incorporating news intent into fake news detection, bridging intents and semantics together. The core insight is that by considering news intents, one can deeply understand the inherent thoughts behind news deception, rather than the surface patterns within words alone. To achieve this goal, we propose Graph-based Intent-Semantic Joint Modeling (InSide) for fake news detection, which models deception clues from both semantic and intent signals via graph-based joint learning. Specifically, InSide reformulates news semantic and intent signals into heterogeneous graph structures, enabling long-range context interaction through entity guidance and capturing both holistic and implementation-level intent via coarse-to-fine intent modeling. To achieve better alignment between semantics and intents, we further develop a dynamic pathway-based graph alignment strategy for effective message passing and aggregation across these signals by establishing a common space. Extensive experiments on four benchmark datasets demonstrate the superiority of the proposed InSide compared to state-of-the-art methods.
Authors: Abdelkrime Aries
Abstract: Pre-trained language models (PLMs) have substantially advanced natural language processing by providing context-sensitive text representations. However, the Algerian dialect remains under-represented, with few dedicated models available. Processing this dialect is challenging due to its complex morphology, frequent code-switching, multiple scripts, and strong lexical influences from other languages. These characteristics complicate tokenization and reduce the effectiveness of conventional word- or subword-level approaches. To address this gap, we introduce chDzDT, a character-level pre-trained language model tailored for Algerian morphology. Unlike conventional PLMs that rely on token sequences, chDzDT is trained on isolated words. This design allows the model to encode morphological patterns robustly, without depending on token boundaries or standardized orthography. The training corpus draws from diverse sources, including YouTube comments, French, English, and Berber Wikipedia, as well as the Tatoeba project. It covers multiple scripts and linguistic varieties, resulting in a substantial pre-training workload. Our contributions are threefold: (i) a detailed morphological analysis of Algerian dialect using YouTube comments; (ii) the construction of a multilingual Algerian lexicon dataset; and (iii) the development and extensive evaluation of a character-level PLM as a morphology-focused encoder for downstream tasks. The proposed approach demonstrates the potential of character-level modeling for morphologically rich, low-resource dialects and lays a foundation for more inclusive and adaptable NLP systems.
Authors: Andong Hua, Kenan Tang, Chenhe Gu, Jindong Gu, Eric Wong, Yao Qin
Abstract: Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs. In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes? To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates. We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases. When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts. Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.
Authors: Shreyas Tirumala, Nishant Jain, Danny D. Leybzon, Trent D. Buskirk
Abstract: Transformer-based Large Language Models (LLMs) have paved the way for "AI interviewers" that can administer voice-based surveys with respondents in real-time. This position paper reviews emerging evidence to understand when such AI interviewing systems are fit for purpose for collecting data within quantitative and qualitative research contexts. We evaluate the capabilities of AI interviewers as well as current Interactive Voice Response (IVR) systems across two dimensions: input/output performance (i.e., speech recognition, answer recording, emotion handling) and verbal reasoning (i.e., ability to probe, clarify, and handle branching logic). Field studies suggest that AI interviewers already exceed IVR capabilities for both quantitative and qualitative data collection, but real-time transcription error rates, limited emotion detection abilities, and uneven follow-up quality indicate that the utility, use and adoption of current AI interviewer technology may be context-dependent for qualitative data collection efforts.
Authors: Zhimeng Luo, Abhibha Gupta, Adam Frisch, Daqing He
Abstract: The extraction of critical patient information from Electronic Health Records (EHRs) poses significant challenges due to the complexity and unstructured nature of the data. Traditional machine learning approaches often fail to capture pertinent details efficiently, making it difficult for clinicians to utilize these tools effectively in patient care. This paper introduces a novel approach to extracting the OPQRST assessment from EHRs by leveraging the capabilities of Large Language Models (LLMs). We propose to reframe the task from sequence labeling to text generation, enabling the models to provide reasoning steps that mimic a physician's cognitive processes. This approach enhances interpretability and adapts to the limited availability of labeled data in healthcare settings. Furthermore, we address the challenge of evaluating the accuracy of machine-generated text in clinical contexts by proposing a modification to traditional Named Entity Recognition (NER) metrics. This includes the integration of semantic similarity measures, such as the BERT Score, to assess the alignment between generated text and the clinical intent of the original records. Our contributions demonstrate a significant advancement in the use of AI in healthcare, offering a scalable solution that improves the accuracy and usability of information extraction from EHRs, thereby aiding clinicians in making more informed decisions and enhancing patient care outcomes.
Authors: Zhimeng Luo, Zhendong Wang, Rui Meng, Diyang Xue, Adam Frisch, Daqing He
Abstract: A Chief complaint (CC) is the reason for the medical visit as stated in the patient's own words. It helps medical professionals to quickly understand a patient's situation, and also serves as a short summary for medical text mining. However, chief complaint records often take a variety of entering methods, resulting in a wide variation of medical notations, which makes it difficult to standardize across different medical institutions for record keeping or text mining. In this study, we propose a weakly supervised method to automatically extract and link entities in chief complaints in the absence of human annotation. We first adopt a split-and-match algorithm to produce weak annotations, including entity mention spans and class labels, on 1.2 million real-world de-identified and IRB approved chief complaint records. Then we train a BERT-based model with generated weak labels to locate entity mentions in chief complaint text and link them to a pre-defined ontology. We conducted extensive experiments, and the results showed that our Weakly Supervised Entity Extraction and Linking (\ours) method produced superior performance over previous methods without any human annotation.
Authors: Sachin Pawar, Manoj Apte, Girish K. Palshikar, Basit Ali, Nitin Ramrakhiyani
Abstract: Disputes between two parties occur in almost all domains such as taxation, insurance, banking, healthcare, etc. The disputes are generally resolved in a specific forum (e.g., consumer court) where facts are presented, points of disagreement are discussed, arguments as well as specific demands of the parties are heard, and finally a human judge resolves the dispute by often favouring one of the two parties. In this paper, we explore the use of large language models (LLMs) as assistants for the human judge to resolve such disputes, as part of our DRAssist system. We focus on disputes from two specific domains -- automobile insurance and domain name disputes. DRAssist identifies certain key structural elements (e.g., facts, aspects or disagreement, arguments) of the disputes and summarizes the unstructured dispute descriptions to produce a structured summary for each dispute. We then explore multiple prompting strategies with multiple LLMs for their ability to assist in resolving the disputes in these domains. In DRAssist, these LLMs are prompted to produce the resolution output at three different levels -- (i) identifying an overall stronger party in a dispute, (ii) decide whether each specific demand of each contesting party can be accepted or not, (iii) evaluate whether each argument by each contesting party is strong or weak. We evaluate the performance of LLMs on all these tasks by comparing them with relevant baselines using suitable evaluation metrics.
Authors: Chao Xue, Ziyuan Gao
Abstract: Text semantic matching requires nuanced understanding of both structural relationships and fine-grained semantic distinctions. While pre-trained language models excel at capturing token-level interactions, they often overlook hierarchical structural patterns and struggle with subtle semantic discrimination. In this paper, we proposed StructCoh, a graph-enhanced contrastive learning framework that synergistically combines structural reasoning with representation space optimization. Our approach features two key innovations: (1) A dual-graph encoder constructs semantic graphs via dependency parsing and topic modeling, then employs graph isomorphism networks to propagate structural features across syntactic dependencies and cross-document concept nodes. (2) A hierarchical contrastive objective enforces consistency at multiple granularities: node-level contrastive regularization preserves core semantic units, while graph-aware contrastive learning aligns inter-document structural semantics through both explicit and implicit negative sampling strategies. Experiments on three legal document matching benchmarks and academic plagiarism detection datasets demonstrate significant improvements over state-of-the-art methods. Notably, StructCoh achieves 86.7% F1-score (+6.2% absolute gain) on legal statute matching by effectively identifying argument structure similarities.
Authors: Hexian Zhang, Xinyu Yan, Yanqi Yang, Lijian Jin, Ping Yang, Junwen Wang
Abstract: Large language models (LLMs) hold transformative potential in healthcare, yet their capacity to interpret longitudinal patient narratives remains inadequately explored. Dentistry, with its rich repository of structured clinical data, presents a unique opportunity to rigorously assess LLMs' reasoning abilities. While several commercial LLMs already exist, DeepSeek, a model that gained significant attention earlier this year, has also joined the competition. This study evaluated four state-of-the-art LLMs (GPT-4o, Gemini 2.0 Flash, Copilot, and DeepSeek V3) on their ability to analyze longitudinal dental case vignettes through open-ended clinical tasks. Using 34 standardized longitudinal periodontal cases (comprising 258 question-answer pairs), we assessed model performance via automated metrics and blinded evaluations by licensed dentists. DeepSeek emerged as the top performer, demonstrating superior faithfulness (median score = 0.528 vs. 0.367-0.457) and higher expert ratings (median = 4.5/5 vs. 4.0/5), without significantly compromising readability. Our study positions DeepSeek as the leading LLM for case analysis, endorses its integration as an adjunct tool in both medical education and research, and highlights its potential as a domain-specific agent.
Authors: Bashar Talafha, Hawau Olamide Toyin, Peter Sullivan, AbdelRahim Elmadany, Abdurrahman Juma, Amirbek Djanibekov, Chiyu Zhang, Hamad Alshehhi, Hanan Aldarmaki, Mustafa Jarrar, Nizar Habash, Muhammad Abdul-Mageed
Abstract: We present the findings of the sixth Nuanced Arabic Dialect Identification (NADI 2025) Shared Task, which focused on Arabic speech dialect processing across three subtasks: spoken dialect identification (Subtask 1), speech recognition (Subtask 2), and diacritic restoration for spoken dialects (Subtask 3). A total of 44 teams registered, and during the testing phase, 100 valid submissions were received from eight unique teams. The distribution was as follows: 34 submissions for Subtask 1 "five teams{\ae}, 47 submissions for Subtask 2 "six teams", and 19 submissions for Subtask 3 "two teams". The best-performing systems achieved 79.8% accuracy on Subtask 1, 35.68/12.20 WER/CER (overall average) on Subtask 2, and 55/13 WER/CER on Subtask 3. These results highlight the ongoing challenges of Arabic dialect speech processing, particularly in dialect identification, recognition, and diacritic restoration. We also summarize the methods adopted by participating teams and briefly outline directions for future editions of NADI.
Authors: Guangzeng Han, Weisi Liu, Xiaolei Huang
Abstract: Large Language Models (LLMs) excel at generating synthetic data, but ensuring its quality and diversity remains challenging. We propose Genetic Prompt, a novel framework that combines genetic algorithms with LLMs to augment synthetic data generation. Our approach treats semantic text attributes as gene sequences and leverages the LLM to simulate crossover and mutation operations. This genetic process enhances data quality and diversity by creating novel attribute combinations, yielding synthetic distributions closer to real-world data. To optimize parent selection, we also integrate an active learning scheme that expands the offspring search space. Our experiments on multiple NLP tasks reveal several key findings: Genetic Prompt not only significantly outperforms state-of-the-art baselines but also shows robust performance across various generator model sizes and scales. Moreover, we demonstrate that fusing our synthetic data with the original training set significantly boosts downstream model performance, particularly for class-imbalanced scenarios. Our findings validate that Genetic Prompt is an effective method for producing high-quality synthetic data for a wide range of NLP applications.
Authors: Elisabetta Rocchetti, Alfio Ferrara
Abstract: Adhering to explicit length constraints, such as generating text with a precise word count, remains a significant challenge for Large Language Models (LLMs). This study aims at investigating the differences between foundation models and their instruction-tuned counterparts, on length-controlled text generation in English and Italian. We analyze both performance and internal component contributions using Cumulative Weighted Attribution, a metric derived from Direct Logit Attribution. Our findings reveal that instruction-tuning substantially improves length control, primarily by specializing components in deeper model layers. Specifically, attention heads in later layers of IT models show increasingly positive contributions, particularly in English. In Italian, while attention contributions are more attenuated, final-layer MLPs exhibit a stronger positive role, suggesting a compensatory mechanism. These results indicate that instruction-tuning reconfigures later layers for task adherence, with component-level strategies potentially adapting to linguistic context.
Authors: Juhyeon Lee, Wonduk Seo, Hyunjin An, Seunghyun Lee, Yi Bu
Abstract: Automatic prompt optimization has recently emerged as a strategy for improving the quality of prompts used in Large Language Models (LLMs), with the goal of generating more accurate and useful responses. However, most prior work focuses on direct prompt refinement or model fine-tuning, overlooking the potential of leveraging LLMs' inherent reasoning capability to learn from contrasting examples. In this paper, we present Contrastive Reasoning Prompt Optimization (CRPO), a novel framework that formulates prompt optimization as a retrieval augmented reasoning process. Our approach retrieves top k reference prompts from the HelpSteer2 dataset, an open-source collection annotated for helpfulness, correctness, coherence, complexity, and verbosity, and constructs two complementary optimization paradigms: (1) tiered contrastive reasoning, where the LLM compares high, medium, and low quality prompts to refine its own generation through reflective reasoning, and (2) multi-metric contrastive reasoning, where the LLM analyzes the best prompts along each evaluation dimension and integrates their strengths into an optimized prompt. By explicitly contrasting high and low quality exemplars, CRPO enables the model to deduce why certain prompts succeed while others fail, thereby achieving more robust and interpretable optimization. Experimental results on the HelpSteer2 benchmark demonstrate that CRPO significantly outperforms baselines. Our findings highlight the promise of contrastive, retrieval-augmented reasoning for advancing automatic prompt optimization.
Authors: Zhichao Shi, Xuhui Jiang, Chengjin Xu, Cangli Yao, Zhenxin Huang, Shengjie Ma, Yinghan Shen, Yuanzhuo Wang
Abstract: Evaluating the capabilities of large language models (LLMs) is an essential step to ensure the successful application of LLMs across various domains. The current evaluation of LLMs is based on a paradigm that involves querying them with predefined question sets and assessing their outputs. This paradigm offers controllable processes and simplicity, but faces challenges such as limited interaction with targets, insufficient difficulty control, and difficulties in verifying the validity of evaluation results, making it hard to precisely determine the knowledge and capability boundaries of target models. To address these challenges, we propose JudgeAgent, a knowledge-target adaptive dynamic evaluation framework based on a new interviewer-style evaluation paradigm. JudgeAgent employs a comprehensive evaluation approach consisting of benchmark grading, interactive extension, and evaluation feedback. It utilizes knowledge-driven data synthesis and target-adaptive difficulty adjustment methods to conduct extended testing, providing accurate and effective evaluation results. We also introduce a novel insight into validating evaluation methods, demonstrating the effectiveness of JudgeAgent and its dynamic evaluation paradigm through extensive experiments.
Authors: Wang Chen, Guanqiang Qi, Weikang Li, Yang Li
Abstract: Retrieval-Augmented Generation (RAG) has become a core paradigm in document question answering tasks. However, existing methods have limitations when dealing with multimodal documents: one category of methods relies on layout analysis and text extraction, which can only utilize explicit text information and struggle to capture images or unstructured content; the other category treats document segmentation as visual input and directly passes it to visual language models (VLMs) for processing, yet it ignores the semantic advantages of text, leading to suboptimal generation results. This paper proposes co-modality-based RAG (CMRAG), which can simultaneously leverage text and images for efficient retrieval and generation. Specifically, we first perform structured parsing on documents to obtain co-modality representations of text segments and image regions. Subsequently, in response to user queries, we retrieve candidate evidence from text and image channels, respectively, and aggregate the results at the cross-modal retrieval level. Finally, we prompt the VLM to generate the final response based on the co-modality retrieval results. Experiments demonstrate that our method significantly outperforms pure-vision-based RAG in visual document question answering tasks. The findings of this paper show that integrating co-modality information into the RAG framework in a unified manner is an effective approach to improving the performance of complex document visual question-answering (VQA) systems.
Authors: Snehasis Mukhopadhyay, Aryan Kasat, Shivam Dubey, Rahul Karthikeyan, Dhruv Sood, Vinija Jain, Aman Chadha, Amitava Das
Abstract: Large Language Models (LLMs) can inadvertently reflect societal biases present in their training data, leading to harmful or prejudiced outputs. In the Indian context, our empirical evaluations across a suite of models reveal that biases around caste and religion are particularly salient. Yet, most existing mitigation strategies are Western-centric and fail to address these local nuances. We propose AMBEDKAR, a framework inspired by the egalitarian vision of Dr B. R. Ambedkar, architect of the Indian Constitution, to guide LLM outputs toward fairness, neutrality, and inclusion in line with Articles 14 to 17. Our approach introduces a Constitution-Aware Decoding Layer, guided by the AI Constitution of India and applied only at inference time, without any parameter updates to the base model. We incorporate a speculative decoding algorithm that proactively reduces casteist and communal bias during generation. This mitigation layer operates directly within the decoding process, avoiding changes to model internals and lowering the computational and infrastructural costs associated with retraining. We reinterpret speculative decoding not merely as an efficiency tool but as a mechanism for fairness. In this framework, a Small Language Model (SLM) acts as a potentially biased generator, while a constitutionally guided Large Language Model (LLM) serves as the verifier. Rather than accelerating generation, the LLM enforces bias-robust trajectories in the SLM outputs. This inversion of roles gives rise to a fairness-by-speculation paradigm. Our approach yields an absolute reduction of bias up to 26.41 percent compared to baseline. Our source code, datasets, and results are available at https://anonymous.4open.science/r/AMBEDKAR-983B/
Authors: David Demitri Africa, Suchir Salhan, Yuval Weiss, Paula Buttery, Richard Diehl Martinez
Abstract: Named-entity recognition (NER) in low-resource languages is usually tackled by finetuning very large multilingual LMs, an option that is often infeasible in memory- or latency-constrained settings. We ask whether small decoder LMs can be pretrained so that they adapt quickly and transfer zero-shot to languages unseen during pretraining. To this end we replace part of the autoregressive objective with first-order model-agnostic meta-learning (MAML). Tagalog and Cebuano are typologically similar yet structurally different in their actor/non-actor voice systems, and hence serve as a challenging test-bed. Across four model sizes (11 M - 570 M) MAML lifts zero-shot micro-F1 by 2-6 pp under head-only tuning and 1-3 pp after full tuning, while cutting convergence time by up to 8%. Gains are largest for single-token person entities that co-occur with Tagalog case particles si/ni, highlighting the importance of surface anchors.
Authors: Kyeongman Park, Nakyeong Yang, Kyomin Jung
Abstract: Large Language Models (LLMs) often generate repetitive and monotonous outputs, especially in tasks like story generation, due to limited creative diversity when given the same input prompt. To address this challenge, we propose a novel decoding strategy, Avoidance Decoding, that modifies token logits by penalizing similarity to previously generated outputs, thereby encouraging more diverse multi-branch stories. This penalty adaptively balances two similarity measures: (1) Concept-level Similarity Penalty, which is prioritized in early stages to diversify initial story concepts, and (2) Narrative-level Similarity Penalty, which is increasingly emphasized later to ensure natural yet diverse plot development. Notably, our method achieves up to 2.6 times higher output diversity and reduces repetition by an average of 30% compared to strong baselines, while effectively mitigating text degeneration. Furthermore, we reveal that our method activates a broader range of neurons, demonstrating that it leverages the model's intrinsic creativity.
Authors: Anum Afzal, Juraj Vladika, Florian Matthes
Abstract: Large Language Models tend to struggle when dealing with specialized domains. While all aspects of evaluation hold importance, factuality is the most critical one. Similarly, reliable fact-checking tools and data sources are essential for hallucination mitigation. We address these issues by providing a comprehensive Fact-checking Benchmark FActBench covering four generation tasks and six state-of-the-art Large Language Models (LLMs) for the Medical domain. We use two state-of-the-art Fact-checking techniques: Chain-of-Thought (CoT) Prompting and Natural Language Inference (NLI). Our experiments show that the fact-checking scores acquired through the Unanimous Voting of both techniques correlate best with Domain Expert Evaluation.
Authors: Jaime Collado-Monta\~nez, L. Alfonso Ure\~na-L\'opez, Arturo Montejo-R\'aez
Abstract: Large Language Models offer impressive language capabilities but suffer from well-known limitations, including hallucinations, biases, privacy concerns, and high computational costs. These issues are largely driven by the combination of linguistic competence and factual memorization within a single monolithic model. This paper introduces and empirically supports the Fundamental Language Model (FLM) paradigm, which advocates for smaller, linguistically competent models that offload factual retrieval to external tools. We evaluate models ranging from 135M to 32B parameters across three dimensions: linguistic competence, external factual knowledge, and internal factual knowledge. Our findings reveal that while both linguistic competence and factual knowledge improve with scale, internal factual knowledge grows significantly faster, suggesting that model size is more closely tied to memorization than to core language ability. These results support a modular approach to language modeling, where compact, linguistically proficient models serve as the foundation for tool-augmented systems. The FLM paradigm offers a path toward more efficient, interpretable, and sustainable NLP solutions.
Authors: Katharine Kowalyshyn, Matthias Scheutz
Abstract: What if large language models could not only infer human mindsets but also expose every blind spot in team dialogue such as discrepancies in the team members' joint understanding? We present a novel, two-step framework that leverages large language models (LLMs) both as human-style annotators of team dialogues to track the team's shared mental models (SMMs) and as automated discrepancy detectors among individuals' mental states. In the first step, an LLM generates annotations by identifying SMM elements within task-oriented dialogues from the Cooperative Remote Search Task (CReST) corpus. Then, a secondary LLM compares these LLM-derived annotations and human annotations against gold-standard labels to detect and characterize divergences. We define an SMM coherence evaluation framework for this use case and apply it to six CReST dialogues, ultimately producing: (1) a dataset of human and LLM annotations; (2) a reproducible evaluation framework for SMM coherence; and (3) an empirical assessment of LLM-based discrepancy detection. Our results reveal that, although LLMs exhibit apparent coherence on straightforward natural-language annotation tasks, they systematically err in scenarios requiring spatial reasoning or disambiguation of prosodic cues.
Authors: Shihui Yang, Chengfeng Dou, Peidong Guo, Kai Lu, Qiang Ju, Fei Deng, Rihui Xin
Abstract: Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a promising framework for enhancing the reasoning capabilities of large language models. However, existing approaches such as GRPO often suffer from zero gradients. This problem arises primarily due to fixed clipping bounds for token-level probability ratios and the standardization of identical rewards, which can lead to ineffective gradient updates and underutilization of generated responses. In this work, we propose Dynamic Clipping Policy Optimization (DCPO), which introduces a dynamic clipping strategy that adaptively adjusts the clipping bounds based on token-specific prior probabilities to enhance token-level exploration, and a smooth advantage standardization technique that standardizes rewards across cumulative training steps to improve the response-level effective utilization of generated responses. DCPO achieved state-of-the-art performance on four benchmarks based on four different models. In particular, DCPO achieved an Avg@1 of 46.7 under greedy decoding and an Avg@32 of 38.8 under 32 times sampling on the AIME24 benchmark, surpassing both DAPO (36.7/31.6) and GRPO (36.7/32.1) on the Qwen2.5-Math-7B model. On the AIME25 benchmark based on Qwen2.5-14B, DCPO achieves a performance of (23.3/19.0), surpassing GRPO (13.3/10.5) and DAPO (20.0/15.3). Furthermore, DCPO achieved an average 28% improvement in the nonzero advantage over GRPO in four models, doubled the training efficiency over DAPO, and significantly reduced the token clipping ratio by an order of magnitude compared to both GRPO and DAPO, while achieving superior performance. These results highlight DCPO's effectiveness in leveraging generated data more efficiently for reinforcement learning in large language models.
Authors: Jindong Li, Yali Fu, Li Fan, Jiahong Liu, Yao Shu, Chengwei Qin, Menglin Yang, Irwin King, Rex Ying
Abstract: Large Language Models (LLMs) have demonstrated strong generalization across a wide range of tasks. Reasoning with LLMs is central to solving multi-step problems and complex decision-making. To support efficient reasoning, recent studies have shifted attention from explicit chain-of-thought prompting toward implicit reasoning, where reasoning occurs silently via latent structures without emitting intermediate textual steps. Implicit reasoning brings advantages such as lower generation cost, faster inference, and better alignment with internal computation. Although prior surveys have discussed latent representations in the context of reasoning, a dedicated and mechanism-level examination of how reasoning unfolds internally within LLMs remains absent. This survey fills that gap by introducing a taxonomy centered on execution paradigms, shifting the focus from representational forms to computational strategies. We organize existing methods into three execution paradigms based on \textbf{\textit{how and where internal computation unfolds}}: latent optimization, signal-guided control, and layer-recurrent execution. We also review structural, behavioral and representation-based evidence that supports the presence of implicit reasoning in LLMs. We further provide a structured overview of the evaluation metrics and benchmarks used in existing works to assess the effectiveness and reliability of implicit reasoning.We maintain a continuously updated project at: https://github.com/digailab/awesome-llm-implicit-reasoning.
URLs: https://github.com/digailab/awesome-llm-implicit-reasoning.
Authors: Gaurav Negi, Atul Kr. Ojha, Omnia Zayed, Paul Buitelaar
Abstract: We propose a scalable method for constructing a temporal opinion knowledge base with large language models (LLMs) as automated annotators. Despite the demonstrated utility of time-series opinion analysis of text for downstream applications such as forecasting and trend analysis, existing methodologies underexploit this potential due to the absence of temporally grounded fine-grained annotations. Our approach addresses this gap by integrating well-established opinion mining formulations into a declarative LLM annotation pipeline, enabling structured opinion extraction without manual prompt engineering. We define three data models grounded in sentiment and opinion mining literature, serving as schemas for structured representation. We perform rigorous quantitative evaluation of our pipeline using human-annotated test samples. We carry out the final annotations using two separate LLMs, and inter-annotator agreement is computed label-wise across the fine-grained opinion dimensions, analogous to human annotation protocols. The resulting knowledge base encapsulates time-aligned, structured opinions and is compatible with applications in Retrieval-Augmented Generation (RAG), temporal question answering, and timeline summarisation.
Authors: Ali Hamdi, Malak Mohamed, Rokaia Emad, Khaled Shaban
Abstract: Social telehealth has made remarkable progress in healthcare by allowing patients to post symptoms and participate in medical consultations remotely. Users frequently post symptoms on social media and online health platforms, creating a huge repository of medical data that can be leveraged for disease classification. Large language models (LLMs) such as LLAMA3 and GPT-3.5, along with transformer-based models like BERT, have demonstrated strong capabilities in processing complex medical text. In this study, we evaluate three Arabic medical text preprocessing methods such as summarization, refinement, and Named Entity Recognition (NER) before applying fine-tuned Arabic transformer models (CAMeLBERT, AraBERT, and AsafayaBERT). To enhance robustness, we adopt a majority voting ensemble that combines predictions from original and preprocessed text representations. This approach achieved the best classification accuracy of 80.56%, thus showing its effectiveness in leveraging various text representations and model predictions to improve the understanding of medical texts. To the best of our knowledge, this is the first work that integrates LLM-based preprocessing with fine-tuned Arabic transformer models and ensemble learning for disease classification in Arabic social telehealth data.
Authors: Lingzhi Shen, Xiaohao Cai, Yunfei Long, Imran Razzak, Guanming Chen, Shoaib Jameel
Abstract: Personality detection from text is commonly performed by analysing users' social media posts. However, existing methods heavily rely on large-scale annotated datasets, making it challenging to obtain high-quality personality labels. Moreover, most studies treat emotion and personality as independent variables, overlooking their interactions. In this paper, we propose a novel self-supervised framework, EmoPerso, which improves personality detection through emotion-aware modelling. EmoPerso first leverages generative mechanisms for synthetic data augmentation and rich representation learning. It then extracts pseudo-labeled emotion features and jointly optimizes them with personality prediction via multi-task learning. A cross-attention module is employed to capture fine-grained interactions between personality traits and the inferred emotional representations. To further refine relational reasoning, EmoPerso adopts a self-taught strategy to enhance the model's reasoning capabilities iteratively. Extensive experiments on two benchmark datasets demonstrate that EmoPerso surpasses state-of-the-art models. The source code is available at https://github.com/slz0925/EmoPerso.
Authors: Seyedali Mohammadi, Bhaskara Hanuma Vedula, Hemank Lamba, Edward Raff, Ponnurangam Kumaraguru, Francis Ferraro, Manas Gaur
Abstract: Do LLMs genuinely incorporate external definitions, or do they primarily rely on their parametric knowledge? To address these questions, we conduct controlled experiments across multiple explanation benchmark datasets (general and domain-specific) and label definition conditions, including expert-curated, LLM-generated, perturbed, and swapped definitions. Our results reveal that while explicit label definitions can enhance accuracy and explainability, their integration into an LLM's task-solving processes is neither guaranteed nor consistent, suggesting reliance on internalized representations in many cases. Models often default to their internal representations, particularly in general tasks, whereas domain-specific tasks benefit more from explicit definitions. These findings underscore the need for a deeper understanding of how LLMs process external knowledge alongside their pre-existing capabilities.
Authors: Ahmed Ahmed, Kevin Klyman, Yi Zeng, Sanmi Koyejo, Percy Liang
Abstract: Companies that develop foundation models publish behavioral guidelines they pledge their models will follow, but it remains unclear if models actually do so. While providers such as OpenAI, Anthropic, and Google have published detailed specifications describing both desired safety constraints and qualitative traits for their models, there has been no systematic audit of adherence to these guidelines. We introduce an automated framework that audits models against their providers specifications by parsing behavioral statements, generating targeted prompts, and using models to judge adherence. Our central focus is on three way consistency between a provider specification, its model outputs, and its own models as judges; an extension of prior two way generator validator consistency. This establishes a necessary baseline: at minimum, a foundation model should consistently satisfy the developer behavioral specifications when judged by the developer evaluator models. We apply our framework to 16 models from six developers across more than 100 behavioral statements, finding systematic inconsistencies including compliance gaps of up to 20 percent across providers.
Authors: Chenglong Wang, Yongyu Mu, Hang Zhou, Yifu Huo, Ziming Zhu, Jiali Zeng, Murun Yang, Bei Li, Tong Xiao, Xiaoyang Hao, Chunliang Zhang, Fandong Meng, Jingbo Zhu
Abstract: Significant progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs towards generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on abundant unlabeled data offers a promising direction, but existing approaches fall short of instilling explicit reasoning into reward models. To bridge this gap, we propose a self-training approach that leverages unlabeled data to elicit reward reasoning in reward models. Based on this approach, we develop GRAM-R$^2$, a generative reward model trained to produce not only preference labels but also accompanying reward rationales. GRAM-R$^2$ can serve as a foundation model for reward reasoning and can be applied to a wide range of tasks with minimal or no additional fine-tuning. It can support downstream applications such as response ranking and task-specific reward tuning. Experiments on response ranking, task adaptation, and reinforcement learning from human feedback demonstrate that GRAM-R$^2$ consistently delivers strong performance, outperforming several strong discriminative and generative baselines.
Authors: Junxi Wu, Jinpeng Wang, Zheng Liu, Bin Chen, Dongjian Hu, Hao Wu, Shu-Tao Xiu
Abstract: The rapid advancement of large language models has intensified public concerns about the potential misuse. Therefore, it is important to build trustworthy AI-generated text detection systems. Existing methods neglect stylistic modeling and mostly rely on static thresholds, which greatly limits the detection performance. In this paper, we propose the Mixture of Stylistic Experts (MoSEs) framework that enables stylistics-aware uncertainty quantification through conditional threshold estimation. MoSEs contain three core components, namely, the Stylistics Reference Repository (SRR), the Stylistics-Aware Router (SAR), and the Conditional Threshold Estimator (CTE). For input text, SRR can activate the appropriate reference data in SRR and provide them to CTE. Subsequently, CTE jointly models the linguistic statistical properties and semantic features to dynamically determine the optimal threshold. With a discrimination score, MoSEs yields prediction labels with the corresponding confidence level. Our framework achieves an average improvement 11.34% in detection performance compared to baselines. More inspiringly, MoSEs shows a more evident improvement 39.15% in the low-resource case. Our code is available at https://github.com/creator-xi/MoSEs.
Authors: Nishant Tanksale, Tanmay Kokate, Darshan Gohad, Sarvadnyaa Barate, Raviraj Joshi
Abstract: Semantic evaluation in low-resource languages remains a major challenge in NLP. While sentence transformers have shown strong performance in high-resource settings, their effectiveness in Indic languages is underexplored due to a lack of high-quality benchmarks. To bridge this gap, we introduce L3Cube-IndicHeadline-ID, a curated headline identification dataset spanning ten low-resource Indic languages: Marathi, Hindi, Tamil, Gujarati, Odia, Kannada, Malayalam, Punjabi, Telugu, Bengali and English. Each language includes 20,000 news articles paired with four headline variants: the original, a semantically similar version, a lexically similar version, and an unrelated one, designed to test fine-grained semantic understanding. The task requires selecting the correct headline from the options using article-headline similarity. We benchmark several sentence transformers, including multilingual and language-specific models, using cosine similarity. Results show that multilingual models consistently perform well, while language-specific models vary in effectiveness. Given the rising use of similarity models in Retrieval-Augmented Generation (RAG) pipelines, this dataset also serves as a valuable resource for evaluating and improving semantic understanding in such applications. Additionally, the dataset can be repurposed for multiple-choice question answering, headline classification, or other task-specific evaluations of LLMs, making it a versatile benchmark for Indic NLP. The dataset is shared publicly at https://github.com/l3cube-pune/indic-nlp
Authors: Jean-Marie Le Ray
Abstract: A pioneering rule-based mechanical translation system (precursor of modern RBMTs) was first presented in December 1929 by its inventor, Federico Pucci, who later published the full method in a book titled "Il traduttore meccanico ed il metodo per corrispondersi fra Europei conoscendo ciascuno solo la propria lingua: Parte I", in Salerno (Italy), in 1931. This study illustrates how AI breathes new life into the system of international keys and ideograms devised by Pucci to translate from/into any Romance language (at least as a first step). The methodology involves having the AIs retranslate, following Pucci's method, the two text excerpts originally translated in 1931 and clearly documented in his publication: a passage from Dante's La Vita Nuova, translated from Italian into French, and a passage from Voltaire's Zadig, translated from French into Italian. The result is notable: the two texts, translated 94 years apart using the same method--by Pucci in 1931 and by AIs in 2025--show a low average difference, with only minor variations observed. With Pucci's system thus validated, it became feasible to have the AIs reproduce the excerpts in English, Spanish, and German according to his method. The results were consistent, and Pucci--via Artificial Intelligence--was tasked with translating more modern and technical texts, thereby reviving, nearly a century later, an invention that had remained almost entirely unknown and never applied beyond its creator, now brought to wider attention and opened to possible experimentation. Such a demonstration would not only affirm Pucci's historical status but also place him among the precursors and intellectual contributors to machine translation, whose work merits examination alongside figures such as Troyanskij, Booth, and Weaver, with possible consequences for how the history of the field is understood.
Authors: Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu, Massoud Pedram
Abstract: Large language models (LLMs), despite their impressive performance across a wide range of tasks, often struggle to balance two competing objectives in open-ended text generation: fostering diversity and creativity while preserving logical coherence. Existing truncated sampling techniques, including temperature scaling, top-\$p\$ (nucleus) sampling, and min-\$p\$ sampling, aim to manage this trade-off. However, they exhibit limitations, particularly in the effective incorporation of the confidence of the model into the corresponding sampling strategy. For example, min-\$p\$ sampling relies on a single top token as a heuristic for confidence, eventually underutilizing the information of the probability distribution. Toward effective incorporation of the confidence of the model, in this paper, we present **top-H** decoding. We first establish the theoretical foundation of the interplay between creativity and coherence in truncated sampling by formulating an **entropy-constrained minimum divergence** problem. We then prove this minimization problem to be equivalent to an **entropy-constrained mass maximization** (ECMM) problem, which is NP-hard. Finally, we present top-H decoding, a computationally efficient greedy algorithm to solve the ECMM problem. Extensive empirical evaluations demonstrate that top-H outperforms the state-of-the-art (SoTA) alternative of min-\$p\$ sampling by up to **25.63%** on creative writing benchmarks, while maintaining robustness on question-answering datasets such as GPQA, GSM8K, and MT-Bench. Additionally, an *LLM-as-judge* evaluation confirms that top-H indeed produces coherent outputs even at higher temperatures, where creativity is especially critical. In summary, top-H advances SoTA in open-ended text generation and can be *easily integrated* into creative writing applications. The code is available at https://github.com/ErfanBaghaei/Top-H-Decoding.
Authors: Mayur Shirke, Amey Shembade, Pavan Thorat, Madhushri Wagh, Raviraj Joshi
Abstract: Named Entity Recognition (NER) in code-mixed text, particularly Hindi-English (Hinglish), presents unique challenges due to informal structure, transliteration, and frequent language switching. This study conducts a comparative evaluation of code-mixed fine-tuned models and non-code-mixed multilingual models, along with zero-shot generative large language models (LLMs). Specifically, we evaluate HingBERT, HingMBERT, and HingRoBERTa (trained on code-mixed data), and BERT Base Cased, IndicBERT, RoBERTa and MuRIL (trained on non-code-mixed multilingual data). We also assess the performance of Google Gemini in a zero-shot setting using a modified version of the dataset with NER tags removed. All models are tested on a benchmark Hinglish NER dataset using Precision, Recall, and F1-score. Results show that code-mixed models, particularly HingRoBERTa and HingBERT-based fine-tuned models, outperform others - including closed-source LLMs like Google Gemini - due to domain-specific pretraining. Non-code-mixed models perform reasonably but show limited adaptability. Notably, Google Gemini exhibits competitive zero-shot performance, underlining the generalization strength of modern LLMs. This study provides key insights into the effectiveness of specialized versus generalized models for code-mixed NER tasks.
Authors: Jiaming Li, Longze Chen, Ze Gong, Yukun Chen, Lu Wang, Wanwei He, Run Luo, Min Yang
Abstract: Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have empowered large language models (LLMs) to tackle challenging reasoning tasks such as mathematics and programming. RLVR leverages verifiable outcome rewards to guide policy optimization, enabling LLMs to progressively improve output quality in a grounded and reliable manner. Despite its promise, the RLVR paradigm poses significant challenges, as existing methods often suffer from sparse reward signals and unstable policy gradient updates, particularly in RL-based approaches. To address the challenges, we propose $\textbf{PACS}$, a novel RLVR framework that achieves im$\textbf{P}$licit $\textbf{A}$ctor $\textbf{C}$ritic coupling via a $\textbf{S}$upervised learning framework. By treating the outcome reward as a predictable label, we reformulate the RLVR problem into a supervised learning task over a score function parameterized by the policy model and optimized using cross-entropy loss. A detailed gradient analysis shows that this supervised formulation inherently recovers the classical policy gradient update while implicitly coupling actor and critic roles, yielding more stable and efficient training. Benchmarking on challenging mathematical reasoning tasks, PACS outperforms strong RLVR baselines, such as PPO and GRPO, achieving superior reasoning performance. For instance, PACS achieves 59.78\% at pass@256 on AIME 2025, representing improvements of 13.32 and 14.36 points over PPO and GRPO. This simple yet powerful framework offers a promising avenue for LLMs post-training with verifiable rewards. Our code and data are available as open source at https://github.com/ritzz-ai/PACS.
Authors: Evan King, Adam Sabra, Manjunath Kudlur, James Wang, Pete Warden
Abstract: We present the Flavors of Moonshine, a suite of tiny automatic speech recognition (ASR) models specialized for a range of underrepresented languages. Prevailing wisdom suggests that multilingual ASR models outperform monolingual counterparts by exploiting cross-lingual phonetic similarities. We challenge this assumption, showing that for sufficiently small models (27M parameters), training monolingual systems on a carefully balanced mix of high-quality human-labeled, pseudo-labeled, and synthetic data yields substantially superior performance. On average, our models achieve error rates 48% lower than the comparably sized Whisper Tiny model, outperform the 9x larger Whisper Small model, and in most cases match or outperform the 28x larger Whisper Medium model. These results advance the state of the art for models of this size, enabling accurate on-device ASR for languages that previously had limited support. We release Arabic, Chinese, Japanese, Korean, Ukrainian, and Vietnamese Moonshine models under a permissive open-source license.
Authors: Tianjian Li, Yiming Zhang, Ping Yu, Swarnadeep Saha, Daniel Khashabi, Jason Weston, Jack Lanchantin, Tianlu Wang
Abstract: Post-training of Large Language Models (LMs) often prioritizes accuracy and helpfulness at the expense of diversity. This creates a tension: while post-training improves response quality, it also sharpens output distributions and reduces the range of ideas, limiting the usefulness of LMs in creative and exploratory tasks such as brainstorming, storytelling, or problem solving. We address this challenge with Diversity-Aware Reinforcement Learning (DARLING), a framework that jointly optimizes for response quality and semantic diversity. At its core, DARLING introduces a learned partition function to measure diversity beyond surface-level lexical variations. This diversity signal is then combined with a quality reward during online reinforcement learning, encouraging models to generate outputs that are both high-quality and distinct. Experiments across multiple model families and sizes show that DARLING generalizes to two regimes: non-verifiable tasks (instruction following and creative writing) and verifiable tasks (competition math). On five benchmarks in the first setting, DARLING consistently outperforms quality-only RL baselines, producing outputs that are simultaneously of higher quality and novelty. In the second setting, DARLING achieves higher pass@1 (solution quality) and pass@k (solution variety). Most strikingly, explicitly optimizing for diversity catalyzes exploration in online RL, which manifests itself as higher-quality responses.
Authors: Fakhraddin Alwajih, Abdellah El Mekki, Hamdy Mubarak, Majd Hawasly, Abubakr Mohamed, Muhammad Abdul-Mageed
Abstract: Large Language Models (LLMs) inherently reflect the vast data distributions they encounter during their pre-training phase. As this data is predominantly sourced from the web, there is a high chance it will be skewed towards high-resourced languages and cultures, such as those of the West. Consequently, LLMs often exhibit a diminished understanding of certain communities, a gap that is particularly evident in their knowledge of Arabic and Islamic cultures. This issue becomes even more pronounced with increasingly under-represented topics. To address this critical challenge, we introduce PalmX 2025, the first shared task designed to benchmark the cultural competence of LLMs in these specific domains. The task is composed of two subtasks featuring multiple-choice questions (MCQs) in Modern Standard Arabic (MSA): General Arabic Culture and General Islamic Culture. These subtasks cover a wide range of topics, including traditions, food, history, religious practices, and language expressions from across 22 Arab countries. The initiative drew considerable interest, with 26 teams registering for Subtask 1 and 19 for Subtask 2, culminating in nine and six valid submissions, respectively. Our findings reveal that task-specific fine-tuning substantially boosts performance over baseline models. The top-performing systems achieved an accuracy of 72.15% on cultural questions and 84.22% on Islamic knowledge. Parameter-efficient fine-tuning emerged as the predominant and most effective approach among participants, while the utility of data augmentation was found to be domain-dependent.
Authors: Shuo Liu, Di Yao, Yan Lin, Gao Cong, Jingping Bi
Abstract: Building a general model capable of analyzing human trajectories across different geographic regions and different tasks becomes an emergent yet important problem for various applications. However, existing works suffer from the generalization problem, \ie, they are either restricted to train for specific regions or only suitable for a few tasks. Given the recent advances of multimodal large language models (MLLMs), we raise the question: can MLLMs reform current trajectory data mining and solve the problem? Nevertheless, due to the modality gap of trajectory, how to generate task-independent multimodal trajectory representations and how to adapt flexibly to different tasks remain the foundational challenges. In this paper, we propose \texttt{Traj-MLLM}}, which is the first general framework using MLLMs for trajectory data mining. By integrating multiview contexts, \texttt{Traj-MLLM}} transforms raw trajectories into interleaved image-text sequences while preserving key spatial-temporal characteristics, and directly utilizes the reasoning ability of MLLMs for trajectory analysis. Additionally, a prompt optimization method is proposed to finalize data-invariant prompts for task adaptation. Extensive experiments on four publicly available datasets show that \texttt{Traj-MLLM}} outperforms state-of-the-art baselines by $48.05\%$, $15.52\%$, $51.52\%$, $1.83\%$ on travel time estimation, mobility prediction, anomaly detection and transportation mode identification, respectively. \texttt{Traj-MLLM}} achieves these superior performances without requiring any training data or fine-tuning the MLLM backbones.
Authors: C\'edric Colas, Tracey Mills, Ben Prystawski, Michael Henry Tessler, Noah Goodman, Jacob Andreas, Joshua Tenenbaum
Abstract: The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models -- revealing how structured, language-compatible representations might enable human-machine collaborative learning.
Authors: Tai Vu
Abstract: Speech Emotion Recognition (SER) presents a significant yet persistent challenge in human-computer interaction. While deep learning has advanced spoken language processing, achieving high performance on limited datasets remains a critical hurdle. This paper confronts this issue by developing and evaluating a suite of machine learning models, including Support Vector Machines (SVMs), Long Short-Term Memory networks (LSTMs), and Convolutional Neural Networks (CNNs), for automated emotion classification in human speech. We demonstrate that by strategically employing transfer learning and innovative data augmentation techniques, our models can achieve impressive performance despite the constraints of a relatively small dataset. Our most effective model, a ResNet34 architecture, establishes a new performance benchmark on the combined RAVDESS and SAVEE datasets, attaining an accuracy of 66.7% and an F1 score of 0.631. These results underscore the substantial benefits of leveraging pre-trained models and data augmentation to overcome data scarcity, thereby paving the way for more robust and generalizable SER systems.
Authors: Tatiana Likhomanenko, Luke Carlson, Richard He Bai, Zijin Gu, Han Tran, Zakaria Aldeneh, Yizhe Zhang, Ruixiang Zhang, Huangjie Zheng, Navdeep Jaitly
Abstract: The emergence of large language models (LLMs) has transformed spoken dialog systems, yet the optimal architecture for real-time on-device voice agents remains an open question. While end-to-end approaches promise theoretical advantages, cascaded systems (CSs) continue to outperform them in language understanding tasks, despite being constrained by sequential processing latency. In this work, we introduce ChipChat, a novel low-latency CS that overcomes traditional bottlenecks through architectural innovations and streaming optimizations. Our system integrates streaming (a) conversational speech recognition with mixture-of-experts, (b) state-action augmented LLM, (c) text-to-speech synthesis, (d) neural vocoder, and (e) speaker modeling. Implemented using MLX, ChipChat achieves sub-second response latency on a Mac Studio without dedicated GPUs, while preserving user privacy through complete on-device processing. Our work shows that strategically redesigned CSs can overcome their historical latency limitations, offering a promising path forward for practical voice-based AI agents.
Authors: Laksh Patel, Neel Shanbhag
Abstract: Modern generative models risk overfitting and unintentionally memorizing rare training examples, which can be extracted by adversaries or inflate benchmark performance. We propose Generative Data Cartography (GenDataCarto), a data-centric framework that assigns each pretraining sample a difficulty score (early-epoch loss) and a memorization score (frequency of ``forget events''), then partitions examples into four quadrants to guide targeted pruning and up-/down-weighting. We prove that our memorization score lower-bounds classical influence under smoothness assumptions and that down-weighting high-memorization hotspots provably decreases the generalization gap via uniform stability bounds. Empirically, GenDataCarto reduces synthetic canary extraction success by over 40\% at just 10\% data pruning, while increasing validation perplexity by less than 0.5\%. These results demonstrate that principled data interventions can dramatically mitigate leakage with minimal cost to generative performance.
Authors: Qibin Wang, Pu Zhao, Shaohan Huang, Fangkai Yang, Lu Wang, Furu Wei, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang
Abstract: To further enhance the ability of Large Language Models (LLMs) to solve complex, multi-step reasoning problems, test-time scaling (TTS) methods have gained widespread attention. Existing approaches such as Best-of-N and majority voting are limited as their performance depends on the quality of candidate responses, making them unable to produce a correct solution when all candidates are incorrect. Introducing an additional model to select the best response also incurs significant deployment costs. To this end, we introduce Generative Self-Refinement (GSR), a novel parallel test-time scaling framework where a unified model first generates a set of candidate responses in parallel and then performs self-refinement to synthesize a new superior solution based on a prompt consisting of the problem and these candidates. However, LLMs struggle to perform refinement effectively when prompted directly. Therefore, we design a hybrid training pipeline by jointly optimizing for two complementary objectives, solving problems directly and refining candidate responses. Experimental results demonstrate that our method achieves state-of-the-art performance across five mathematical benchmarks. We further show that this learned self-refinement skill is a model-agnostic enhancement, robust across different model scales and generalizing to out-of-distribution reasoning tasks.
Authors: Ephraiem Sarabamoun
Abstract: As large language models (LLMs) take on greater roles in high-stakes decisions, alignment with human values is essential. Reliance on proprietary APIs limits reproducibility and broad participation. We study whether local open-source ensemble debates can improve alignmentoriented reasoning. Across 150 debates spanning 15 scenarios and five ensemble configurations, ensembles outperform single-model baselines on a 7-point rubric (overall: 3.48 vs. 3.13), with the largest gains in reasoning depth (+19.4%) and argument quality (+34.1%). Improvements are strongest for truthfulness (+1.25 points) and human enhancement (+0.80). We provide code, prompts, and a debate data set, providing an accessible and reproducible foundation for ensemble-based alignment evaluation.
Authors: Abdullah Abdelfattah, Mahmoud I. Khalil, Hazem Abbas
Abstract: Assessing spoken language is challenging, and quantifying pronunciation metrics for machine learning models is even harder. However, for the Holy Quran, this task is simplified by the rigorous recitation rules (tajweed) established by Muslim scholars, enabling highly effective assessment. Despite this advantage, the scarcity of high-quality annotated data remains a significant barrier. In this work, we bridge these gaps by introducing: (1) A 98% automated pipeline to produce high-quality Quranic datasets -- encompassing: Collection of recitations from expert reciters, Segmentation at pause points (waqf) using our fine-tuned wav2vec2-BERT model, Transcription of segments, Transcript verification via our novel Tasmeea algorithm; (2) 850+ hours of audio (~300K annotated utterances); (3) A novel ASR-based approach for pronunciation error detection, utilizing our custom Quran Phonetic Script (QPS) to encode Tajweed rules (unlike the IPA standard for Modern Standard Arabic). QPS uses a two-level script: (Phoneme level): Encodes Arabic letters with short/long vowels. (Sifa level): Encodes articulation characteristics of every phoneme. We further include comprehensive modeling with our novel multi-level CTC Model which achieved 0.16% average Phoneme Error Rate (PER) on the testset. We release all code, data, and models as open-source: https://obadx.github.io/prepare-quran-dataset/
Authors: Yao Fu, Runchao Li, Xianxuan Long, Haotian Yu, Xiaotian Han, Yu Yin, Pan Li
Abstract: Neural network pruning has emerged as a promising approach for deploying LLMs in low-resource scenarios while preserving downstream task performance. However, for the first time, we reveal that such pruning disrupts LLMs' internal activation features crucial for lie detection, where probing classifiers (typically small logistic regression models) trained on these features assess the truthfulness of LLM-generated statements. This discovery raises a crucial open question: how can we prune LLMs without sacrificing these critical lie detection capabilities? Our investigation further reveals that naively adjusting layer-wise pruning sparsity based on importance inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. To address this issue, we propose Truthful Pruning aligned by Layer-wise Outliers (TPLO), which places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. This preserves LLMs' original performance while retaining critical features of inner states needed for robust lie detection. Moreover, we introduce a prompting rule to enrich the TruthfulQA benchmark for better calibrating LLM pruning. Empirical results show that our approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA.
Authors: Manish Shukla
Abstract: Agentic artificial intelligence (AI) -- multi-agent systems that combine large language models with external tools and autonomous planning -- are rapidly transitioning from research laboratories into high-stakes domains. Our earlier "Basic" paper introduced a five-axis framework and proposed preliminary metrics such as goal drift and harm reduction but did not provide an algorithmic instantiation or empirical evidence. This "Advanced" sequel fills that gap. First, we revisit recent benchmarks and industrial deployments to show that technical metrics still dominate evaluations: a systematic review of 84 papers from 2023--2025 found that 83% report capability metrics while only 30% consider human-centred or economic axes [2]. Second, we formalise an Adaptive Multi-Dimensional Monitoring (AMDM) algorithm that normalises heterogeneous metrics, applies per-axis exponentially weighted moving-average thresholds and performs joint anomaly detection via the Mahalanobis distance. Third, we conduct simulations and real-world experiments. AMDM cuts anomaly-detection latency from 12.3 s to 5.6 s on simulated goal drift and reduces false-positive rates from 4.5% to 0.9% compared with static thresholds. We present a comparison table and ROC/PR curves, and we reanalyse case studies to surface missing metrics. Code, data and a reproducibility checklist accompany this paper to facilitate replication.
Authors: Linus Stuhlmann, Michael Alexander Saxer
Abstract: This study evaluates the performance of three advanced speech encoder models, Wav2Vec 2.0, XLS-R, and Whisper, in speaker identification tasks. By fine-tuning these models and analyzing their layer-wise representations using SVCCA, k-means clustering, and t-SNE visualizations, we found that Wav2Vec 2.0 and XLS-R capture speaker-specific features effectively in their early layers, with fine-tuning improving stability and performance. Whisper showed better performance in deeper layers. Additionally, we determined the optimal number of transformer layers for each model when fine-tuned for speaker identification tasks.
Authors: Ziyi Guan, Jason Chun Lok Li, Zhijian Hou, Pingping Zhang, Donglai Xu, Yuzhi Zhao, Mengyang Wu, Jinpeng Chen, Thanh-Toan Nguyen, Pengfei Xian, Wenao Ma, Shengchao Qin, Graziano Chesi, Ngai Wong
Abstract: Despite recent progress, Graphic User Interface (GUI) agents powered by Large Language Models (LLMs) struggle with complex mobile tasks due to limited app-specific knowledge. While UI Transition Graphs (UTGs) offer structured navigation representations, they are underutilized due to poor extraction and inefficient integration. We introduce KG-RAG, a Knowledge Graph-driven Retrieval-Augmented Generation framework that transforms fragmented UTGs into structured vector databases for efficient real-time retrieval. By leveraging an intent-guided LLM search method, KG-RAG generates actionable navigation paths, enhancing agent decision-making. Experiments across diverse mobile apps show that KG-RAG outperforms existing methods, achieving a 75.8% success rate (8.9% improvement over AutoDroid), 84.6% decision accuracy (8.1% improvement), and reducing average task steps from 4.5 to 4.1. Additionally, we present KG-Android-Bench and KG-Harmony-Bench, two benchmarks tailored to the Chinese mobile ecosystem for future research. Finally, KG-RAG transfers to web/desktop (+40% SR on Weibo-web; +20% on QQ Music-desktop), and a UTG cost ablation shows accuracy saturates at ~4h per complex app, enabling practical deployment trade-offs.
Authors: Li Weigang, Pedro Carvalho Brom, Lucas Ramson Siefert
Abstract: We propose a novel SuperBrain framework for collective intelligence, grounded in the co-evolution of large language models (LLMs) and human users. Unlike static prompt engineering or isolated agent simulations, our approach emphasizes a dynamic pathway from Subclass Brain to Superclass Brain: (1) A Subclass Brain arises from persistent, personalized interaction between a user and an LLM, forming a cognitive dyad with adaptive learning memory. (2) Through GA-assisted forward-backward evolution, these dyads iteratively refine prompts and task performance. (3) Multiple Subclass Brains coordinate via Swarm Intelligence, optimizing across multi-objective fitness landscapes and exchanging distilled heuristics. (4) Their standardized behaviors and cognitive signatures integrate into a Superclass Brain, an emergent meta-intelligence capable of abstraction, generalization and self-improvement. We outline the theoretical constructs, present initial implementations (e.g., UAV scheduling, KU/KI keyword filtering) and propose a registry for cross-dyad knowledge consolidation. This work provides both a conceptual foundation and an architectural roadmap toward scalable, explainable and ethically aligned collective AI.
Authors: Yuzheng Cai, Yanzhao Zhang, Dingkun Long, Mingxin Li, Pengjun Xie, Weiguo Zheng
Abstract: Text reranking models are a crucial component in modern systems like Retrieval-Augmented Generation, tasked with selecting the most relevant documents prior to generation. However, current Large Language Models (LLMs) powered rerankers often face a fundamental trade-off. On one hand, Supervised Fine-Tuning based pointwise methods that frame relevance as a binary classification task lack the necessary scoring discrimination, particularly for those built on reasoning LLMs. On the other hand, approaches designed for complex reasoning often employ powerful yet inefficient listwise formulations, rendering them impractical for low latency applications. To resolve this dilemma, we introduce ERank, a highly effective and efficient pointwise reranker built from a reasoning LLM that excels across diverse relevance scenarios. We propose a novel two-stage training pipeline that begins with Supervised Fine-Tuning (SFT). In this stage, we move beyond binary labels and train the model generatively to output fine grained integer scores, which significantly enhances relevance discrimination. The model is then further refined using Reinforcement Learning (RL) with a novel, listwise derived reward. This technique instills global ranking awareness into the efficient pointwise architecture. We evaluate the ERank reranker on the BRIGHT, FollowIR, TREC DL, and BEIR benchmarks, demonstrating superior effectiveness and robustness compared to existing approaches. On the reasoning-intensive BRIGHT benchmark, our ERank-4B achieves an nDCG@10 of 38.7, while a larger 32B variant reaches a state of the art nDCG@10 of 40.2.
Authors: Lu Han, Mengyan Li, Jiping Qiang, Zhi Su
Abstract: Heterogeneous data, which encompass both numerical financial variables and textual records, present substantial challenges for credit monitoring. To address this issue, we propose Advanced Spectral Clustering (ASC), a method that integrates financial and textual similarities through an optimized weight parameter and selects eigenvectors using a novel eigenvalue-silhouette optimization approach. Evaluated on a dataset comprising 1,428 small and medium-sized enterprises (SMEs), ASC achieves a Silhouette score that is 18% higher than that of a single-type data baseline method. Furthermore, the resulting clusters offer actionable insights; for instance, 51% of low-risk firms are found to include the term 'social recruitment' in their textual records. The robustness of ASC is confirmed across multiple clustering algorithms, including k-means, k-medians, and k-medoids, with {\Delta}Intra/Inter < 0.13 and {\Delta}Silhouette Coefficient < 0.02. By bridging spectral clustering theory with heterogeneous data applications, ASC enables the identification of meaningful clusters, such as recruitment-focused SMEs exhibiting a 30% lower default risk, thereby supporting more targeted and effective credit interventions.
Authors: Albert Sadowski, Jaros{\l}aw A. Chudziak
Abstract: Legal reasoning requires both precise interpretation of statutory language and consistent application of complex rules, presenting significant challenges for AI systems. This paper introduces a modular multi-agent framework that decomposes legal reasoning into distinct knowledge acquisition and application stages. In the first stage, specialized agents extract legal concepts and formalize rules to create verifiable intermediate representations of statutes. The second stage applies this knowledge to specific cases through three steps: analyzing queries to map case facts onto the ontology schema, performing symbolic inference to derive logically entailed conclusions, and generating final answers using a programmatic implementation that operationalizes the ontological knowledge. This bridging of natural language understanding with symbolic reasoning provides explicit and verifiable inspection points, significantly enhancing transparency compared to end-to-end approaches. Evaluation on statutory tax calculation tasks demonstrates substantial improvements, with foundational models achieving 76.4\% accuracy compared to 18.8\% baseline performance, effectively narrowing the performance gap between reasoning and foundational models. These findings suggest that modular architectures with formalized knowledge representations can make sophisticated legal reasoning more accessible through computationally efficient models while enhancing consistency and explainability in AI legal reasoning, establishing a foundation for future research into more transparent, trustworthy, and effective AI systems for legal domain.
Authors: Ziqi Wang, Boqin Yuan
Abstract: We present L-MARS (Legal Multi-Agent Workflow with Orchestrated Reasoning and Agentic Search), a system that reduces hallucination and uncertainty in legal question answering through coordinated multi-agent reasoning and retrieval. Unlike single-pass retrieval-augmented generation (RAG), L-MARS decomposes queries into subproblems, issues targeted searches across heterogeneous sources (Serper web, local RAG, CourtListener case law), and employs a Judge Agent to verify sufficiency, jurisdiction, and temporal validity before answer synthesis. This iterative reasoning-search-verification loop maintains coherence, filters noisy evidence, and grounds answers in authoritative law. We evaluated L-MARS on LegalSearchQA, a new benchmark of 200 up-to-date multiple choice legal questions in 2025. Results show that L-MARS substantially improves factual accuracy, reduces uncertainty, and achieves higher preference scores from both human experts and LLM-based judges. Our work demonstrates that multi-agent reasoning with agentic search offers a scalable and reproducible blueprint for deploying LLMs in high-stakes domains requiring precise legal retrieval and deliberation.
Authors: Lee Hyun, Sohee Yoon, Jinwoo Park, Sue In Chae, Seongeon Park, Jooyeon Ahn, Yebin Jung, Youjung Chung, Hogeun Chang, Myeonginn Kang, Jina Kim, Ho-Gyeong Kim, Myeonghun Jeong
Abstract: AI-driven materials discovery that couples automated experimentation with algorithmic decision-making requires process aware recipe to property predictors that are accurate, calibrated, and physically admissible. We approach this as a reasoning problem with large reasoning models (LRMs). To instill reasoning capability into language models, we curate reasoning traces from a teacher model to train a student model. However, most training pipelines select reasoning traces using binary correctness or learned preference signals that poorly reflect physical admissibility. We introduce Physics-aware Rejection Sampling (PaRS), a training-time trace selection scheme that favors traces consistent with fundamental physics and numerically close to targets, with lightweight halting to control compute. We instantiate our framework with a large student model fine-tuned on traces synthesized by a larger teacher model, and evaluate under matched token budgets against various rejection sampling baselines. Our method improves accuracy and calibration, reduces physics-violation rates, and lowers sampling cost relative to baselines. These results indicate that modest, domain-aware constraints combined with trace-level selection provide a practical path toward reliable, efficient LRMs for process-aware property prediction and closed-loop materials design.
Authors: Zonghai Yao, Talha Chafekar, Junda Wang, Shuo Han, Feiyun Ouyang, Junhui Qian, Lingxi Li, Hong Yu
Abstract: Real-world adoption of closed-loop insulin delivery systems (CLIDS) in type 1 diabetes remains low, driven not by technical failure, but by diverse behavioral, psychosocial, and social barriers. We introduce ChatCLIDS, the first benchmark to rigorously evaluate LLM-driven persuasive dialogue for health behavior change. Our framework features a library of expert-validated virtual patients, each with clinically grounded, heterogeneous profiles and realistic adoption barriers, and simulates multi-turn interactions with nurse agents equipped with a diverse set of evidence-based persuasive strategies. ChatCLIDS uniquely supports longitudinal counseling and adversarial social influence scenarios, enabling robust, multi-dimensional evaluation. Our findings reveal that while larger and more reflective LLMs adapt strategies over time, all models struggle to overcome resistance, especially under realistic social pressure. These results highlight critical limitations of current LLMs for behavior change, and offer a high-fidelity, scalable testbed for advancing trustworthy persuasive AI in healthcare and beyond.
Authors: Aman Sharma, Saeed Najafi, Parsa Farinneya, Benyamin Jamialahmadi, Marzieh S. Tahaei, Yuhe Fan, Mehdi Rezagholizadeh, Boxing Chen, Aref Jafari
Abstract: Transformers achieve state-of-the-art results across many tasks, but their uniform application of quadratic self-attention to every token at every layer makes them computationally expensive. We introduce DTRNet (Dynamic Token Routing Network), an improved Transformer architecture that allows tokens to dynamically skip the quadratic cost of cross-token mixing while still receiving lightweight linear updates. By preserving the MLP module and reducing the attention cost for most tokens to linear, DTRNet ensures that every token is explicitly updated while significantly lowering overall computation. This design offers an efficient and effective alternative to standard dense attention. Once trained, DTRNet blocks routes only ~10% of tokens through attention at each layer while maintaining performance comparable to a full Transformer. It consistently outperforms routing-based layer skipping methods such as MoD and D-LLM in both accuracy and memory at matched FLOPs, while routing fewer tokens to full attention. Its efficiency gains, scales with sequence length, offering significant reduction in FLOPs for long-context inputs. By decoupling token updates from attention mixing, DTRNet substantially reduces the quadratic share of computation, providing a simple, efficient, and scalable alternative to Transformers.
Authors: Zifeng Ding, Shenyang Huang, Zeyu Cao, Emma Kondrup, Zachary Yang, Xingyue Huang, Yuan Sui, Zhangdie Yuan, Yuqicheng Zhu, Xianglong Hu, Yuan He, Farimah Poursafaei, Michael Bronstein, Andreas Vlachos
Abstract: Forecasting future links is a central task in temporal graph (TG) reasoning, requiring models to leverage historical interactions to predict upcoming ones. Traditional neural approaches, such as temporal graph neural networks, achieve strong performance but lack explainability and cannot be applied to unseen graphs without retraining. Recent studies have begun to explore using large language models (LLMs) for graph reasoning, but most of them are constrained to static graphs or small synthetic TGs and lack the evaluation of the quality of reasoning traces generated by LLMs. In this work, we present Reasoning-Enhanced Learning for Temporal Graphs (ReaL-TG), a reinforcement learning framework that fine-tunes LLMs to perform explainable link forecasting on real-world TGs. ReaL-TG uses outcome-based reward to encourage models to self-explore reasoning strategies from graph structure and to produce explanations that directly justify their predictions. To enable evaluation on LLM-generated reasoning traces, we propose a new evaluation protocol combining ranking metrics with an LLM-as-a-Judge system that assesses both the quality of reasoning and the impact of hallucinations. Experiments with ReaL-TG-4B, obtained by fine-tuning Qwen3-4B under our framework, show that it outperforms much larger frontier LLMs, including GPT-5 mini, on ranking metrics, while producing high-quality explanations confirmed by both the LLM judge and human evaluation.
Authors: Deepak Bastola, Woohyeok Choi
Abstract: Legal documents pose unique challenges for text classification due to their domain-specific language and often limited labeled data. This paper proposes a hybrid approach for classifying legal texts by combining unsupervised topic and graph embeddings with a supervised model. We employ Top2Vec to learn semantic document embeddings and automatically discover latent topics, and Node2Vec to capture structural relationships via a bipartite graph of legal documents. The embeddings are combined and clustered using KMeans, yielding coherent groupings of documents. Our computations on a legal document dataset demonstrate that the combined Top2Vec+Node2Vec approach improves clustering quality over text-only or graph-only embeddings. We conduct a sensitivity analysis of hyperparameters, such as the number of clusters and the dimensionality of the embeddings, and demonstrate that our method achieves competitive performance against baseline Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF) models. Key findings indicate that while the pipeline presents an innovative approach to unsupervised legal document analysis by combining semantic topic modeling with graph embedding techniques, its efficacy is contingent upon the quality of initial topic generation and the representational power of the chosen embedding models for specialized legal language. Strategic recommendations include the exploration of domain-specific embeddings, more comprehensive hyperparameter tuning for Node2Vec, dynamic determination of cluster numbers, and robust human-in-the-loop validation processes to enhance legal relevance and trustworthiness. The pipeline demonstrates potential for exploratory legal data analysis and as a precursor to supervised learning tasks but requires further refinement and domain-specific adaptation for practical legal applications.
Authors: Runjia Zeng, Guangyan Sun, Qifan Wang, Tong Geng, Sohail Dianat, Xiaotian Han, Raghuveer Rao, Xueling Zhang, Cheng Han, Lifu Huang, Dongfang Liu
Abstract: Considering deep neural networks as manifold mappers, the pretrain-then-fine-tune paradigm can be interpreted as a two-stage process: pretrain establishes a broad knowledge base, and fine-tune adjusts the model parameters to activate specific neural pathways to align with the target manifold. Although prior fine-tuning approaches demonstrate success, their rigid parameter space limits their ability to dynamically activate appropriate neural pathways, rendering them ill-equipped to adapt flexibly to the diverse and evolving data distributions. In light of this view, we propose a novel approach, Mixture of Expert Prompt Tuning (MEPT), as an effective and efficient manifold-mapping framework. MEPT leverages the Mixture of Experts architecture by integrating multiple prompt experts to adaptively learn diverse and non-stationary data distributions. Empirical evaluations demonstrate that MEPT outperforms several state-of-the-art parameter efficient baselines on SuperGLUE, achieving notable improvements in mean accuracy (e.g., 1.94%) while significantly reducing activated prompts by 79.25%. The effectiveness of MEPT is further supported by theoretical insights from manifold learning and validated through neural activation pathway visualization results. Our code is avaliable at https://github.com/runtsang/MEPT.
Authors: Aishni Parab, Hongjing Lu, Ying Nian Wu, Sumit Gulwani
Abstract: Inductive reasoning enables humans to infer abstract rules from limited examples and apply them to novel situations. In this work, we compare an LLM-based hypothesis search framework with direct program generation approaches on few-shot rule induction tasks. Our findings show that hypothesis search achieves performance comparable to humans, while direct program generation falls notably behind. An error analysis reveals key bottlenecks in hypothesis generation and suggests directions for advancing program induction methods. Overall, this paper underscores the potential of LLM-based hypothesis search for modeling inductive reasoning and the challenges in building more efficient systems.
Authors: Matte Lim, Catherine Yeh, Martin Wattenberg, Fernanda Vi\'egas, Panagiotis Michalatos
Abstract: Many real-world datasets -- from an artist's body of work to a person's social media history -- exhibit meaningful semantic changes over time that are difficult to capture with existing dimensionality reduction methods. To address this gap, we introduce a visualization technique that combines force-based projection and streaming clustering methods to build a spatial-temporal map of embeddings. Applying this technique, we create Chronotome, a tool for interactively exploring evolving themes in time-based data -- in real time. We demonstrate the utility of our approach through use cases on text and image data, showing how it offers a new lens for understanding the aesthetics and semantics of temporal datasets.
Authors: Jaewoo Ahn, Junseo Kim, Heeseung Yun, Jaehyeon Son, Dongmin Park, Jaewoong Cho, Gunhee Kim
Abstract: GUI agents powered by LLMs show promise in interacting with diverse digital environments. Among these, video games offer a valuable testbed due to their varied interfaces, with adventure games posing additional challenges through complex, narrative-driven interactions. Existing game benchmarks, however, lack diversity and rarely evaluate agents on completing entire storylines. To address this, we introduce FlashAdventure, a benchmark of 34 Flash-based adventure games designed to test full story arc completion and tackle the observation-behavior gap: the challenge of remembering and acting on earlier gameplay information. We also propose CUA-as-a-Judge, an automated gameplay evaluator, and COAST, an agentic framework leveraging long-term clue memory to better plan and solve sequential tasks. Experiments show current GUI agents struggle with full story arcs, while COAST improves milestone completion by bridging the observation-behavior gap. Nonetheless, a marked discrepancy between humans and best-performing agents warrants continued research efforts to narrow this divide.
Authors: Dongfu Jiang, Yi Lu, Zhuofeng Li, Zhiheng Lyu, Ping Nie, Haozhe Wang, Alex Su, Hui Chen, Kai Zou, Chao Du, Tianyu Pang, Wenhu Chen
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2$\times$ speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research. Our code is open-sourced at https://github.com/TIGER-AI-Lab/verl-tool.
Authors: Hyunjong Ok, Jaeho Lee
Abstract: Recent advances in multimodal large language models (MLLMs) have led to much progress in video understanding tasks. To avoid the heavy computational cost of processing all frames, these models typically rely on keyframe sampling methods guided by vision-language encoders (\textit{e.g.,} SigLIP). However, it remains unclear whether such encoders can truly identify the most informative frames. In this work, we provide several empirical pieces of evidence revealing that popular vision encoders critically suffer from their limited capability to identify where the MLLM should look inside the video to handle the given textual query appropriately. Our findings suggest that the development of better keyframe identification techniques may be necessary for efficient video MLLMs.
Authors: Wonduk Seo, Taesub Shin, Hyunjin An, Dokyun Kim, Seunghyun Lee
Abstract: Identifying whether two product listings refer to the same Stock Keeping Unit (SKU) is a persistent challenge in ecommerce, especially when explicit identifiers are missing and product names vary widely across platforms. Rule based heuristics and keyword similarity often misclassify products by overlooking subtle distinctions in brand, specification, or bundle configuration. To overcome these limitations, we propose Question to Knowledge (Q2K), a multi agent framework that leverages Large Language Models (LLMs) for reliable SKU mapping. Q2K integrates: (1) a Reasoning Agent that generates targeted disambiguation questions, (2) a Knowledge Agent that resolves them via focused web searches, and (3) a Deduplication Agent that reuses validated reasoning traces to reduce redundancy and ensure consistency. A human in the loop mechanism further refines uncertain cases. Experiments on real world consumer goods datasets show that Q2K surpasses strong baselines, achieving higher accuracy and robustness in difficult scenarios such as bundle identification and brand origin disambiguation. By reusing retrieved reasoning instead of issuing repeated searches, Q2K balances accuracy with efficiency, offering a scalable and interpretable solution for product integration.
Authors: Mattia Tritto, Giuseppe Farano, Dario Di Palma, Gaetano Rossiello, Fedelucio Narducci, Dharmashankar Subramanian, Tommaso Di Noia
Abstract: Text-to-SQL, the task of translating natural language questions into SQL queries, has significantly advanced with the introduction of Large Language Models (LLMs), broadening database accessibility for a wide range of users. Despite substantial progress in generating valid SQL, current LLMs still struggle with complex queries that require precise alignment between user intent and the database schema. To mitigate this, test-time strategies such as Best-of-N (BoN) and Majority Voting (Maj) are often employed, based on the assumption that LLMs can generate correct answers but may require multiple attempts. However, these methods rely on surface-level heuristics, selecting either the syntactically correct query through execution-based BoN (ex-BoN) or the most frequently generated query with Maj. Recently, Outcome Reward Models (ORMs), which assign utility scores to generated outputs based on semantic correctness, have emerged as a promising approach for better aligning model predictions with user intent. Nevertheless, their application to Text-to-SQL remains largely underexplored. In this work, we evaluate ORMs as an effective heuristic for BoN, compare them with ex-BoN and Maj, and introduce a framework for training ORMs for the Text-to-SQL task. We evaluate our ORMs on the BIRD and SPIDER benchmarks, finetuning various open-source LLMs, including the Qwen2, Granite3, and Llama3 model families. Our results show that ORMs outperform ex-BoN and Maj, achieving execution accuracy gains of +4.33% (BIRD) and +2.10% (Spider) over ex-BoN, and +2.91% (BIRD) and +0.93% (Spider) over Maj. We further demonstrate that finetuning models already aligned with SQL generation, such as OmniSQL, yields superior ORM performance. Additionally, we observe that ORMs achieve competitive results on simple queries and benefit more from an increased number of candidates compared to ex-BoN and Maj.
Authors: Xinyu Tang, Zhenduo Zhang, Yurou Liu, Wayne Xin Zhao, Zujie Wen, Zhiqiang Zhang, Jun Zhou
Abstract: Recent advances in large reasoning models have leveraged reinforcement learning with verifiable rewards (RLVR) to improve reasoning capabilities. However, scaling these methods typically requires extensive rollout computation and large datasets, leading to high training costs and low data efficiency. To mitigate this issue, we propose DEPO, a Data-Efficient Policy Optimization pipeline that combines optimized strategies for both offline and online data selection. In the offline phase, we curate a high-quality subset of training samples based on diversity, influence, and appropriate difficulty. During online RLVR training, we introduce a sample-level explorability metric to dynamically filter samples with low exploration potential, thereby reducing substantial rollout computational costs. Furthermore, we incorporate a replay mechanism for under-explored samples to ensure adequate training, which enhances the model's final convergence performance. Experiments across five reasoning benchmarks show that DEPO consistently outperforms existing methods in both offline and online data selection scenarios. Notably, using only 20% of the training data, our approach achieves a 1.85 times speed-up on AIME24 and a 1.66 times speed-up on AIME25 compared to GRPO trained on the full dataset.
Authors: Qianrui Zhou, Hua Xu, Yifan Wang, Xinzhi Dong, Hanlei Zhang
Abstract: Understanding human intents from multimodal signals is critical for analyzing human behaviors and enhancing human-machine interactions in real-world scenarios. However, existing methods exhibit limitations in their modality-level reliance, constraining relational reasoning over fine-grained semantics for complex intent understanding. This paper proposes a novel LLM-Guided Semantic Relational Reasoning (LGSRR) method, which harnesses the expansive knowledge of large language models (LLMs) to establish semantic foundations that boost smaller models' relational reasoning performance. Specifically, an LLM-based strategy is proposed to extract fine-grained semantics as guidance for subsequent reasoning, driven by a shallow-to-deep Chain-of-Thought (CoT) that autonomously uncovers, describes, and ranks semantic cues by their importance without relying on manually defined priors. Besides, we formally model three fundamental types of semantic relations grounded in logical principles and analyze their nuanced interplay to enable more effective relational reasoning. Extensive experiments on multimodal intent and dialogue act recognition tasks demonstrate LGSRR's superiority over state-of-the-art methods, with consistent performance gains across diverse semantic understanding scenarios. The complete data and code are available at https://github.com/thuiar/LGSRR.
Authors: Joonyong Park, Daisuke Saito, Nobuaki Minematsu
Abstract: This study presents a novel approach to voice synthesis that can substitute the traditional grapheme-to-phoneme (G2P) conversion by using a deep learning-based model that generates discrete tokens directly from speech. Utilizing a pre-trained voice SSL model, we train a T5 encoder to produce pseudo-language labels from mixed-script texts (e.g., containing Kanji and Kana). This method eliminates the need for manual phonetic transcription, reducing costs and enhancing scalability, especially for large non-transcribed audio datasets. Our model matches the performance of conventional G2P-based text-to-speech systems and is capable of synthesizing speech that retains natural linguistic and paralinguistic features, such as accents and intonations.
Authors: Ali Abouzeid, Bilal Elbouardi, Mohamed Maged, Shady Shehata
Abstract: Speech emotion recognition is vital for human-computer interaction, particularly for low-resource languages like Arabic, which face challenges due to limited data and research. We introduce ArabEmoNet, a lightweight architecture designed to overcome these limitations and deliver state-of-the-art performance. Unlike previous systems relying on discrete MFCC features and 1D convolutions, which miss nuanced spectro-temporal patterns, ArabEmoNet uses Mel spectrograms processed through 2D convolutions, preserving critical emotional cues often lost in traditional methods. While recent models favor large-scale architectures with millions of parameters, ArabEmoNet achieves superior results with just 1 million parameters, 90 times smaller than HuBERT base and 74 times smaller than Whisper. This efficiency makes it ideal for resource-constrained environments. ArabEmoNet advances Arabic speech emotion recognition, offering exceptional performance and accessibility for real-world applications.
Authors: Yujing Wang, Yiren Chen, Huoran Li, Chunxu Xu, Yuchong Luo, Xianghui Mao, Cong Li, Lun Du, Chunyang Ma, Qiqi Jiang, Yin Wang, Fan Gao, Wenting Mo, Pei Wen, Shantanu Kumar, Taejin Park, Yiwei Song, Vijay Rajaram, Tao Cheng, Sonu Durgia, Pranam Kolari
Abstract: As global e-commerce platforms continue to expand, companies are entering new markets where they encounter cold-start challenges due to limited human labels and user behaviors. In this paper, we share our experiences in Coupang to provide a competitive cold-start performance of relevance matching for emerging e-commerce markets. Specifically, we present a Cold-Start Relevance Matching (CSRM) framework, utilizing a multilingual Large Language Model (LLM) to address three challenges: (1) activating cross-lingual transfer learning abilities of LLMs through machine translation tasks; (2) enhancing query understanding and incorporating e-commerce knowledge by retrieval-based query augmentation; (3) mitigating the impact of training label errors through a multi-round self-distillation training strategy. Our experiments demonstrate the effectiveness of CSRM-LLM and the proposed techniques, resulting in successful real-world deployment and significant online gains, with a 45.8% reduction in defect ratio and a 0.866% uplift in session purchase rate.
Authors: Zetong Zhou, Dongping Chen, Zixian Ma, Zhihan Hu, Mingyang Fu, Sinan Wang, Yao Wan, Zhou Zhao, Ranjay Krishna
Abstract: Visual reasoning, a cornerstone of human intelligence, encompasses complex perceptual and logical processes essential for solving diverse visual problems. While advances in computer vision have produced powerful models for various perceptual tasks, leveraging these for general visual reasoning remains challenging. Prior work demonstrates that augmenting LLMs with vision models via supervised finetuning improves performance, but faces key limitations such as expensive data generation, reliance on careful data filtering, and poor generalization. To address these issues, we propose ReVPT to enhance multi-modal LLMs' abilities to reason about and use visual tools through reinforcement learning. We introduce a novel RL algorithm based on GRPO, designed to train models to reason with a suite of four visual tools. Through extensive experiments, we show that our method achieves state-of-the-art performance on several perception-heavy benchmarks, including SAT, CV-Bench, BLINK and MMStar, significantly outperforming the supervised and text-based RL finetuning baselines. Notably, Our ReVPT-3B and ReVPT-7B outperform the instruct models by 9.03% and 9.44% on CV-Bench. Finally, we bring to the community new insights on RL-based visual tool-usage through extensive ablations. Our code is available at https://github.com/ls-kelvin/REVPT.
Authors: Rui Zhao, Vladyslav Melnychuk, Jun Zhao, Jesse Wright, Nigel Shadbolt
Abstract: In modern times, people have numerous online accounts, but they rarely read the Terms of Service or Privacy Policy of those sites, despite claiming otherwise, due to the practical difficulty in comprehending them. The mist of data privacy practices forms a major barrier for user-centred Web approaches, and for data sharing and reusing in an agentic world. Existing research proposed methods for using formal languages and reasoning for verifying the compliance of a specified policy, as a potential cure for ignoring privacy policies. However, a critical gap remains in the creation or acquisition of such formal policies at scale. We present a semantic-centric approach for using state-of-the-art large language models (LLM), to automatically identify key information about privacy practices from privacy policies, and construct $\mathit{Pr}^2\mathit{Graph}$, knowledge graph with grounding from Data Privacy Vocabulary (DPV) for privacy practices, to support downstream tasks. Along with the pipeline, the $\mathit{Pr}^2\mathit{Graph}$ for the top-100 popular websites is also released as a public resource, by using the pipeline for analysis. We also demonstrate how the $\mathit{Pr}^2\mathit{Graph}$ can be used to support downstream tasks by constructing formal policy representations such as Open Digital Right Language (ODRL) or perennial semantic Data Terms of Use (psDToU). To evaluate the technology capability, we enriched the Policy-IE dataset by employing legal experts to create custom annotations. We benchmarked the performance of different large language models for our pipeline and verified their capabilities. Overall, they shed light on the possibility of large-scale analysis of online services' privacy practices, as a promising direction to audit the Web and the Internet. We release all datasets and source code as public resources to facilitate reuse and improvement.
Authors: Mingxuan Cui, Yilan Jiang, Duo Zhou, Cheng Qian, Yuji Zhang, Qiong Wang
Abstract: Drug shortages pose critical risks to patient care and healthcare systems worldwide, yet the effectiveness of regulatory interventions remains poorly understood due to fundamental information asymmetries in pharmaceutical supply chains. We present \textbf{ShortageSim}, the first Large Language Model (LLM)-based multi-agent simulation framework that captures the complex, strategic interactions between drug manufacturers, institutional buyers, and regulatory agencies in response to shortage alerts. Unlike traditional game-theoretic models that assume perfect rationality and complete information, \textbf{ShortageSim} leverages LLMs to simulate bounded-rational decision-making under uncertainty. Through a sequential production game spanning multiple quarters, we model how FDA announcements, both reactive alerts about existing shortages and proactive warnings about potential disruptions, propagate through the supply chain and influence capacity investment and procurement decisions. Our experiments on historical shortage events reveal that \textbf{ShortageSim} reduces the resolution-lag percentage for discontinued-disclosed cases by 83\%, bringing simulated durations more aligned to ground truth than the zero-shot baseline. We open-source \textbf{ShortageSim} and a dataset of 2,925 FDA shortage events at https://github.com/Lemutisme/Sortage_Management, providing a novel computational framework for designing and testing interventions in complex, information-scarce supply chains.
Authors: Zhenyuan Chen, Chenxi Wang, Ningyu Zhang, Feng Zhang
Abstract: Remote sensing is critical for disaster monitoring, yet existing datasets lack temporal image pairs and detailed textual annotations. While single-snapshot imagery dominates current resources, it fails to capture dynamic disaster impacts over time. To address this gap, we introduce the Remote Sensing Change Caption (RSCC) dataset, a large-scale benchmark comprising 62,315 pre-/post-disaster image pairs (spanning earthquakes, floods, wildfires, and more) paired with rich, human-like change captions. By bridging the temporal and semantic divide in remote sensing data, RSCC enables robust training and evaluation of vision-language models for disaster-aware bi-temporal understanding. Our results highlight RSCC's ability to facilitate detailed disaster-related analysis, paving the way for more accurate, interpretable, and scalable vision-language applications in remote sensing. Code and dataset are available at https://github.com/Bili-Sakura/RSCC.
Authors: Ranjie Duan, Jiexi Liu, Xiaojun Jia, Shiji Zhao, Ruoxi Cheng, Fengxiang Wang, Cheng Wei, Yong Xie, Chang Liu, Defeng Li, Yinpeng Dong, Yichi Zhang, Yuefeng Chen, Chongwen Wang, Xingjun Ma, Xingxing Wei, Yang Liu, Hang Su, Jun Zhu, Xinfeng Li, Yitong Sun, Jie Zhang, Jinzhao Hu, Sha Xu, Yitong Yang, Jialing Tao, Hui Xue
Abstract: Large language models (LLMs) typically deploy safety mechanisms to prevent harmful content generation. Most current approaches focus narrowly on risks posed by malicious actors, often framing risks as adversarial events and relying on defensive refusals. However, in real-world settings, risks also come from non-malicious users seeking help while under psychological distress (e.g., self-harm intentions). In such cases, the model's response can strongly influence the user's next actions. Simple refusals may lead them to repeat, escalate, or move to unsafe platforms, creating worse outcomes. We introduce Constructive Safety Alignment (CSA), a human-centric paradigm that protects against malicious misuse while actively guiding vulnerable users toward safe and helpful results. Implemented in Oyster-I (Oy1), CSA combines game-theoretic anticipation of user reactions, fine-grained risk boundary discovery, and interpretable reasoning control, turning safety into a trust-building process. Oy1 achieves state-of-the-art safety among open models while retaining high general capabilities. On our Constructive Benchmark, it shows strong constructive engagement, close to GPT-5, and unmatched robustness on the Strata-Sword jailbreak dataset, nearing GPT-o1 levels. By shifting from refusal-first to guidance-first safety, CSA redefines the model-user relationship, aiming for systems that are not just safe, but meaningfully helpful. We release Oy1, code, and the benchmark to support responsible, user-centered AI.
Authors: Ruijia Li, Yuan-Hao Jiang, Jiatong Wang, Bo Jiang
Abstract: Heuristic and scaffolded teacher-student dialogues are widely regarded as critical for fostering students' higher-order thinking and deep learning. However, large language models (LLMs) currently face challenges in generating pedagogically rich interactions. This study systematically investigates the structural and behavioral differences between AI-simulated and authentic human tutoring dialogues. We conducted a quantitative comparison using an Initiation-Response-Feedback (IRF) coding scheme and Epistemic Network Analysis (ENA). The results show that human dialogues are significantly superior to their AI counterparts in utterance length, as well as in questioning (I-Q) and general feedback (F-F) behaviors. More importantly, ENA results reveal a fundamental divergence in interactional patterns: human dialogues are more cognitively guided and diverse, centered around a "question-factual response-feedback" teaching loop that clearly reflects pedagogical guidance and student-driven thinking; in contrast, simulated dialogues exhibit a pattern of structural simplification and behavioral convergence, revolving around an "explanation-simplistic response" loop that is essentially a simple information transfer between the teacher and student. These findings illuminate key limitations in current AI-generated tutoring and provide empirical guidance for designing and evaluating more pedagogically effective generative educational dialogue systems.
Authors: Jonathn Chang, Leonard Piff, Suvadip Sana, Jasmine X. Li, Lionel Levine
Abstract: Aligning AI with human values is a pressing unsolved problem. To address the lack of quantitative metrics for value alignment, we propose EigenBench: a black-box method for comparatively benchmarking language models' values. Given an ensemble of models, a constitution describing a value system, and a dataset of scenarios, our method returns a vector of scores quantifying each model's alignment to the given constitution. To produce these scores, each model judges the outputs of other models across many scenarios, and these judgments are aggregated with EigenTrust (Kamvar et al, 2003), yielding scores that reflect a weighted-average judgment of the whole ensemble. EigenBench uses no ground truth labels, as it is designed to quantify traits for which reasonable judges may disagree on the correct label. Using prompted personas, we test whether EigenBench scores are more sensitive to the model or the prompt: we find that most of the variance is explained by the prompt, but a small residual quantifies the disposition of the model itself.
Authors: Nirmalya Thakur, Madeline D Hartel, Lane Michael Boden, Dallas Enriquez, Boston Joyner Ricks
Abstract: This work investigated about 10,000 COVID-19-related YouTube videos published between January 2023 and October 2024 to evaluate how temporal, lexical, linguistic, and structural factors influenced engagement during the late pandemic period. Publishing activity showed consistent weekday effects: in the first window, average views peaked on Mondays at 92,658; in the second, on Wednesdays at 115,479; and in the third, on Fridays at 84,874, reflecting a shift in audience attention toward mid- and late week. Lexical analysis of video titles revealed recurring high-frequency keywords related to COVID-19 and YouTube features, including COVID, coronavirus, shorts, and live. Frequency analysis revealed sharp spikes, with COVID appearing in 799 video titles in August 2024, while engagement analysis showed that videos titled with shorts attracted very high views, peaking at 2.16 million average views per video in June 2023. Analysis of sentiment of video descriptions in English showed weak correlation with views in the raw data (Pearson r = 0.0154, p = 0.2987), but stronger correlations emerged once outliers were addressed, with Spearman r = 0.110 (p < 0.001) and Pearson r = 0.0925 (p < 0.001). Category-level analysis of video durations revealed contrasting outcomes: long videos focusing on people and blogs averaged 209,114 views, short entertainment videos averaged 288,675 views, and medium-to-long news and politics videos averaged 51,309 and 59,226 views, respectively. These results demonstrate that engagement patterns of COVID-19-related videos on YouTube during the late pandemic followed distinct characteristics driven by publishing schedules, title vocabulary, topics, and genre-specific duration effects.
Authors: Refat Othman, Diaeddin Rimawi, Bruno Rossi, Barbara Russo
Abstract: In the domain of security, vulnerabilities frequently remain undetected even after their exploitation. In this work, vulnerabilities refer to publicly disclosed flaws documented in Common Vulnerabilities and Exposures (CVE) reports. Establishing a connection between attacks and vulnerabilities is essential for enabling timely incident response, as it provides defenders with immediate, actionable insights. However, manually mapping attacks to CVEs is infeasible, thereby motivating the need for automation. This paper evaluates 14 state-of-the-art (SOTA) sentence transformers for automatically identifying vulnerabilities from textual descriptions of attacks. Our results demonstrate that the multi-qa-mpnet-base-dot-v1 (MMPNet) model achieves superior classification performance when using attack Technique descriptions, with an F1-score of 89.0, precision of 84.0, and recall of 94.7. Furthermore, it was observed that, on average, 56% of the vulnerabilities identified by the MMPNet model are also represented within the CVE repository in conjunction with an attack, while 61% of the vulnerabilities detected by the model correspond to those cataloged in the CVE repository. A manual inspection of the results revealed the existence of 275 predicted links that were not documented in the MITRE repositories. Consequently, the automation of linking attack techniques to vulnerabilities not only enhances the detection and response capabilities related to software security incidents but also diminishes the duration during which vulnerabilities remain exploitable, thereby contributing to the development of more secure systems.
Authors: Sharjeel Tahir, Judith Johnson, Jumana Abu-Khalaf, Syed Afaq Ali Shah
Abstract: A prevalent shortfall among current empathic AI systems is their inability to recognize when verbal expressions may not fully reflect underlying emotional states. This is because the existing datasets, used for the training of these systems, focus on surface-level emotion recognition without addressing the complex verbal-visual incongruence (mismatch) patterns useful for empathic understanding. In this paper, we present E-THER, the first Person-Centered Therapy-grounded multimodal dataset with multidimensional annotations for verbal-visual incongruence detection, enabling training of AI systems that develop genuine rather than performative empathic capabilities. The annotations included in the dataset are drawn from humanistic approach, i.e., identifying verbal-visual emotional misalignment in client-counsellor interactions - forming a framework for training and evaluating AI on empathy tasks. Additional engagement scores provide behavioral annotations for research applications. Notable gains in empathic and therapeutic conversational qualities are observed in state-of-the-art vision-language models (VLMs), such as IDEFICS and VideoLLAVA, using evaluation metrics grounded in empathic and therapeutic principles. Empirical findings indicate that our incongruence-trained models outperform general-purpose models in critical traits, such as sustaining therapeutic engagement, minimizing artificial or exaggerated linguistic patterns, and maintaining fidelity to PCT theoretical framework.
Authors: Nils Hoehing, Mayug Maniparambil, Ellen Rushe, Noel E. O'Connor, Anthony Ventresque
Abstract: We propose RocketScience, an open-source contrastive VLM benchmark that tests for spatial relation understanding. It is comprised of entirely new real-world image-text pairs covering mostly relative spatial understanding and the order of objects. The benchmark is designed to be very easy for humans and hard for the current generation of VLMs, and this is empirically verified. Our results show a striking lack of spatial relation understanding in open source and frontier commercial VLMs and a surprisingly high performance of reasoning models. Additionally, we perform a disentanglement analysis to separate the contributions of object localization and spatial reasoning in chain-of-thought-based models and find that the performance on the benchmark is bottlenecked by spatial reasoning and not object localization capabilities. We release the dataset with a CC-BY-4.0 license and make the evaluation code available at: https://github.com/nilshoehing/rocketscience
Authors: Luis Felipe Chary, Miguel Arjona Ramirez
Abstract: We present a neural speech codec that challenges the need for complex residual vector quantization (RVQ) stacks by introducing a simpler, single-stage quantization approach. Our method operates directly on the mel-spectrogram, treating it as a 2D data and quantizing non-overlapping 4x4 patches into a single, shared codebook. This patchwise design simplifies the architecture, enables low-latency streaming, and yields a discrete latent grid. To ensure high-fidelity synthesis, we employ a late-stage adversarial fine-tuning for the VQ-VAE and train a HiFi-GAN vocoder from scratch on the codec's reconstructed spectrograms. Operating at approximately 7.5 kbits/s for 16 kHz speech, our system was evaluated against several state-of-the-art neural codecs using objective metrics such as STOI, PESQ, MCD, and ViSQOL. The results demonstrate that our simplified, non-residual architecture achieves competitive perceptual quality and intelligibility, validating it as an effective and open foundation for future low-latency codec designs.
Authors: Haji Gul, Abdul Ghani Naim, Ajaz Ahmad Bhat
Abstract: Accurate estimation of dataset complexity is crucial for evaluating and comparing link prediction models for knowledge graphs (KGs). The Cumulative Spectral Gradient (CSG) metric derived from probabilistic divergence between classes within a spectral clustering framework was proposed as a dataset complexity measure that (1) naturally scales with the number of classes and (2) correlates strongly with downstream classification performance. In this work, we rigorously assess CSG behavior on standard knowledge graph link prediction benchmarks a multi class tail prediction task, using two key parameters governing its computation, M, the number of Monte Carlo sampled points per class, and K, the number of nearest neighbors in the embedding space. Contrary to the original claims, we find that (1) CSG is highly sensitive to the choice of K and therefore does not inherently scale with the number of target classes, and (2) CSG values exhibit weak or no correlation with established performance metrics such as mean reciprocal rank (MRR). Through experiments on FB15k 237, WN18RR, and other standard datasets, we demonstrate that CSG purported stability and generalization predictive power break down in link prediction settings. Our results highlight the need for more robust, classifier agnostic complexity measures in KG link prediction evaluation.
Authors: Jingru Fan, Yufan Dang, Jingyao Wu, Huatao Li, Runde Yang, Xiyuan Yang, Yuheng Wang, Zhong Zhang, Yaxi Lu, Yankai Lin, Zhiyuan Liu, Dahai Li, Chen Qian
Abstract: With the raid evolution of large language models and multimodal foundation models, the mobile-agent landscape has proliferated without converging on the fundamental challenges. This paper identifies four core problems that must be solved for mobile agents to deliver practical, scalable impact: (1) generalization across tasks, modalities, apps, and devices; (2) accuracy, specifically precise on-screen interaction and click targeting; (3) long-horizon capability for sustained, multi-step goals; and (4) efficiency, specifically high-performance runtime on resource-constrained devices. We present AppCopilot, a multimodal, multi-agent, general-purpose on-device assistant that operates across applications and constitutes a full-stack, closed-loop system from data to deployment. AppCopilot operationalizes this position through an end-to-end autonomous pipeline spanning data collection, training, deployment, high-quality and efficient inference, and mobile application development. At the model layer, it integrates multimodal foundation models with robust Chinese-English support. At the reasoning and control layer, it combines chain-of-thought reasoning, hierarchical task planning and decomposition, and multi-agent collaboration. At the execution layer, it enables user personalization and experiential adaptation, voice interaction, function calling, cross-app and cross-device orchestration, and comprehensive mobile app support. The system design incorporates profiling-driven optimization for latency, memory, and energy across heterogeneous hardware. Empirically, AppCopilot achieves significant improvements along all four dimensions: stronger generalization, higher-precision on-screen actions, more reliable long-horizon task completion, and faster, more resource-efficient runtime.
Authors: Yiqun Yao, Xiang Li, Xin Jiang, Xuezhi Fang, Naitong Yu, Wenjia Ma, Aixin Sun, Yequan Wang
Abstract: Full-duplex dialog models are designed to listen and speak simultaneously with rapid responses to fast-changing user input. Among existing approaches, native full-duplex models merges different channels (e.g. listen and speak) in a single time step, overcoming the high response latency inherent to time-division multiplexing time-division multiplexing (TDM) alternatives. Yet, a key challenge remains: aligning textual monologues with audio streams that operate at different bitrates. The prevailing solution relies on word-level alignment, but this can degrade the language ability of large pre-trained models. Moreover, it requires highly accurate timestamps for every token, which introduces cascading errors and increases pre-processing costs. In this paper, we propose textual monologues in continuous tokens sequence, namely "natural" monologues, which mimics humanoid cognitive behavior in dialogs. For temporal alignment, we alternate the position of the natural monologue - leading or trailing the audio - across different training stages. This "dual" training paradigm proves highly effective in building FLM-Audio, our 7B spoken dialog model that demonstrates superior responsiveness, duplexity, and chatting experiences, as confirmed by experimental results.
Authors: Haoming Wang, Haoyang Zou, Huatong Song, Jiazhan Feng, Junjie Fang, Junting Lu, Longxiang Liu, Qinyu Luo, Shihao Liang, Shijue Huang, Wanjun Zhong, Yining Ye, Yujia Qin, Yuwen Xiong, Yuxin Song, Zhiyong Wu, Bo Li, Chen Dun, Chong Liu, Fuxing Leng, Hanbin Wang, Hao Yu, Haobin Chen, Hongyi Guo, Jing Su, Jingjia Huang, Kai Shen, Kaiyu Shi, Lin Yan, Peiyao Zhao, Pengfei Liu, Qinghao Ye, Renjie Zheng, Wayne Xin Zhao, Wen Heng, Wenhao Huang, Wenqian Wang, Xiaobo Qin, Yi Lin, Youbin Wu, Zehui Chen, Zihao Wang, Baoquan Zhong, Xinchun Zhang, Xujing Li, Yuanfan Li, Zhongkai Zhao, Chengquan Jiang, Faming Wu, Haotian Zhou, Jinlin Pang, Li Han, Qianli Ma, Siyao Liu, Songhua Cai, Wenqi Fu, Xin Liu, Zhi Zhang, Bo Zhou, Guoliang Li, Jiajun Shi, Jiale Yang, Jie Tang, Li Li, Taoran Lu, Woyu Lin, Xiaokang Tong, Xinyao Li, Yichi Zhang, Yu Miao, Zhengxuan Jiang, Zili Li, Ziyuan Zhao, Chenxin Li, Dehua Ma, Feng Lin, Ge Zhang, Haihua Yang, Hangyu Guo, Hongda Zhu, Jiaheng Liu, Junda Du, Kai Cai, Kuanye Li, Lichen Yuan, Meilan Han, Minchao Wang, Shuyue Guo, Tianhao Cheng, Xiaobo Ma, Xiaojun Xiao, Xiaolong Huang, Xinjie Chen, Yidi Du, Yilin Chen, Yiwen Wang, Zhaojian Li, Zhenzhu Yang, Zhiyuan Zeng, Chaolin Jin, Chen Li, Hao Chen, Haoli Chen, Jian Chen, Qinghao Zhao, Guang Shi
Abstract: The development of autonomous agents for graphical user interfaces (GUIs) presents major challenges in artificial intelligence. While recent advances in native agent models have shown promise by unifying perception, reasoning, action, and memory through end-to-end learning, open problems remain in data scalability, multi-turn reinforcement learning (RL), the limitations of GUI-only operation, and environment stability. In this technical report, we present UI-TARS-2, a native GUI-centered agent model that addresses these challenges through a systematic training methodology: a data flywheel for scalable data generation, a stabilized multi-turn RL framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts. Empirical evaluation demonstrates that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5. On GUI benchmarks, it reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld, outperforming strong baselines such as Claude and OpenAI agents. In game environments, it attains a mean normalized score of 59.8 across a 15-game suite-roughly 60% of human-level performance-and remains competitive with frontier proprietary models (e.g., OpenAI o3) on LMGame-Bench. Additionally, the model can generalize to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. Detailed analyses of training dynamics further provide insights into achieving stability and efficiency in large-scale agent RL. These results underscore UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios.
Authors: Guibin Zhang, Hejia Geng, Xiaohang Yu, Zhenfei Yin, Zaibin Zhang, Zelin Tan, Heng Zhou, Zhongzhi Li, Xiangyuan Xue, Yijiang Li, Yifan Zhou, Yang Chen, Chen Zhang, Yutao Fan, Zihu Wang, Songtao Huang, Yue Liao, Hongru Wang, Mengyue Yang, Heng Ji, Michael Littman, Jun Wang, Shuicheng Yan, Philip Torr, Lei Bai
Abstract: The emergence of agentic reinforcement learning (Agentic RL) marks a paradigm shift from conventional reinforcement learning applied to large language models (LLM RL), reframing LLMs from passive sequence generators into autonomous, decision-making agents embedded in complex, dynamic worlds. This survey formalizes this conceptual shift by contrasting the degenerate single-step Markov Decision Processes (MDPs) of LLM-RL with the temporally extended, partially observable Markov decision processes (POMDPs) that define Agentic RL. Building on this foundation, we propose a comprehensive twofold taxonomy: one organized around core agentic capabilities, including planning, tool use, memory, reasoning, self-improvement, and perception, and the other around their applications across diverse task domains. Central to our thesis is that reinforcement learning serves as the critical mechanism for transforming these capabilities from static, heuristic modules into adaptive, robust agentic behavior. To support and accelerate future research, we consolidate the landscape of open-source environments, benchmarks, and frameworks into a practical compendium. By synthesizing over five hundred recent works, this survey charts the contours of this rapidly evolving field and highlights the opportunities and challenges that will shape the development of scalable, general-purpose AI agents.
Authors: Monte Hoover, Vatsal Baherwani, Neel Jain, Khalid Saifullah, Joseph Vincent, Chirag Jain, Melissa Kazemi Rad, C. Bayan Bruss, Ashwinee Panda, Tom Goldstein
Abstract: Guardian models are used to supervise and moderate the outputs of user-facing chatbots, enforcing guardrails and detecting bad behaviors. Standard guardian models like LlamaGuard detect predefined, static categories of harms. We propose dynamic guardian models that evaluate text based on user-defined policies, making them useful for different application domains that are not addressed by standard guardian models. Our dynamic guardian models can be used for fast detection of policy violations or with chain-of-thought reasoning that articulates and justifies the model outputs. Our dynamic guardian models match static models in detection accuracy for static harm categories while identifying violations of free-form policies with accuracy comparable to frontier reasoning models in a fraction of the time.
Authors: Wenchuan Mu
Abstract: Capturing the similarities between human language units is crucial for explaining how humans associate different objects, and therefore its computation has received extensive attention, research, and applications. With the ever-increasing amount of information around us, calculating similarity becomes increasingly complex, especially in many cases, such as legal or medical affairs, measuring similarity requires extra care and precision, as small acts within a language unit can have significant real-world effects. My research goal in this thesis is to develop regression models that account for similarities between language units in a more refined way. Computation of similarity has come a long way, but approaches to debugging the measures are often based on continually fitting human judgment values. To this end, my goal is to develop an algorithm that precisely catches loopholes in a similarity calculation. Furthermore, most methods have vague definitions of the similarities they compute and are often difficult to interpret. The proposed framework addresses both shortcomings. It constantly improves the model through catching different loopholes. In addition, every refinement of the model provides a reasonable explanation. The regression model introduced in this thesis is called progressively refined similarity computation, which combines attack testing with adversarial training. The similarity regression model of this thesis achieves state-of-the-art performance in handling edge cases.
Authors: Qisong Li, Ji Lin, Sijia Wei, Neng Liu
Abstract: Recent studies on knowledge graph embedding focus on mapping entities and relations into low-dimensional vector spaces. While most existing models primarily exploit structural information, knowledge graphs also contain rich contextual and textual information that can enhance embedding effectiveness. In this work, we propose a novel model that integrates both contextual and textual signals into entity and relation embeddings through a graph convolutional network. To better utilize context, we introduce two metrics: confidence, computed via a rule-based method, and relatedness, derived from textual representations. These metrics enable more precise weighting of contextual information during embedding learning. Extensive experiments on two widely used benchmark datasets demonstrate the effectiveness of our approach, showing consistent improvements over strong baselines.
Authors: Sijia Wei, Wenwen Zhang, Qisong Li, Jiang Zhao
Abstract: In this paper, we propose a novel method for question answering over knowledge graphs based on graph-to-segment mapping, designed to improve the understanding of natural language questions. Our approach is grounded in semantic parsing, a key technique for interpreting question utterances. The main challenges arise from handling implicit entities and relations, as well as complex constraints such as temporal conditions, ordinality, and aggregation within the context of a knowledge graph. To address these issues, our framework integrates both rule-based and neural methods to parse and construct accurate, comprehensive semantic segment sequences. These sequences are then assembled into semantic query graphs, providing precise representations of question utterances. We formulate question semantic parsing as a sequence generation task, employing an encoder-decoder neural network to map natural language questions into semantic segments. Furthermore, to enhance the identification of implicit entities and relations, we incorporate a graph neural network that leverages knowledge graph context to enrich question representations. Experimental evaluations on two benchmark datasets demonstrate the effectiveness and superior performance of our model in semantic parsing for knowledge graph question answering.
Authors: Adriano Belisario, Scott A. Hale, Luc Rocher
Abstract: Gun violence is a pressing human rights issue that affects nearly every dimension of the social fabric, from healthcare and education to psychology and the economy. Reliable data on firearm events is paramount to developing more effective public policy and emergency responses. However, the lack of comprehensive databases and the risks of in-person surveys prevent human rights organizations from collecting needed data in most countries. Here, we partner with a Brazilian human rights organization to conduct a systematic evaluation of language models to assist with monitoring real-world firearm events from social media data. We propose a fine-tuned BERT-based model trained on Twitter (now X) texts to distinguish gun violence reports from ordinary Portuguese texts. We then incorporate our model into a web application and test it in a live intervention. We study and interview Brazilian analysts who continuously check social media texts to identify new gun violence events. Qualitative assessments show that our solution helped all analysts use their time more efficiently and expanded their search capacities. Quantitative assessments show that the use of our model was associated with analysts having further interactions with online users reporting gun violence. Our findings suggest that human-centered interventions using language models can help support the work of human rights organizations.
Authors: Ariel Rosenfeld, Teddy Lazebnik
Abstract: Large Language Models (LLMs) are capable of generating text that is similar to or surpasses human quality. However, it is unclear whether LLMs tend to exhibit distinctive linguistic styles akin to how human authors do. Through a comprehensive linguistic analysis, we compare the vocabulary, Part-Of-Speech (POS) distribution, dependency distribution, and sentiment of texts generated by three of the most popular LLMS today (GPT-3.5, GPT-4, and Bard) to diverse inputs. The results point to significant linguistic variations which, in turn, enable us to attribute a given text to its LLM origin with a favorable 88\% accuracy using a simple off-the-shelf classification model. Theoretical and practical implications of this intriguing finding are discussed.
Authors: Dayeon Ki, Marine Carpuat
Abstract: Machine Translation (MT) remains one of the last NLP tasks where large language models (LLMs) have not yet replaced dedicated supervised systems. This work exploits the complementary strengths of LLMs and supervised MT by guiding LLMs to automatically post-edit MT with external feedback on its quality, derived from Multidimensional Quality Metric (MQM) annotations. Working with LLaMA-2 models, we consider prompting strategies varying the nature of feedback provided and then fine-tune the LLM to improve its ability to exploit the provided guidance. Through experiments on Chinese-English, English-German, and English-Russian MQM data, we demonstrate that prompting LLMs to post-edit MT improves TER, BLEU and COMET scores, although the benefits of fine-grained feedback are not clear. Fine-tuning helps integrate fine-grained feedback more effectively and further improves translation quality based on both automatic and human evaluation.
Authors: Xiang Geng, Ming Zhu, Jiahuan Li, Zhejian Lai, Wei Zou, Shuaijie She, Jiaxin Guo, Xiaofeng Zhao, Yinglu Li, Yuang Li, Chang Su, Yanqing Zhao, Xinglin Lyu, Min Zhang, Jiajun Chen, Hao Yang, Shujian Huang
Abstract: The scarcity of non-English data limits the development of non-English large language models (LLMs). Transforming English-centric LLMs to non-English has been identified as an effective and resource-efficient method. Previous works start from base LLMs and perform knowledge distillation (KD) with data generated by stronger LLMs, e.g. GPT-4. Compared to base LLMs, chat LLMs are further optimized for advanced abilities, e.g. multi-turn conversation and human preference alignment, and thus more powerful in both helpfulness and safety. However, transforming a chat LLM involves two critical issues: (1) How can we effectively transfer advanced abilities without their supervised data? (2) How can we prevent the original knowledge from catastrophic forgetting during transformation? We target these issues by introducing a simple framework called TransLLM. For the first issue, TransLLM divides the transfer problem into some common sub-tasks with the translation chain-of-thought, which uses the translation as the bridge between English and non-English step-by-step. We further enhance the performance of sub-tasks with publicly available data. For the second issue, we propose a method comprising two synergistic components: low-rank adaptation for training to maintain the original LLM parameters, and recovery KD, which utilizes data generated by the chat LLM itself to recover the original knowledge from the frozen parameters. In the experiments, we transform the LLaMA-2-chat-7B to the Thai language. Our method, using only single-turn data, outperforms strong baselines and ChatGPT on multi-turn benchmark MT-bench. Furthermore, our method, without safety data, rejects more harmful queries of safety benchmark AdvBench than both ChatGPT and GPT-4. Code is available at https://github.com/hy5468/TransLLM.
Authors: Yihuai Hong, Lei Yu, Haiqin Yang, Shauli Ravfogel, Mor Geva
Abstract: The task of "unlearning" certain concepts in large language models (LLMs) has attracted immense attention recently, due to its importance in mitigating undesirable model behaviours, such as the generation of harmful, private, or incorrect information. Current protocols to evaluate unlearning methods largely rely on behavioral tests, without monitoring the presence of unlearned knowledge within the model's parameters. This residual knowledge can be adversarially exploited to recover the erased information post-unlearning. We argue that unlearning should also be evaluated internally, by considering changes in the parametric knowledge traces of the unlearned concepts. To this end, we propose a general evaluation methodology that leverages vocabulary projections to inspect concepts encoded in model parameters. We use this approach to localize "concept vectors" - parameter vectors that encode concrete concepts - and construct ConceptVectors, a benchmark dataset containing hundreds of common concepts and their parametric knowledge traces within two open-source LLMs. Evaluation on ConceptVectors shows that existing unlearning methods minimally impact concept vectors and mostly suppress them during inference, while directly ablating these vectors demonstrably removes the associated knowledge and significantly reduces the model's susceptibility to adversarial manipulation. Our results highlight limitations in behavioral-based unlearning evaluations and call for future work to include parameter-based evaluations. To support this, we release our code and benchmark at https://github.com/yihuaihong/ConceptVectors.
Authors: Jiyang Qiu, Xinbei Ma, Zhuosheng Zhang, Hai Zhao, Yun Li, Qianren Wang
Abstract: Large language models (LLMs) have exhibited remarkable versatility and adaptability, while their widespread adoption across various applications also raises critical safety concerns. This paper focuses on the impact of backdoored LLMs. Traditional backdoor injection methods are primarily limited to yes-or-no discriminative tasks, leading users to underestimate the potential risks of backdoored LLMs. Given the inherently generative nature of LLMs, this paper reveals that a generative backdoor injected into LLMs can expose the true safety risks in their applications. We propose an editing-based generative backdoor, named MEGen, aiming to expand the backdoor to generative tasks in a unified format of any text-to any text, leading to natural generations with a specific intention. Experiments show that MEGen achieves a high attack success rate by adjusting only a small set of local parameters with few-shot samples. Notably, we show that the backdoored model, when triggered, can freely output pre-set dangerous information while completing downstream tasks. Our work highlights that MEGen enables backdoors in LLMs to exhibit generative capabilities, causing potential safety risks by altering the generative style. The code is available at https://github.com/MonoQ-hub/MEGen.
Authors: Yifan Zhang, Yang Yuan, Andrew Chi-Chih Yao
Abstract: Large Language Models (LLMs) excel at many tasks but often falter on complex problems that require structured, multi-step reasoning. We introduce the Diagram of Thought (DoT), a new framework that enables a single LLM to build and navigate a mental map of its reasoning. Instead of thinking in a straight line, the model constructs a dynamic diagram of ideas, where it can propose different lines of thought, critique its own steps, and synthesize validated insights into a final conclusion. This entire process is self-contained within the model, making it highly efficient by avoiding the complex external controllers or search algorithms required by other methods. To ensure the reliability of this process, we ground DoT in a rigorous mathematical framework from category theory. This foundation guarantees that the way the model combines information is logical, consistent, and robust, regardless of the order in which ideas were explored. The result is a more powerful and transparent reasoning process that produces a fully auditable, step-by-step trace of the LLM's thinking, bridging the gap between fluent language and formal reasoning.
Authors: Dayeon Ki, Cheonbok Park, Hyunjoong Kim
Abstract: Accurately aligning contextual representations in cross-lingual sentence embeddings is key for effective parallel data mining. A common strategy for achieving this alignment involves disentangling semantics and language in sentence embeddings derived from multilingual pre-trained models. However, we discover that current disentangled representation learning methods suffer from semantic leakage - a term we introduce to describe when a substantial amount of language-specific information is unintentionally leaked into semantic representations. This hinders the effective disentanglement of semantic and language representations, making it difficult to retrieve embeddings that distinctively represent the meaning of the sentence. To address this challenge, we propose a novel training objective, ORthogonAlity Constraint LEarning (ORACLE), tailored to enforce orthogonality between semantic and language embeddings. ORACLE builds upon two components: intra-class clustering and inter-class separation. Through experiments on cross-lingual retrieval and semantic textual similarity tasks, we demonstrate that training with the ORACLE objective effectively reduces semantic leakage and enhances semantic alignment within the embedding space.
Authors: Daniele Gambetta, Gizem Gezici, Fosca Giannotti, Dino Pedreschi, Alistair Knott, Luca Pappalardo
Abstract: As synthetic content increasingly infiltrates the web, generative AI models may be retrained on their own outputs: a process termed "autophagy". This leads to model collapse: a progressive loss of performance and diversity across generations. Recent studies have examined the emergence of model collapse across various generative AI models and data types, and have proposed mitigation strategies that rely on incorporating human-authored content. However, current characterizations of model collapse remain limited, and existing mitigation methods assume reliable knowledge of whether training data is human-authored or AI-generated. In this paper, we address these gaps by introducing new measures that characterise collapse directly from a model's next-token probability distributions, rather than from properties of AI-generated text. Using these measures, we show that the degree of collapse depends on the complexity of the initial training set, as well as on the extent of autophagy. Our experiments prompt a new suggestion: that model collapse occurs when a model trains on data that does not "surprise" it. We express this hypothesis in terms of the well-known Free Energy Principle in cognitive science. Building on this insight, we propose a practical mitigation strategy: filtering training items by high surplexity, maximising the surprise of the model. Unlike existing methods, this approach does not require distinguishing between human- and AI-generated data. Experiments across datasets and models demonstrate that our strategy is at least as effective as human-data baselines, and even more effective in reducing distributional skewedness. Our results provide a richer understanding of model collapse and point toward more resilient approaches for training generative AI systems in environments increasingly saturated with synthetic data.
Authors: Xiangci Li, Jessica Ouyang
Abstract: Retrieval-augmented generation (RAG) is a powerful method for enhancing natural language generation by integrating external knowledge into a model's output. While prior work has demonstrated the importance of improving knowledge retrieval for boosting generation quality, the role of knowledge selection, a.k.a. reranking or filtering, remains less clear. This paper empirically analyzes how knowledge selection influences downstream generation performance in RAG systems. By simulating different retrieval and selection conditions through a controlled mixture of gold and distractor knowledge, we assess the impact of these factors on generation outcomes. Our findings indicate that the downstream generator model's capability, as well as the complexity of the task and dataset, significantly influence the impact of knowledge selection on the overall RAG system performance. In typical scenarios, improving the knowledge recall score is key to enhancing generation outcomes, with the knowledge selector providing limited benefit when a strong generator model is used on clear, well-defined tasks. For weaker generator models or more ambiguous tasks and datasets, the knowledge F1 score becomes a critical factor, and the knowledge selector plays a more prominent role in improving overall performance.
Authors: Wei He, Zhiheng Xi, Wanxu Zhao, Xiaoran Fan, Yiwen Ding, Zifei Shan, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract: Solving complex chart Q&A tasks requires advanced visual reasoning abilities in multimodal large language models (MLLMs), including recognizing key information from visual inputs and conducting reasoning over it. While fine-tuning MLLMs for reasoning is critical, collecting and annotating charts and questions is expensive, hard to scale, and often results in low-quality annotations. To address this, we propose Code-as-Intermediary Translation (CIT), a cost-effective, efficient and scalable data synthesis method for distilling visual reasoning abilities from LLMs to MLLMs. The code serves as an intermediary that translates visual chart representations into textual representations, enabling language models to understand cross-modal information and generate reasoning chains accordingly. In this way, we can employ text-based synthesizing techniques to expand chart-plotting code and generate high-quality Q&A pairs for training models. This produces ReachQA, a dataset containing 3k reasoning-intensive charts and 20k Q&A pairs to enhance both recognition and reasoning abilities of MLLMs. Experiments show that models fine-tuned with ReachQA not only perform well on chart-related tasks but also show performance gains on general reasoning benchmarks. The code and dataset are publicly available at https://github.com/hewei2001/ReachQA.
Authors: John Chen, Alexandros Lotsos, Sihan Cheng, Caiyi Wang, Lexie Zhao, Jessica Hullman, Bruce Sherin, Uri Wilensky, Michael Horn
Abstract: Qualitative analysis is critical to understanding human datasets in many social science disciplines. A central method in this process is inductive coding, where researchers identify and interpret codes directly from the datasets themselves. Yet, this exploratory approach poses challenges for meeting methodological expectations (such as ``depth'' and ``variation''), especially as researchers increasingly adopt Generative AI (GAI) for support. Ground-truth-based metrics are insufficient because they contradict the exploratory nature of inductive coding, while manual evaluation can be labor-intensive. This paper presents a theory-informed computational method for measuring inductive coding results from humans and GAI. Our method first merges individual codebooks using an LLM-enriched algorithm. It measures each coder's contribution against the merged result using four novel metrics: Coverage, Overlap, Novelty, and Divergence. Through two experiments on a human-coded online conversation dataset, we 1) reveal the merging algorithm's impact on metrics; 2) validate the metrics' stability and robustness across multiple runs and different LLMs; and 3) showcase the metrics' ability to diagnose coding issues, such as excessive or irrelevant (hallucinated) codes. Our work provides a reliable pathway for ensuring methodological rigor in human-AI qualitative analysis.
Authors: Junhua Liu, Yong Keat Tan, Bin Fu, Kwan Hui Lim
Abstract: In conversational AI systems, a critical challenge in training effective multi-turn intent classification models lies in the generation of large-scale, domain-specific, multilingual dialogue datasets. In this paper, we introduce Chain-of-Intent, a novel framework that integrates Hidden Markov Models (HMMs) with Large Language Models (LLMs) to generate intent-driven, context-aware dialogues through self-play. Our method first extracts domain-specific intent transition patterns from real-world e-commerce chat logs, which guide the modeling of turn-level dynamics and intent sequences. LLMs are then employed to parameterize the emission probabilities of HMMs, enabling the generation of natural, coherent utterances aligned with predicted intents and dialogue context. We also propose MINT-CL, a multi-task contrastive learning framework for multi-turn intent classification, which improves performance while reducing dependence on large-scale annotated datasets. Empirical results demonstrate that our approach outperforms competitive baselines in dialogue generation quality and classification accuracy, particularly in multilingual settings. To facilitate future research, we release MINT-E, a comprehensive, multilingual, intent-aware multi-turn dialogue corpus derived from the e-commerce domain\footnote{The reproduced source code and dataset are available at https://github.com/junhua/chain-of-intent.
Authors: Seungone Kim, Juyoung Suk, Xiang Yue, Vijay Viswanathan, Seongyun Lee, Yizhong Wang, Kiril Gashteovski, Carolin Lawrence, Sean Welleck, Graham Neubig
Abstract: Given the increasing use of synthetic data in language model (LM) post-training, an LM's ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs' data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs' data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM's data generation ability doesn't necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality-including response quality, perplexity, and instruction difficulty-collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness.
Authors: Tianyang Zhong, Zhenyuan Yang, Zhengliang Liu, Ruidong Zhang, Yiheng Liu, Haiyang Sun, Yi Pan, Yiwei Li, Yifan Zhou, Hanqi Jiang, Junhao Chen, Tianming Liu
Abstract: Low-resource languages serve as invaluable repositories of human history, embodying cultural evolution and intellectual diversity. Despite their significance, these languages face critical challenges, including data scarcity and technological limitations, which hinder their comprehensive study and preservation. Recent advancements in large language models (LLMs) offer transformative opportunities for addressing these challenges, enabling innovative methodologies in linguistic, historical, and cultural research. This study systematically evaluates the applications of LLMs in low-resource language research, encompassing linguistic variation, historical documentation, cultural expressions, and literary analysis. By analyzing technical frameworks, current methodologies, and ethical considerations, this paper identifies key challenges such as data accessibility, model adaptability, and cultural sensitivity. Given the cultural, historical, and linguistic richness inherent in low-resource languages, this work emphasizes interdisciplinary collaboration and the development of customized models as promising avenues for advancing research in this domain. By underscoring the potential of integrating artificial intelligence with the humanities to preserve and study humanity's linguistic and cultural heritage, this study fosters global efforts towards safeguarding intellectual diversity.
Authors: Jing Liu, Abdellah Fourtassi
Abstract: LLMs can generate human-like dialogues, yet their ability to simulate early child-adult interactions remains largely unexplored. In this paper, we examined how effectively LLMs can capture the distinctive features of child-caregiver language in interaction, using both static and interactive benchmarking methods. We found that state-of-the-art LLMs like Llama 3 and GPT-4o can approximate child-caregiver dialogues at the word and utterance level, but they struggle to reproduce the child and caregiver's discursive patterns, exaggerate alignment, and fail to reach the level of diversity shown by humans. The broader goal of this work is to initiate the development of a comprehensive benchmark for LLMs in child-oriented applications.
Authors: Ildik\'o Pil\'an, Benet Manzanares-Salor, David S\'anchez, Pierre Lison
Abstract: Text sanitization aims to rewrite parts of a document to prevent disclosure of personal information. The central challenge of text sanitization is to strike a balance between privacy protection (avoiding the leakage of personal information) and utility preservation (retaining as much as possible of the document's original content). To this end, we introduce a novel text sanitization method based on generalizations, that is, broader but still informative terms that subsume the semantic content of the original text spans. The approach relies on the use of instruction-tuned large language models (LLMs) and is divided into two stages. Given a document including text spans expressing personally identifiable information (PII), the LLM is first applied to obtain truth-preserving replacement candidates for each text span and rank those according to their abstraction level. Those candidates are then evaluated for their ability to protect privacy by conducting inference attacks with the LLM. Finally, the system selects the most informative replacement candidate shown to be resistant to those attacks. This two-stage process produces replacements that effectively balance privacy and utility. We also present novel metrics to evaluate these two aspects without needing to manually annotate documents. Results on the Text Anonymization Benchmark show that the proposed approach, implemented with Mistral 7B Instruct, leads to enhanced utility, with only a marginal (< 1 p.p.) increase in re-identification risk compared to fully suppressing the original spans. Furthermore, our approach is shown to be more truth-preserving than existing methods such as Microsoft Presidio's synthetic replacements.
Authors: Chenxi Fu, Xiaoyi Wang, Zaijiang Man, Caimei Yang
Abstract: As recursion has been underlying any linguistic work for the last 60 years, the acquisition of recursive structures by children during language learning has become a focal point of inquiry. This study delves into the developmental trajectory of Mandarin-speaking children's acquisition of recursive possessives and locatives, assessing the impact of structural diversity on language acquisition. The research contrasts the comprehension of two-level recursive structures among children aged 3 to 7 years, employing answering question while seeing a picture task to elicit responses. The findings indicate that children do not attain adult-like proficiency in two-level recursion until the age of 6, and there exists a notable asymmetry in the acquisition of recursive possessives versus locatives. These results underscore the primacy of structural complexity and cognitive factors in the acquisition process, enhancing our comprehension of the cognitive foundations of language development and the pivotal role of recursion in child language acquisition.
Authors: Saughmon Boujkian
Abstract: This study examines the cross-linguistic effectiveness of transfer learning for low-resource machine translation by fine-tuning models initially trained on typologically similar high-resource languages, using limited data from the target low-resource language. We hypothesize that linguistic similarity enables efficient adaptation, reducing the need for extensive training data. To test this, we conduct experiments on five typologically diverse language pairs spanning distinct families: Semitic (Modern Standard Arabic to Levantine Arabic), Bantu (Hausa to Zulu), Romance (Spanish to Catalan), Slavic (Slovak to Macedonian), and a language isolate (Eastern Armenian to Western Armenian). Results show that transfer learning consistently improves translation quality across all pairs, confirming its applicability beyond closely related languages. As a secondary analysis, we vary key hyperparameters learning rate, batch size, number of epochs, and weight decay to ensure results are not dependent on a single configuration. We find that moderate batch sizes (e.g., 32) are often optimal for similar pairs, smaller sizes benefit less similar pairs, and excessively high learning rates can destabilize training. These findings provide empirical evidence for the generalizability of transfer learning across language families and offer practical guidance for building machine translation systems in low-resource settings with minimal tuning effort.
Authors: Weijia Xu, Nebojsa Jojic, Sudha Rao, Chris Brockett, Bill Dolan
Abstract: With rapid advances in large language models (LLMs), there has been an increasing application of LLMs in creative content ideation and generation. A critical question emerges: can current LLMs provide ideas that are diverse enough to truly bolster collective creativity? We examine two state-of-the-art LLMs, GPT-4 and LLaMA-3, on story generation and discover that LLM-generated stories often consist of plot elements that are echoed across a number of generations. To quantify this phenomenon, we introduce the Sui Generis score, an automatic metric that measures the uniqueness of a plot element among alternative storylines generated using the same prompt under an LLM. Evaluating on 100 short stories, we find that LLM-generated stories often contain combinations of idiosyncratic plot elements echoed frequently across generations and across different LLMs, while plots from the original human-written stories are rarely recreated or even echoed in pieces. Moreover, our human evaluation shows that the ranking of Sui Generis scores among story segments correlates moderately with human judgment of surprise level, even though score computation is completely automatic without relying on human judgment.
Authors: DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, Qinqing Zheng
Abstract: Large Language Models (LLMs) excel at reasoning and planning when trained on chainof-thought (CoT) data, where the step-by-step thought process is explicitly outlined by text tokens. However, this results in lengthy inputs where many words support textual coherence rather than core reasoning information, and processing these inputs consumes substantial computation resources. In this work, we propose a hybrid representation of the reasoning process, where we partially abstract away the initial reasoning steps using latent discrete tokens generated by VQ-VAE, significantly reducing the length of reasoning traces. We explore the use of latent trace abstractions in two scenarios: 1) training the model from scratch for the Keys-Finding Maze problem, 2) fine-tuning LLMs on this hybrid data with an extended vocabulary including unseen latent tokens, for both logical and mathematical reasoning problems. To facilitate effective learning, we introduce a simple training procedure that randomly mixes latent and text tokens, which enables fast adaptation to new latent tokens. Our approach consistently outperforms the baselines methods in various benchmarks.
Authors: Zirui Song, Bin Yan, Yuhan Liu, Miao Fang, Mingzhe Li, Rui Yan, Xiuying Chen
Abstract: Large Language Models (LLMs) have demonstrated remarkable success in various tasks such as natural language understanding, text summarization, and machine translation. However, their general-purpose nature often limits their effectiveness in domain-specific applications that require specialized knowledge, such as healthcare, chemistry, or legal analysis. To address this, researchers have explored diverse methods to enhance LLMs by integrating domain-specific knowledge. In this survey, we provide a comprehensive overview of these methods, which we categorize into four key approaches: dynamic knowledge injection, static knowledge embedding, modular adapters, and prompt optimization. Each approach offers unique mechanisms to equip LLMs with domain expertise, balancing trade-offs between flexibility, scalability, and efficiency. We discuss how these methods enable LLMs to tackle specialized tasks, compare their advantages and disadvantages, evaluate domain-specific LLMs against general LLMs, and highlight the challenges and opportunities in this emerging field. For those interested in delving deeper into this area, we also summarize the commonly used datasets and benchmarks. To keep researchers updated on the latest studies, we maintain an open-source at: https://github.com/abilliyb/Knowledge_Injection_Survey_Papers, dedicated to documenting research in the field of specialized LLM.
URLs: https://github.com/abilliyb/Knowledge_Injection_Survey_Papers,
Authors: Tianshi Zheng, Jiayang Cheng, Chunyang Li, Haochen Shi, Zihao Wang, Jiaxin Bai, Yangqiu Song, Ginny Y. Wong, Simon See
Abstract: Modern large language models (LLMs) employ various forms of logical inference, both implicitly and explicitly, when addressing reasoning tasks. Understanding how to optimally leverage these inference paradigms is critical for advancing LLMs' reasoning capabilities. This paper adopts an exploratory approach by introducing a controlled evaluation environment for analogical reasoning -- a fundamental cognitive task -- that is systematically parameterized across three dimensions: modality (textual, visual, symbolic), difficulty (easy, medium, hard), and task format (multiple-choice or free-text generation). We analyze the comparative dynamics of inductive, abductive, and deductive inference pipelines across these dimensions, and demonstrate that our findings generalize to broader in-context learning tasks. Additionally, we investigate advanced paradigms such as hypothesis selection, verification, and refinement, revealing their potential to scale up logical inference in LLM reasoning. This exploratory study provides a foundation for future research in enhancing LLM reasoning through systematic logical inference strategies. Resources are available at https://github.com/HKUST-KnowComp/LogiDynamics.
Authors: Jiayi Shi, Yiwei Li, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Huan Ren, Yao Hu, Kan Li
Abstract: Large language models (LLMs) typically undergo instruction tuning to enhance alignment. Recent studies emphasize that quality and diversity of instruction data are more crucial than quantity, highlighting the need to select diverse, high-quality subsets to reduce training costs. However, how to evolve these selected subsets alongside the development of new instruction data remains insufficiently explored. To achieve LLMs' ongoing alignment, we introduce Instruction Bank (\textbf{InsBank}), a continuously updated repository that integrates the latest valuable instruction data. We further propose Progressive Instruction Bank Evolution (\textbf{PIBE}), a novel framework designed to evolve InsBank effectively and efficiently over time. PIBE employs a gradual data selection strategy to maintain long-term efficiency, leveraging a representation-based diversity score to capture relationships between data points and retain historical information for comprehensive diversity evaluation. This also allows for flexible combination of diversity and quality scores during data selection and ranking. Extensive experiments demonstrate that PIBE significantly outperforms baselines in InsBank evolution and is able to extract budget-specific subsets, demonstrating its effectiveness and adaptability.
Authors: Nurkhan Laiyk, Daniil Orel, Rituraj Joshi, Maiya Goloburda, Yuxia Wang, Preslav Nakov, Fajri Koto
Abstract: Instruction tuning in low-resource languages remains underexplored due to limited text data, particularly in government and cultural domains. To address this, we introduce and open-source a large-scale (10,600 samples) instruction-following (IFT) dataset, covering key institutional and cultural knowledge relevant to Kazakhstan. Our dataset enhances LLMs' understanding of procedural, legal, and structural governance topics. We employ LLM-assisted data generation, comparing open-weight and closed-weight models for dataset construction, and select GPT-4o as the backbone. Each entity of our dataset undergoes full manual verification to ensure high quality. We also show that fine-tuning Qwen, Falcon, and Gemma on our dataset leads to consistent performance improvements in both multiple-choice and generative tasks, demonstrating the potential of LLM-assisted instruction tuning for low-resource languages.
Authors: Jonathan Rystr{\o}m, Hannah Rose Kirk, Scott Hale
Abstract: Large Language Models (LLMs) are becoming increasingly capable across global languages. However, the ability to communicate across languages does not necessarily translate to appropriate cultural representations. A key concern is US-centric bias, where LLMs reflect US rather than local cultural values. We propose a novel methodology that compares LLM-generated response distributions against population-level opinion data from the World Value Survey across four languages (Danish, Dutch, English, and Portuguese). Using a rigorous linear mixed-effects regression framework, we compare two families of models: Google's Gemma models (2B--27B parameters) and successive iterations of OpenAI's turbo-series. Across the families of models, we find no consistent relationships between language capabilities and cultural alignment. While the Gemma models have a positive correlation between language capability and cultural alignment across languages, the OpenAI models do not. Importantly, we find that self-consistency is a stronger predictor of multicultural alignment than multilingual capabilities. Our results demonstrate that achieving meaningful cultural alignment requires dedicated effort beyond improving general language capabilities.
Authors: Dayeon Ki, Marine Carpuat
Abstract: Can we improve machine translation (MT) with LLMs by rewriting their inputs automatically? Users commonly rely on the intuition that well-written text is easier to translate when using off-the-shelf MT systems. LLMs can rewrite text in many ways but in the context of MT, these capabilities have been primarily exploited to rewrite outputs via post-editing. We present an empirical study of 21 input rewriting methods with 3 open-weight LLMs for translating from English into 6 target languages. We show that text simplification is the most effective MT-agnostic rewrite strategy and that it can be improved further when using quality estimation to assess translatability. Human evaluation further confirms that simplified rewrites and their MT outputs both largely preserve the original meaning of the source and MT. These results suggest LLM-assisted input rewriting as a promising direction for improving translations.
Authors: Kerem Zaman, Shashank Srivastava
Abstract: Large Language Models (LLMs) offer natural language explanations as an alternative to feature attribution methods for model interpretability. However, despite their plausibility, they may not reflect the model's true reasoning faithfully, which is crucial for understanding the model's true decision-making processes. Although several faithfulness metrics have been proposed, they are often evaluated in isolation, making direct, principled comparisons between them difficult. Here, we present Causal Diagnosticity, a framework that serves as a common testbed to evaluate faithfulness metrics for natural language explanations. Our framework employs the concept of diagnosticity, and uses model-editing methods to generate faithful-unfaithful explanation pairs. Our benchmark includes four tasks: fact-checking, analogy, object counting, and multi-hop reasoning. We evaluate prominent faithfulness metrics, including post-hoc explanation and chain-of-thought-based methods. We find that diagnostic performance varies across tasks and models, with Filler Tokens performing best overall. Additionally, continuous metrics are generally more diagnostic than binary ones but can be sensitive to noise and model choice. Our results highlight the need for more robust faithfulness metrics.
Authors: Zhongqi Yang, Amir Rahmani
Abstract: Large Language Models (LLMs) excel at general-purpose reasoning by leveraging broad commonsense knowledge, but they remain limited in tasks requiring personalized reasoning over multifactorial personal data. This limitation constrains their applicability in domains such as healthcare, where decisions must adapt to individual contexts. We introduce Personalized Causal Graph Reasoning, a framework that enables LLMs to reason over individual-specific causal graphs constructed from longitudinal data. Each graph encodes how user-specific factors influence targeted outcomes. In response to a query, the LLM traverses the graph to identify relevant causal pathways, rank them by estimated impact, simulate potential outcomes, and generate tailored responses. We implement this framework in the context of nutrient-oriented dietary recommendations, where variability in metabolic responses demands personalized reasoning. Using counterfactual evaluation, we assess the effectiveness of LLM-generated food suggestions for glucose control. Our method reduces postprandial glucose iAUC across three time windows compared to prior approaches. Additional LLM-as-a-judge evaluations further confirm improvements in personalization quality.
Authors: Rinku Dewri
Abstract: This article explores the gaps that can manifest when using a large language model (LLM) to obtain simplified interpretations of data practices from a complex privacy policy. We exemplify these gaps to showcase issues in accuracy, completeness, clarity and representation, while advocating for continued research to realize an LLM's true potential in revolutionizing privacy management through personal assistants and automated compliance checking.
Authors: Zhongyuan Wang, Richong Zhang, Zhijie Nie, Hangyu Mao
Abstract: Advanced table question answering (TableQA) methods prompt large language models (LLMs) to generate answer text, SQL query, Python code, or custom operation, which impressively improve the complex reasoning problems in the TableQA task. However, these methods lack the versatility to cope with specific question types or table structures. In contrast, the Spreadsheet Formula, the widely used and well-defined operation language for tabular data, has not been thoroughly explored to solve TableQA. In this paper, we first attempt to use the Formula as the executable representation for solving complex reasoning on tables with different structures. Specifically, we construct \texttt{FromulaQA}, a large Formula-annotated TableQA dataset from existing datasets. In addition, we propose \texttt{TabAF}, a general table answering framework to solve multiple types of tasks over multiple types of tables simultaneously, which decodes answers and Formulas with a single LLM backbone. Extensive experiments demonstrate the versatility and generalization of \texttt{TabAF}. Under the same model size, \texttt{TabAF} achieves new state-of-the-art performance on the WikiTableQuestion, HiTab, and TabFact.
Authors: Andrei-Marius Avram, Marian Lupa\c{s}cu, Dumitru-Clementin Cercel, Ionu\c{t} Mironic\u{a}, \c{S}tefan Tr\u{a}u\c{s}an-Matu
Abstract: This paper presents UniBERT, a compact multilingual language model that uses an innovative training framework that integrates three components: masked language modeling, adversarial training, and knowledge distillation. Pre-trained on a meticulously curated Wikipedia corpus spanning 107 languages, UniBERT is designed to reduce the computational demands of large-scale models while maintaining competitive performance across various natural language processing tasks. Comprehensive evaluations on four tasks - named entity recognition, natural language inference, question answering, and semantic textual similarity - demonstrate that our multilingual training strategy enhanced by an adversarial objective significantly improves cross-lingual generalization. Specifically, UniBERT models show an average relative improvement of 7.72% over traditional baselines, which achieved an average relative improvement of only 1.17%, and statistical analysis confirms the significance of these gains (p-value = 0.0181). This work highlights the benefits of combining adversarial training and knowledge distillation to build scalable and robust language models, thus advancing the field of multilingual and cross-lingual natural language processing.
Authors: Zixin Chen, Sicheng Song, Kashun Shum, Yanna Lin, Rui Sheng, Huamin Qu
Abstract: Misleading visualizations, which manipulate chart representations to support specific claims, can distort perception and lead to incorrect conclusions. Despite decades of research, they remain a widespread issue-posing risks to public understanding and raising safety concerns for AI systems involved in data-driven communication. While recent multimodal large language models (MLLMs) show strong chart comprehension abilities, their capacity to detect and interpret misleading charts remains unexplored. We introduce Misleading ChartQA benchmark, a large-scale multimodal dataset designed to evaluate MLLMs on misleading chart reasoning. It contains 3,026 curated examples spanning 21 misleader types and 10 chart types, each with standardized chart code, CSV data, multiple-choice questions, and labeled explanations, validated through iterative MLLM checks and exhausted expert human review. We benchmark 24 state-of-the-art MLLMs, analyze their performance across misleader types and chart formats, and propose a novel region-aware reasoning pipeline that enhances model accuracy. Our work lays the foundation for developing MLLMs that are robust, trustworthy, and aligned with the demands of responsible visual communication.
Authors: Wenhan Liu, Xinyu Ma, Yutao Zhu, Lixin Su, Shuaiqiang Wang, Dawei Yin, Zhicheng Dou
Abstract: Large Language Models (LLMs) have demonstrated superior listwise ranking performance. However, their superior performance often relies on large-scale parameters (\eg, GPT-4) and a repetitive sliding window process, which introduces significant efficiency challenges. In this paper, we propose \textbf{CoRanking}, a novel collaborative ranking framework that combines small and large ranking models for efficient and effective ranking. CoRanking first employs a small-size reranker to pre-rank all the candidate passages, bringing relevant ones to the top part of the list (\eg, top-20). Then, the LLM listwise reranker is applied to only rerank these top-ranked passages instead of the whole list, substantially enhancing overall ranking efficiency. Although more efficient, previous studies have revealed that the LLM listwise reranker have significant positional biases on the order of input passages. Directly feed the top-ranked passages from small reranker may result in the sub-optimal performance of LLM listwise reranker. To alleviate this problem, we introduce a passage order adjuster trained via reinforcement learning, which reorders the top passages from the small reranker to align with the LLM's preferences of passage order. Extensive experiments on three IR benchmarks demonstrate that CoRanking significantly improves efficiency (reducing ranking latency by about 70\%) while achieving even better effectiveness compared to using only the LLM listwise reranker.
Authors: Bingxiang He, Wenbin Zhang, Jiaxi Song, Cheng Qian, Zixuan Fu, Bowen Sun, Ning Ding, Haiwen Hong, Longtao Huang, Hui Xue, Ganqu Cui, Wanxiang Che, Zhiyuan Liu, Maosong Sun
Abstract: Preference learning is critical for aligning large language models (LLMs) with human values, yet its success hinges on high-quality datasets comprising three core components: Preference \textbf{A}nnotations, \textbf{I}nstructions, and \textbf{R}esponse Pairs. Current approaches conflate these components, obscuring their individual impacts and hindering systematic optimization. In this work, we propose \textbf{AIR}, a component-wise analysis framework that systematically isolates and optimizes each component while evaluating their synergistic effects. Through rigorous experimentation, AIR reveals actionable principles: annotation simplicity (point-wise generative scoring), instruction inference stability (variance-based filtering across LLMs), and response pair quality (moderate margins + high absolute scores). When combined, these principles yield +5.3 average gains over baseline method, even with only 14k high-quality pairs. Our work shifts preference dataset design from ad hoc scaling to component-aware optimization, offering a blueprint for efficient, reproducible alignment.
Authors: Mihnea-Alexandru V\^irlan, R\u{a}zvan-Alexandru Sm\u{a}du, Dumitru-Clementin Cercel, Florin Pop, Mihaela-Claudia Cercel
Abstract: The primary goal of a news headline is to summarize an event in as few words as possible. Depending on the media outlet, a headline can serve as a means to objectively deliver a summary or improve its visibility. For the latter, specific publications may employ stylistic approaches that incorporate the use of sarcasm, irony, and exaggeration, key elements of a satirical approach. As such, even the headline must reflect the tone of the satirical main content. Current approaches for the Romanian language tend to detect the non-conventional tone (i.e., satire and clickbait) of the news content by combining both the main article and the headline. Because we consider a headline to be merely a brief summary of the main article, we investigate in this paper the presence of satirical tone in headlines alone, testing multiple baselines ranging from standard machine learning algorithms to deep learning models. Our experiments show that Bidirectional Transformer models outperform both standard machine-learning approaches and Large Language Models (LLMs), particularly when the meta-learning Reptile approach is employed.
Authors: Juan Diego Rodriguez, Wenxuan Ding, Katrin Erk, Greg Durrett
Abstract: Although large language models (LLMs) have become more capable and accurate across many tasks, some fundamental sources of unreliability remain in their behavior. One key limitation is their inconsistency at reporting the same information when prompts are changed. In this paper, we consider the discrepancy between a model's generated answer and their own verification of that answer, the generator-validator gap. We define this gap in a more stringent way than prior work: we expect correlation of scores from a generator and a validator over the entire set of candidate answers, i.e., candidate completions that could possibly arise during ordinary language use without breaking Gricean norms. We show that according to this measure, a large gap exists in various settings, including question answering, lexical semantics tasks, and next-word prediction. We then propose RankAlign, a ranking-based training method, and show that it significantly closes the gap, surpassing all baseline methods. Moreover, this approach generalizes well to out-of-domain tasks and lexical items.
Authors: Dayeon Ki, Kevin Duh, Marine Carpuat
Abstract: How can a monolingual English speaker determine whether an automatic translation in French is good enough to be shared? Existing MT error detection and quality estimation (QE) techniques do not address this practical scenario. We introduce AskQE, a question generation and answering framework designed to detect critical MT errors and provide actionable feedback, helping users decide whether to accept or reject MT outputs even without the knowledge of the target language. Using ContraTICO, a dataset of contrastive synthetic MT errors in the COVID-19 domain, we explore design choices for AskQE and develop an optimized version relying on LLaMA-3 70B and entailed facts to guide question generation. We evaluate the resulting system on the BioMQM dataset of naturally occurring MT errors, where AskQE has higher Kendall's Tau correlation and decision accuracy with human ratings compared to other QE metrics.
Authors: Minwoo Kang, Suhong Moon, Seung Hyeong Lee, Ayush Raj, Joseph Suh, David M. Chan, John Canny
Abstract: Large language models (LLMs) are increasingly capable of simulating human behavior, offering cost-effective ways to estimate user responses to various surveys and polls. However, the questions in these surveys usually reflect socially understood attitudes: the patterns of attitudes of old/young, liberal/conservative, as understood by both members and non-members of those groups. It is not clear whether the LLM binding is \emph{deep}, meaning the LLM answers as a member of a particular in-group would, or \emph{shallow}, meaning the LLM responds as an out-group member believes an in-group member would. To explore this difference, we use questions that expose known in-group/out-group biases. This level of fidelity is critical for applying LLMs to various political science studies, including timely topics on polarization dynamics, inter-group conflict, and democratic backsliding. To this end, we propose a novel methodology for constructing virtual personas with synthetic user "backstories" generated as extended, multi-turn interview transcripts. This approach is justified by the theory of \emph{narrative identity} which argues that personality at the highest level is \emph{constructed} from self-narratives. Our generated backstories are longer, rich in detail, and consistent in authentically describing a singular individual, compared to previous methods. We show that virtual personas conditioned on our backstories closely replicate human response distributions (up to an 87% improvement as measured by Wasserstein Distance) and produce effect sizes that closely match those observed in the original studies of in-group/out-group biases. Altogether, our work extends the applicability of LLMs beyond estimating socially understood responses, enabling their use in a broader range of human studies.
Authors: Takuma Udagawa, Yang Zhao, Hiroshi Kanayama, Bishwaranjan Bhattacharjee
Abstract: Large language models (LLMs) acquire general linguistic knowledge from massive-scale pretraining. However, pretraining data mainly comprised of web-crawled texts contain undesirable social biases which can be perpetuated or even amplified by LLMs. In this study, we propose an efficient yet effective annotation pipeline to investigate social biases in the pretraining corpora. Our pipeline consists of protected attribute detection to identify diverse demographics, followed by regard classification to analyze the language polarity towards each attribute. Through our experiments, we demonstrate the effect of our bias analysis and mitigation measures, focusing on Common Crawl as the most representative pretraining corpus.
Authors: Adrian Benton, Alexander Gutkin, Christo Kirov, Brian Roark
Abstract: The Latin script is often used to informally write languages with non-Latin native scripts. In many cases (e.g., most languages in India), the lack of conventional spelling in the Latin script results in high spelling variability. Such romanization renders languages that are normally easily distinguished due to being written in different scripts - Hindi and Urdu, for example - highly confusable. In this work, we increase language identification (LID) accuracy for romanized text by improving the methods used to synthesize training sets. We find that training on synthetic samples which incorporate natural spelling variation yields higher LID system accuracy than including available naturally occurring examples in the training set, or even training higher capacity models. We demonstrate new state-of-the-art LID performance on romanized text from 20 Indic languages in the Bhasha-Abhijnaanam evaluation set (Madhani et al., 2023a), improving test F1 from the reported 74.7% (using a pretrained neural model) to 85.4% using a linear classifier trained solely on synthetic data and 88.2% when also training on available harvested text.
Authors: Miaomiao Ji, Yanqiu Wu, Zhibin Wu, Shoujin Wang, Jian Yang, Mark Dras, Usman Naseem
Abstract: Reward design plays a pivotal role in aligning large language models (LLMs) with human values, serving as the bridge between feedback signals and model optimization. This survey provides a structured organization of reward modeling and addresses three key aspects: mathematical formulation, construction practices, and interaction with optimization paradigms. Building on this, it develops a macro-level taxonomy that characterizes reward mechanisms along complementary dimensions, thereby offering both conceptual clarity and practical guidance for alignment research. The progression of LLM alignment can be understood as a continuous refinement of reward design strategies, with recent developments highlighting paradigm shifts from reinforcement learning (RL)-based to RL-free optimization and from single-task to multi-objective and complex settings.
Authors: Yaorui Shi, Sihang Li, Chang Wu, Zhiyuan Liu, Junfeng Fang, Hengxing Cai, An Zhang, Xiang Wang
Abstract: Large language models have demonstrated impressive reasoning capabilities but are inherently limited by their knowledge reservoir. Retrieval-augmented reasoning mitigates this limitation by allowing LLMs to query external resources, but existing methods often retrieve irrelevant or noisy information, hindering accurate reasoning. In this paper, we propose AutoRefine, a reinforcement learning post-training framework that adopts a new ``search-and-refine-during-think'' paradigm. AutoRefine introduces explicit knowledge refinement steps between successive search calls, enabling the model to iteratively filter, distill, and organize evidence before generating an answer. Furthermore, we incorporate tailored retrieval-specific rewards alongside answer correctness rewards using group relative policy optimization. Experiments on single-hop and multi-hop QA benchmarks demonstrate that AutoRefine significantly outperforms existing approaches, particularly in complex, multi-hop reasoning scenarios. Detailed analysis shows that AutoRefine issues frequent, higher-quality searches and synthesizes evidence effectively.
Authors: Xiaomin Li, Zhou Yu, Zhiwei Zhang, Xupeng Chen, Ziji Zhang, Yingying Zhuang, Narayanan Sadagopan, Anurag Beniwal
Abstract: Reasoning-enhanced large language models (RLLMs), whether explicitly trained for reasoning or prompted via chain-of-thought (CoT), have achieved state-of-the-art performance on many complex reasoning tasks. However, we uncover a surprising and previously overlooked phenomenon: explicit CoT reasoning can significantly degrade instruction-following accuracy. Evaluating 15 models on two benchmarks: IFEval (with simple, rule-verifiable constraints) and ComplexBench (with complex, compositional constraints), we consistently observe performance drops when CoT prompting is applied. Through large-scale case studies and an attention-based analysis, we identify common patterns where reasoning either helps (e.g., with formatting or lexical precision) or hurts (e.g., by neglecting simple constraints or introducing unnecessary content). We propose a metric, constraint attention, to quantify model focus during generation and show that CoT reasoning often diverts attention away from instruction-relevant tokens. To mitigate these effects, we introduce and evaluate four strategies: in-context learning, self-reflection, self-selective reasoning, and classifier-selective reasoning. Our results demonstrate that selective reasoning strategies, particularly classifier-selective reasoning, can substantially recover lost performance. To our knowledge, this is the first work to systematically expose reasoning-induced failures in instruction-following and offer practical mitigation strategies.
Authors: Tianshi Zheng, Zheye Deng, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Zihao Wang, Yangqiu Song
Abstract: Large Language Models (LLMs) are catalyzing a paradigm shift in scientific discovery, evolving from task-specific automation tools into increasingly autonomous agents and fundamentally redefining research processes and human-AI collaboration. This survey systematically charts this burgeoning field, placing a central focus on the changing roles and escalating capabilities of LLMs in science. Through the lens of the scientific method, we introduce a foundational three-level taxonomy-Tool, Analyst, and Scientist-to delineate their escalating autonomy and evolving responsibilities within the research lifecycle. We further identify pivotal challenges and future research trajectories such as robotic automation, self-improvement, and ethical governance. Overall, this survey provides a conceptual architecture and strategic foresight to navigate and shape the future of AI-driven scientific discovery, fostering both rapid innovation and responsible advancement. Github Repository: https://github.com/HKUST-KnowComp/Awesome-LLM-Scientific-Discovery.
URLs: https://github.com/HKUST-KnowComp/Awesome-LLM-Scientific-Discovery.
Authors: Huihao Jing, Haoran Li, Wenbin Hu, Qi Hu, Heli Xu, Tianshu Chu, Peizhao Hu, Yangqiu Song
Abstract: As Model Context Protocol (MCP) introduces an easy-to-use ecosystem for users and developers, it also brings underexplored safety risks. Its decentralized architecture, which separates clients and servers, poses unique challenges for systematic safety analysis. This paper proposes a novel framework to enhance MCP safety. Guided by the MAESTRO framework, we first analyze the missing safety mechanisms in MCP, and based on this analysis, we propose the Model Contextual Integrity Protocol (MCIP), a refined version of MCP that addresses these gaps. Next, we develop a fine-grained taxonomy that captures a diverse range of unsafe behaviors observed in MCP scenarios. Building on this taxonomy, we develop benchmark and training data that support the evaluation and improvement of LLMs' capabilities in identifying safety risks within MCP interactions. Leveraging the proposed benchmark and training data, we conduct extensive experiments on state-of-the-art LLMs. The results highlight LLMs' vulnerabilities in MCP interactions and demonstrate that our approach substantially improves their safety performance.
Authors: Ryang Heo, Yongsik Seo, Junseong Lee, Dongha Lee
Abstract: The surge of user-generated online content presents a wealth of insights into customer preferences and market trends. However, the highly diverse, complex, and context-rich nature of such contents poses significant challenges to traditional opinion mining approaches. To address this, we introduce Online Opinion Mining Benchmark (OOMB), a novel dataset and evaluation protocol designed to assess the ability of large language models (LLMs) to mine opinions effectively from diverse and intricate online environments. OOMB provides extensive (entity, feature, opinion) tuple annotations and a comprehensive opinion-centric summary that highlights key opinion topics within each content, thereby enabling the evaluation of both the extractive and abstractive capabilities of models. Through our proposed benchmark, we conduct a comprehensive analysis of which aspects remain challenging and where LLMs exhibit adaptability, to explore whether they can effectively serve as opinion miners in realistic online scenarios. This study lays the foundation for LLM-based opinion mining and discusses directions for future research in this field.
Authors: Kristin Qi, Jiali Cheng, Youxiang Zhu, Hadi Amiri, Xiaohui Liang
Abstract: Detecting Mild Cognitive Impairment from picture descriptions is critical yet challenging, especially in multilingual and multiple picture settings. Prior work has primarily focused on English speakers describing a single picture (e.g., the 'Cookie Theft'). The TAUKDIAL-2024 challenge expands this scope by introducing multilingual speakers and multiple pictures, which presents new challenges in analyzing picture-dependent content. To address these challenges, we propose a framework with three components: (1) enhancing discriminative representation learning via supervised contrastive learning, (2) involving image modality rather than relying solely on speech and text modalities, and (3) applying a Product of Experts (PoE) strategy to mitigate spurious correlations and overfitting. Our framework improves MCI detection performance, achieving a +7.1% increase in Unweighted Average Recall (UAR) (from 68.1% to 75.2%) and a +2.9% increase in F1 score (from 80.6% to 83.5%) compared to the text unimodal baseline. Notably, the contrastive learning component yields greater gains for the text modality compared to speech. These results highlight our framework's effectiveness in multilingual and multi-picture MCI detection.
Authors: Chen Shani, Liron Soffer, Dan Jurafsky, Yann LeCun, Ravid Shwartz-Ziv
Abstract: Humans organize knowledge into compact categories through semantic compression by mapping diverse instances to abstract representations while preserving meaning (e.g., robin and blue jay are both birds; most birds can fly). These concepts reflect a trade-off between expressive fidelity and representational simplicity. Large Language Models (LLMs) demonstrate remarkable linguistic abilities, yet whether their internal representations strike a human-like trade-off between compression and semantic fidelity is unclear. We introduce a novel information-theoretic framework, drawing from Rate-Distortion Theory and the Information Bottleneck principle, to quantitatively compare these strategies. Analyzing token embeddings from a diverse suite of LLMs against seminal human categorization benchmarks, we uncover key divergences. While LLMs form broad conceptual categories that align with human judgment, they struggle to capture the fine-grained semantic distinctions crucial for human understanding. More fundamentally, LLMs demonstrate a strong bias towards aggressive statistical compression, whereas human conceptual systems appear to prioritize adaptive nuance and contextual richness, even if this results in lower compressional efficiency by our measures. These findings illuminate critical differences between current AI and human cognitive architectures, guiding pathways toward LLMs with more human-aligned conceptual representations.
Authors: Kristin Qi, Youxiang Zhu, Caroline Summerour, John A. Batsis, Xiaohui Liang
Abstract: Early detection of cognitive decline is crucial for enabling interventions that can slow neurodegenerative disease progression. Traditional diagnostic approaches rely on labor-intensive clinical assessments, which are impractical for frequent monitoring. Our pilot study investigates voice assistant systems (VAS) as non-invasive tools for detecting cognitive decline through longitudinal analysis of speech patterns in voice commands. Over an 18-month period, we collected voice commands from 35 older adults, with 15 participants providing daily at-home VAS interactions. To address the challenges of analyzing these short, unstructured and noisy commands, we propose Cog-TiPRO, a framework that combines (1) LLM-driven iterative prompt refinement for linguistic feature extraction, (2) HuBERT-based acoustic feature extraction, and (3) transformer-based temporal modeling. Using iTransformer, our approach achieves 73.80% accuracy and 72.67% F1-score in detecting MCI, outperforming its baseline by 27.13%. Through our LLM approach, we identify linguistic features that uniquely characterize everyday command usage patterns in individuals experiencing cognitive decline.
Authors: Abhijit Chakraborty, Chahana Dahal, Vivek Gupta
Abstract: Federated Retrieval-Augmented Generation (Federated RAG) combines Federated Learning (FL), which enables distributed model training without exposing raw data, with Retrieval-Augmented Generation (RAG), which improves the factual accuracy of language models by grounding outputs in external knowledge. As large language models are increasingly deployed in privacy-sensitive domains such as healthcare, finance, and personalized assistance, Federated RAG offers a promising framework for secure, knowledge-intensive natural language processing (NLP). To the best of our knowledge, this paper presents the first systematic mapping study of Federated RAG, covering literature published between 2020 and 2025. Following Kitchenham's guidelines for evidence-based software engineering, we develop a structured classification of research focuses, contribution types, and application domains. We analyze architectural patterns, temporal trends, and key challenges, including privacy-preserving retrieval, cross-client heterogeneity, and evaluation limitations. Our findings synthesize a rapidly evolving body of research, identify recurring design patterns, and surface open questions, providing a foundation for future work at the intersection of RAG and federated systems.
Authors: Christopher Ormerod
Abstract: This study illustrates how incorporating feedback-oriented annotations into the scoring pipeline can enhance the accuracy of automated essay scoring (AES). This approach is demonstrated with the Persuasive Essays for Rating, Selecting, and Understanding Argumentative and Discourse Elements (PERSUADE) corpus. We integrate two types of feedback-driven annotations: those that identify spelling and grammatical errors, and those that highlight argumentative components. To illustrate how this method could be applied in real-world scenarios, we employ two LLMs to generate annotations -- a generative language model used for spell correction and an encoder-based token-classifier trained to identify and mark argumentative elements. By incorporating annotations into the scoring process, we demonstrate improvements in performance using encoder-based large language models fine-tuned as classifiers.
Authors: Dayeon Ki, Rachel Rudinger, Tianyi Zhou, Marine Carpuat
Abstract: Large Language Models (LLMs) need to adapt their predictions to diverse cultural contexts to benefit diverse communities across the world. While previous efforts have focused on single-LLM, single-turn approaches, we propose to exploit the complementary strengths of multiple LLMs to promote cultural adaptability. We introduce a Multi-Agent Debate framework, where two LLM-based agents debate over a cultural scenario and collaboratively reach a final decision. We propose two variants: one where either LLM agents exclusively debate and another where they dynamically choose between self-reflection and debate during their turns. We evaluate these approaches on 7 open-weight LLMs (and 21 LLM combinations) using the NormAd-ETI benchmark for social etiquette norms in 75 countries. Experiments show that debate improves both overall accuracy and cultural group parity over single-LLM baselines. Notably, multi-agent debate enables relatively small LLMs (7-9B) to achieve accuracies comparable to that of a much larger model (27B parameters).
Authors: Dayeon Ki, Kevin Duh, Marine Carpuat
Abstract: As people increasingly use AI systems in work and daily life, feedback mechanisms that help them use AI responsibly are urgently needed, particularly in settings where users are not equipped to assess the quality of AI predictions. We study a realistic Machine Translation (MT) scenario where monolingual users decide whether to share an MT output, first without and then with quality feedback. We compare four types of quality feedback: explicit feedback that directly give users an assessment of translation quality using (1) error highlights and (2) LLM explanations, and implicit feedback that helps users compare MT inputs and outputs through (3) backtranslation and (4) question-answer (QA) tables. We find that all feedback types, except error highlights, significantly improve both decision accuracy and appropriate reliance. Notably, implicit feedback, especially QA tables, yields significantly greater gains than explicit feedback in terms of decision accuracy, appropriate reliance, and user perceptions, receiving the highest ratings for helpfulness and trust, and the lowest for mental burden.
Authors: Anshun Asher Zheng, Junyi Jessy Li, David I. Beaver
Abstract: Language is often used strategically, particularly in high-stakes, adversarial settings, yet most work on pragmatics and LLMs centers on cooperativity. This leaves a gap in the systematic understanding of strategic communication in adversarial settings. To address this, we introduce SDA (Strategic Discourse Assessment), a framework grounded in Gricean and game-theoretic pragmatics to assess strategic use of language. It adapts the ME Game jury function to make it empirically estimable for analyzing dialogue. Our approach incorporates two key adaptations: a commitment-based taxonomy of discourse moves, which provides a finer-grained account of strategic effects, and the use of estimable proxies grounded in Gricean maxims to operationalize abstract constructs such as credibility. Together, these adaptations build on discourse theory by treating discourse as the strategic management of commitments, enabling systematic evaluation of how conversational moves advance or undermine discourse goals. We further derive three interpretable metrics-Benefit at Turn (BAT), Penalty at Turn (PAT), and Normalized Relative Benefit at Turn (NRBAT)-to quantify the perceived strategic effects of discourse moves. We also present CPD (the Crooked Path Dataset), an annotated dataset of real courtroom cross-examinations, to demonstrate the framework's effectiveness. Using these tools, we evaluate a range of LLMs and show that LLMs generally exhibit limited pragmatic understanding of strategic language. While model size shows an increase in performance on our metrics, reasoning ability does not help and largely hurts, introducing overcomplication and internal confusion.
Authors: Feng Wang, Yiding Sun, Jiaxin Mao, Wei Xue, Danqing Xu
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across various professional domains, with their performance typically evaluated through standardized benchmarks. In the financial field, the stringent demands for professional accuracy and real-time data processing often necessitate the use of retrieval-augmented generation (RAG) techniques. However, the development of financial RAG benchmarks has been constrained by data confidentiality issues and the lack of dynamic data integration. To address this issue, we introduce FinS-Pilot, a novel benchmark for evaluating RAG systems in online financial applications. Constructed from real-world financial assistant interactions, our benchmark incorporates both real-time API data and text data, organized through an intent classification framework covering critical financial domains. The benchmark enables comprehensive evaluation of financial assistants' capabilities in handling both static knowledge and time-sensitive market information.Through systematic experiments with multiple Chinese leading LLMs, we demonstrate FinS-Pilot's effectiveness in identifying models suitable for financial applications while addressing the current gap in specialized evaluation tools for the financial domain. Our work contributes both a practical evaluation framework and a curated dataset to advance research in financial NLP systems. The code and dataset are accessible on GitHub.
Authors: Zetong Tang, Qian Ma, Di Wu
Abstract: Using the best Text-to-SQL methods in resource-constrained environments is challenging due to their reliance on resource-intensive open-source models. This paper introduces Auto Prompt SQL(AP-SQL), a novel architecture designed to bridge the gap between resource-efficient small open-source models and the powerful capabilities of large closed-source models for Text-to-SQL translation. Our method decomposes the task into schema filtering, retrieval-augmented text-to-SQL generation based on in-context examples, and prompt-driven schema linking and SQL generation. To improve schema selection accuracy, we fine-tune large language models. Crucially, we also explore the impact of prompt engineering throughout the process, leveraging Chain-of-Thought(CoT) and Graph-of-Thought(GoT) templates to significantly enhance the model's reasoning for accurate SQL generation. Comprehensive evaluations on the Spider benchmarks demonstrate the effectiveness of AP-SQL.
Authors: Ming Zhang, Yujiong Shen, Zelin Li, Huayu Sha, Binze Hu, Yuhui Wang, Chenhao Huang, Shichun Liu, Jingqi Tong, Changhao Jiang, Mingxu Chai, Zhiheng Xi, Shihan Dou, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract: Evaluating large language models (LLMs) in medicine is crucial because medical applications require high accuracy with little room for error. Current medical benchmarks have three main types: medical exam-based, comprehensive medical, and specialized assessments. However, these benchmarks have limitations in question design (mostly multiple-choice), data sources (often not derived from real clinical scenarios), and evaluation methods (poor assessment of complex reasoning). To address these issues, we present LLMEval-Med, a new benchmark covering five core medical areas, including 2,996 questions created from real-world electronic health records and expert-designed clinical scenarios. We also design an automated evaluation pipeline, incorporating expert-developed checklists into our LLM-as-Judge framework. Furthermore, our methodology validates machine scoring through human-machine agreement analysis, dynamically refining checklists and prompts based on expert feedback to ensure reliability. We evaluate 13 LLMs across three categories (specialized medical models, open-source models, and closed-source models) on LLMEval-Med, providing valuable insights for the safe and effective deployment of LLMs in medical domains. The dataset is released in https://github.com/llmeval/LLMEval-Med.
Authors: Aleksandra Sorokovikova, Pavel Chizhov, Iuliia Eremenko, Ivan P. Yamshchikov
Abstract: Modern language models are trained on large amounts of data. These data inevitably include controversial and stereotypical content, which contains all sorts of biases related to gender, origin, age, etc. As a result, the models express biased points of view or produce different results based on the assigned personality or the personality of the user. In this paper, we investigate various proxy measures of bias in large language models (LLMs). We find that evaluating models with pre-prompted personae on a multi-subject benchmark (MMLU) leads to negligible and mostly random differences in scores. However, if we reformulate the task and ask a model to grade the user's answer, this shows more significant signs of bias. Finally, if we ask the model for salary negotiation advice, we see pronounced bias in the answers. With the recent trend for LLM assistant memory and personalization, these problems open up from a different angle: modern LLM users do not need to pre-prompt the description of their persona since the model already knows their socio-demographics.
Authors: Abdullah Al Shafi, Rowzatul Zannat, Abdul Muntakim, Mahmudul Hasan
Abstract: Disease-symptom datasets are significant and in demand for medical research, disease diagnosis, clinical decision-making, and AI-driven health management applications. These datasets help identify symptom patterns associated with specific diseases, thus improving diagnostic accuracy and enabling early detection. The dataset presented in this study systematically compiles disease-symptom relationships from various online sources, medical literature, and publicly available health databases. The data was gathered through analyzing peer-reviewed medical articles, clinical case studies, and disease-symptom association reports. Only the verified medical sources were included in the dataset, while those from non-peer-reviewed and anecdotal sources were excluded. The dataset is structured in a tabular format, where the first column represents diseases, and the remaining columns represent symptoms. Each symptom cell contains a binary value, indicating whether a symptom is associated with a disease. Thereby, this structured representation makes the dataset very useful for a wide range of applications, including machine learning-based disease prediction, clinical decision support systems, and epidemiological studies. Although there are some advancements in the field of disease-symptom datasets, there is a significant gap in structured datasets for the Bangla language. This dataset aims to bridge that gap by facilitating the development of multilingual medical informatics tools and improving disease prediction models for underrepresented linguistic communities. Further developments should include region-specific diseases and further fine-tuning of symptom associations for better diagnostic performance
Authors: Fabien Furfaro
Abstract: Transformer-based large language models (LLMs) have achieved strong performance across many natural language processing tasks. Nonetheless, their quadratic computational and memory requirements, particularly in self-attention layers, pose challenges for efficient inference on long contexts and for deployment in resource-limited environments. We present TPTT (Transforming Pretrained Transformers into Titans), a framework designed to augment pretrained Transformers with linearized attention (LiZA) and internal memory gating via Memory as Gate (MaG), applied without full retraining. TPTT supports parameter-efficient fine-tuning (LoRA) and integrates with standard toolkits such as Hugging Face Transformers. We evaluated TPTT on several pretrained models, including Llama-1B, OlMoE-1B-7B, Qwen2.5-1.5B, Gemma3-270m, OpenELM-1.3B, and Mistral-7B, in order to assess applicability across architectures of different scales.Experiments on models with approximately 1 billion parameters, evaluated primarily on the MMLU benchmark, suggest potential improvements in both efficiency and accuracy compared to baseline models. For example, Titans-Llama-1B exhibited up to a 20\% relative increase in Exact Match scores in one-shot evaluation. An additional finding is that it is possible to convert a quadratic-attention model into a purely linear-attention model using the DeltaProduct mechanism. All training runs were carried out with modest computational resources.These preliminary findings indicate that TPTT may help adapt pretrained LLMs for long-context tasks with limited overhead. Further studies on larger models and a broader set of benchmarks will be necessary to evaluate the generality and robustness of the framework. Code is available at https://github.com/fabienfrfr/tptt . Python package at https://pypi.org/project/tptt/ .
URLs: https://github.com/fabienfrfr/tptt, https://pypi.org/project/tptt/
Authors: Siyuan Li, Ruitong Liu, Yan Wen, Te Sun, Andi Zhang, Yanbiao Ma, Xiaoshuai Hao
Abstract: Knowledge graph completion demands effective modeling of multifaceted semantic relationships between entities. Yet, prevailing methods, which rely on static scoring functions over learned embeddings, struggling to simultaneously capture rich semantic context and the dynamic nature of relations. To overcome this limitation, we propose the Flow-Modulated Scoring (FMS) framework, conceptualizing a relation as a dynamic evolutionary process governed by its static semantic environment. FMS operates in two stages: it first learns context-aware entity embeddings via a Semantic Context Learning module, and then models a dynamic flow between them using a Conditional Flow-Matching module. This learned flow dynamically modulates a base static score for the entity pair. By unifying context-rich static representations with a conditioned dynamic flow, FMS achieves a more comprehensive understanding of relational semantics. Extensive experiments demonstrate that FMS establishes a new state of the art across both canonical knowledge graph completion tasks: relation prediction and entity prediction. On the standard relation prediction benchmark FB15k-237, FMS achieves a near-perfect MRR of 99.8\% and Hits@1 of 99.7\% using a mere 0.35M parameters, while also attaining a 99.9\% MRR on WN18RR. Its dominance extends to entity prediction, where it secures a 25.2\% relative MRR gain in the transductive setting and substantially outperforms all baselines in challenging inductive settings. By unifying a dynamic flow mechanism with rich static contexts, FMS offers a highly effective and parameter-efficient new paradigm for knowledge graph completion. Code published at: https://github.com/yuanwuyuan9/FMS.
Authors: Asad Aali, Vasiliki Bikia, Maya Varma, Nicole Chiou, Sophie Ostmeier, Arnav Singhvi, Magdalini Paschali, Ashwin Kumar, Andrew Johnston, Karimar Amador-Martinez, Eduardo Juan Perez Guerrero, Paola Naovi Cruz Rivera, Sergios Gatidis, Christian Bluethgen, Eduardo Pontes Reis, Eddy D. Zandee van Rilland, Poonam Laxmappa Hosamani, Kevin R Keet, Minjoung Go, Evelyn Ling, David B. Larson, Curtis Langlotz, Roxana Daneshjou, Jason Hom, Sanmi Koyejo, Emily Alsentzer, Akshay S. Chaudhari
Abstract: With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the "LM-as-judge" paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a self-supervised framework that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset containing 840 outputs annotated by physicians, following a physician-defined taxonomy of risk levels and error categories. Across 6 diverse medical tasks and 10 state-of-the-art LMs spanning open-source, proprietary, and medically adapted models, MedVAL fine-tuning significantly improves (p < 0.001) alignment with physicians on both seen and unseen tasks, increasing average F1 scores from 66% to 83%, with per-sample safety classification scores up to 86%. MedVAL improves the performance of even the best-performing proprietary LM (GPT-4o) by 8%. To support a scalable, risk-aware pathway towards clinical integration, we open-source the 1) codebase (https://github.com/StanfordMIMI/MedVAL), 2) MedVAL-Bench (https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench), and 3) MedVAL-4B (https://huggingface.co/stanfordmimi/MedVAL-4B), the best-performing open-source LM. Our research provides the first evidence of LMs approaching expert-level validation ability for medical text.
URLs: https://github.com/StanfordMIMI/MedVAL),, https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench),, https://huggingface.co/stanfordmimi/MedVAL-4B),
Authors: Weihua Du, Pranjal Aggarwal, Sean Welleck, Yiming Yang
Abstract: Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces. Tool-augmented agents address arithmetic via code execution, but often falter on complex logical tasks. We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model. Using this approach, we train Agentic-R1, which dynamically selects the optimal strategy for each query, invoking tools for arithmetic and algorithmic problems, and using text-based reasoning for abstract ones. Our method improves accuracy across a range of tasks, including both computation-intensive and standard benchmarks, demonstrating the effectiveness of multi-strategy distillation in achieving robust and efficient reasoning. Our project is available at https://github.com/StigLidu/DualDistill
Authors: Yinzhu Quan, Xinrui Li, Ying Chen
Abstract: In e-commerce private-domain channels such as instant messaging and e-mail, merchants engage customers directly as part of their Customer Relationship Management (CRM) programmes to drive retention and conversion. While a few top performers excel at crafting outbound messages, most merchants struggle to write persuasive copy because they lack both expertise and scalable tools. We introduce CRMAgent, a multi-agent system built on large language models (LLMs) that generates high-quality message templates and actionable writing guidance through three complementary modes. First, group-based learning enables the agent to learn from a merchant's own top-performing messages within the same audience segment and rewrite low-performing ones. Second, retrieval-and-adaptation fetches templates that share the same audience segment and exhibit high similarity in voucher type and product category, learns their successful patterns, and adapts them to the current campaign. Third, a rule-based fallback provides a lightweight zero-shot rewrite when no suitable references are available. Extensive experiments show that CRMAgent consistently outperforms merchants' original templates, delivering significant gains in both audience-match and marketing-effectiveness metrics.
Authors: Patr\'icia Schmidtov\'a, Niyati Bafna, Seth Aycock, Gianluca Vico, Wiktor Kamzela, Katharina H\"ammerl, Vil\'em Zouhar
Abstract: Large language models (LLMs) have achieved top results in recent machine translation evaluations, but they are also known to be sensitive to errors and perturbations in their prompts. We systematically evaluate how both humanly plausible and synthetic errors in user prompts affect LLMs' performance on two related tasks: Machine translation and machine translation evaluation. We provide both a quantitative analysis and qualitative insights into how the models respond to increasing noise in the user prompt. The prompt quality strongly affects the translation performance: With many errors, even a good prompt can underperform a minimal or poor prompt without errors. However, different noise types impact translation quality differently, with character-level and combined noisers degrading performance more than phrasal perturbations. Qualitative analysis reveals that lower prompt quality largely leads to poorer instruction following, rather than directly affecting translation quality itself. Further, LLMs can still translate in scenarios with overwhelming random noise that would make the prompt illegible to humans.
Authors: Junhong Ye, Xu Yuan, Xinying Qiu
Abstract: Accurate recognition of personally identifiable information (PII) is central to automated text anonymization. This paper investigates the effectiveness of cross-domain model transfer, multi-domain data fusion, and sample-efficient learning for PII recognition. Using annotated corpora from healthcare (I2B2), legal (TAB), and biography (Wikipedia), we evaluate models across four dimensions: in-domain performance, cross-domain transferability, fusion, and few-shot learning. Results show legal-domain data transfers well to biographical texts, while medical domains resist incoming transfer. Fusion benefits are domain-specific, and high-quality recognition is achievable with only 10% of training data in low-specialization domains.
Authors: Maciej Jalocha, Johan Hausted Schmidt, William Michelseen
Abstract: The field of cybersecurity NER lacks standardized labels, making it challenging to combine datasets. We investigate label unification across four cybersecurity datasets to increase data resource usability. We perform a coarse-grained label unification and conduct pairwise cross-dataset evaluations using BiLSTM models. Qualitative analysis of predictions reveals errors, limitations, and dataset differences. To address unification limitations, we propose alternative architectures including a multihead model and a graph-based transfer model. Results show that models trained on unified datasets generalize poorly across datasets. The multihead model with weight sharing provides only marginal improvements over unified training, while our graph-based transfer model built on BERT-base-NER shows no significant performance gains compared BERT-base-NER.
Authors: Bhishma Dedhia, Yuval Kansal, Niraj K. Jha
Abstract: Language models traditionally used for cross-domain generalization have recently demonstrated task-specific reasoning. However, their top-down training approach on general corpora is insufficient for acquiring abstractions needed for deep domain expertise. This may require a bottom-up approach that acquires expertise by learning to compose simple domain concepts into more complex ones. A knowledge graph (KG) provides this compositional structure, where domain primitives are represented as head-relation-tail edges and their paths encode higher-level concepts. We present a task generation pipeline that synthesizes tasks directly from KG primitives, enabling models to acquire and compose them for reasoning. We fine-tune language models on the resultant KG-grounded curriculum to demonstrate domain-specific superintelligence. While broadly applicable, we validate our approach in medicine, where reliable KGs exist. Using a medical KG, we curate 24,000 reasoning tasks paired with thinking traces derived from diverse medical primitives. We fine-tune the QwQ-32B model on this curriculum to obtain QwQ-Med-3 that takes a step towards medical superintelligence. We also introduce ICD-Bench, an evaluation suite to quantify reasoning abilities across 15 medical domains. Our experiments demonstrate that QwQ-Med-3 significantly outperforms state-of-the-art reasoning models on ICD-Bench categories. Further analysis reveals that QwQ-Med-3 utilizes acquired primitives to widen the performance gap on the hardest tasks of ICD-Bench. Finally, evaluation on medical question-answer benchmarks shows that QwQ-Med-3 transfers acquired expertise to enhance the base model's performance. While the industry's approach to artificial general intelligence (AGI) emphasizes broad expertise, we envision a future in which AGI emerges from the composable interaction of efficient domain-specific superintelligent agents.
Authors: Chathuri Jayaweera, Bonnie J. Dorr
Abstract: This position paper argues that annotation disagreement in Natural Language Inference (NLI) is not mere noise but often reflects meaningful variation, especially when triggered by ambiguity in the premise or hypothesis. While underspecified guidelines and annotator behavior contribute to variation, content-based ambiguity provides a process-independent signal of divergent human perspectives. We call for a shift toward ambiguity-aware NLI that first identifies ambiguous input pairs, classifies their types, and only then proceeds to inference. To support this shift, we present a framework that incorporates ambiguity detection and classification prior to inference. We also introduce a unified taxonomy that synthesizes existing taxonomies, illustrates key subtypes with examples, and motivates targeted detection methods that better align models with human interpretation. Although current resources lack datasets explicitly annotated for ambiguity and subtypes, this gap presents an opportunity: by developing new annotated resources and exploring unsupervised approaches to ambiguity detection, we enable more robust, explainable, and human-aligned NLI systems.
Authors: Shahriar Golchin, Yanfei Chen, Rujun Han, Manan Gandhi, Tianli Yu, Swaroop Mishra, Mihai Surdeanu, Rishabh Agarwal, Chen-Yu Lee, Tomas Pfister
Abstract: Long-context large language models (LLMs) are able to process inputs containing up to several million tokens. In the scope of in-context learning (ICL), this translates into using hundreds/thousands of demonstrations in the input prompt, enabling many-shot ICL. In practice, a fixed set of demonstrations is often selected at random in many-shot settings due to (1) high inference costs, (2) the benefits of caching and reusing computations, and (3) the similar performance offered by this strategy compared to others when scaled. In this work, we propose two straightforward strategies for demonstration selection in many-shot ICL that improve performance with minimal computational overhead. Our first method combines a small number of demonstrations, selected based on their similarity to each test sample, with a disproportionately larger set of random demonstrations that are cached. The second strategy improves the first by replacing random demonstrations with those selected using centroids derived from test sample representations via k-means clustering. Our experiments with Gemini Pro and Flash across several datasets indicate that our strategies consistently outperform random selection and surpass or match the most performant selection approach while supporting caching and reducing inference cost by up to an order of magnitude. We also show that adjusting the proportion of demonstrations selected based on different criteria can balance performance and inference cost in many-shot ICL.
Authors: Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, Jack Lindsey
Abstract: Large language models interact with users through a simulated 'Assistant' persona. While the Assistant is typically trained to be helpful, harmless, and honest, it sometimes deviates from these ideals. In this paper, we identify directions in the model's activation space-persona vectors-underlying several traits, such as evil, sycophancy, and propensity to hallucinate. We confirm that these vectors can be used to monitor fluctuations in the Assistant's personality at deployment time. We then apply persona vectors to predict and control personality shifts that occur during training. We find that both intended and unintended personality changes after finetuning are strongly correlated with shifts along the relevant persona vectors. These shifts can be mitigated through post-hoc intervention, or avoided in the first place with a new preventative steering method. Moreover, persona vectors can be used to flag training data that will produce undesirable personality changes, both at the dataset level and the individual sample level. Our method for extracting persona vectors is automated and can be applied to any personality trait of interest, given only a natural-language description.
Authors: Vasudha Varadarajan, Hui Xu, Rebecca Astrid Boehme, Mariam Marlan Mirstrom, Sverker Sikstrom, H. Andrew Schwartz
Abstract: Recent advances in large language models (LLMs) offer new opportunities for scalable, interactive mental health assessment, but excessive querying by LLMs burdens users and is inefficient for real-world screening across transdiagnostic symptom profiles. We introduce MAQuA, an adaptive question-asking framework for simultaneous, multidimensional mental health screening. Combining multi-outcome modeling on language responses with item response theory (IRT) and factor analysis, MAQuA selects the questions with most informative responses across multiple dimensions at each turn to optimize diagnostic information, improving accuracy and potentially reducing response burden. Empirical results on a novel dataset reveal that MAQuA reduces the number of assessment questions required for score stabilization by 50-87% compared to random ordering (e.g., achieving stable depression scores with 71% fewer questions and eating disorder scores with 85% fewer questions). MAQuA demonstrates robust performance across both internalizing (depression, anxiety) and externalizing (substance use, eating disorder) domains, with early stopping strategies further reducing patient time and burden. These findings position MAQuA as a powerful and efficient tool for scalable, nuanced, and interactive mental health screening, advancing the integration of LLM-based agents into real-world clinical workflows.
Authors: Kaiyu Wang, Lin Mu, Zhiyao Yang, Ximing Li, Xiaotang Zhou Wanfu Gao, Huimao Zhang
Abstract: Quality Assurance (QA) for radiology reports refers to judging whether the junior reports (written by junior doctors) are qualified. The QA scores of one junior report are given by the senior doctor(s) after reviewing the image and junior report. This process requires intensive labor costs for senior doctors. Additionally, the QA scores may be inaccurate for reasons like diagnosis bias, the ability of senior doctors, and so on. To address this issue, we propose a Span-level Quality Assurance EvaluaTOR (Sqator) to mark QA scores automatically. Unlike the common document-level semantic comparison method, we try to analyze the semantic difference by exploring more fine-grained text spans. Specifically, Sqator measures QA scores by measuring the importance of revised spans between junior and senior reports, and outputs the final QA scores by merging all revised span scores. We evaluate Sqator using a collection of 12,013 radiology reports. Experimental results show that Sqator can achieve competitive QA scores. Moreover, the importance scores of revised spans can be also consistent with the judgments of senior doctors.
Authors: Xian Gao, Jiacheng Ruan, Zongyun Zhang, Jingsheng Gao, Ting Liu, Yuzhuo Fu
Abstract: With the rapid growth of academic publications, peer review has become an essential yet time-consuming responsibility within the research community. Large Language Models (LLMs) have increasingly been adopted to assist in the generation of review comments; however, current LLM-based review tasks lack a unified evaluation benchmark to rigorously assess the models' ability to produce comprehensive, accurate, and human-aligned assessments, particularly in scenarios involving multimodal content such as figures and tables. To address this gap, we propose \textbf{MMReview}, a comprehensive benchmark that spans multiple disciplines and modalities. MMReview includes multimodal content and expert-written review comments for 240 papers across 17 research domains within four major academic disciplines: Artificial Intelligence, Natural Sciences, Engineering Sciences, and Social Sciences. We design a total of 13 tasks grouped into four core categories, aimed at evaluating the performance of LLMs and Multimodal LLMs (MLLMs) in step-wise review generation, outcome formulation, alignment with human preferences, and robustness to adversarial input manipulation. Extensive experiments conducted on 16 open-source models and 5 advanced closed-source models demonstrate the thoroughness of the benchmark. We envision MMReview as a critical step toward establishing a standardized foundation for the development of automated peer review systems.
Authors: NVIDIA, :, Aarti Basant, Abhijit Khairnar, Abhijit Paithankar, Abhinav Khattar, Adithya Renduchintala, Aditya Malte, Akhiad Bercovich, Akshay Hazare, Alejandra Rico, Aleksander Ficek, Alex Kondratenko, Alex Shaposhnikov, Alexander Bukharin, Ali Taghibakhshi, Amelia Barton, Ameya Sunil Mahabaleshwarkar, Amy Shen, Andrew Tao, Ann Guan, Anna Shors, Anubhav Mandarwal, Arham Mehta, Arun Venkatesan, Ashton Sharabiani, Ashwath Aithal, Ashwin Poojary, Ayush Dattagupta, Balaram Buddharaju, Banghua Zhu, Barnaby Simkin, Bilal Kartal, Bita Darvish Rouhani, Bobby Chen, Boris Ginsburg, Brandon Norick, Brian Yu, Bryan Catanzaro, Charles Wang, Charlie Truong, Chetan Mungekar, Chintan Patel, Chris Alexiuk, Christian Munley, Christopher Parisien, Dan Su, Daniel Afrimi, Daniel Korzekwa, Daniel Rohrer, Daria Gitman, David Mosallanezhad, Deepak Narayanan, Dima Rekesh, Dina Yared, Dmytro Pykhtar, Dong Ahn, Duncan Riach, Eileen Long, Elliott Ning, Eric Chung, Erick Galinkin, Evelina Bakhturina, Gargi Prasad, Gerald Shen, Haifeng Qian, Haim Elisha, Harsh Sharma, Hayley Ross, Helen Ngo, Herman Sahota, Hexin Wang, Hoo Chang Shin, Hua Huang, Iain Cunningham, Igor Gitman, Ivan Moshkov, Jaehun Jung, Jan Kautz, Jane Polak Scowcroft, Jared Casper, Jian Zhang, Jiaqi Zeng, Jimmy Zhang, Jinze Xue, Jocelyn Huang, Joey Conway, John Kamalu, Jonathan Cohen, Joseph Jennings, Julien Veron Vialard, Junkeun Yi, Jupinder Parmar, Kari Briski, Katherine Cheung, Katherine Luna, Keith Wyss, Keshav Santhanam, Kezhi Kong, Krzysztof Pawelec, Kumar Anik, Kunlun Li, Kushan Ahmadian, Lawrence McAfee, Laya Sleiman, Leon Derczynski, Luis Vega, Maer Rodrigues de Melo, Makesh Narsimhan Sreedhar, Marcin Chochowski, Mark Cai, Markus Kliegl, Marta Stepniewska-Dziubinska, Matvei Novikov, Mehrzad Samadi, Meredith Price, Meriem Boubdir, Michael Boone, Michael Evans, Michal Bien, Michal Zawalski, Miguel Martinez, Mike Chrzanowski, Mohammad Shoeybi, Mostofa Patwary, Namit Dhameja, Nave Assaf, Negar Habibi, Nidhi Bhatia, Nikki Pope, Nima Tajbakhsh, Nirmal Kumar Juluru, Oleg Rybakov, Oleksii Hrinchuk, Oleksii Kuchaiev, Oluwatobi Olabiyi, Pablo Ribalta, Padmavathy Subramanian, Parth Chadha, Pavlo Molchanov, Peter Dykas, Peter Jin, Piotr Bialecki, Piotr Januszewski, Pradeep Thalasta, Prashant Gaikwad, Prasoon Varshney, Pritam Gundecha, Przemek Tredak, Rabeeh Karimi Mahabadi, Rajen Patel, Ran El-Yaniv, Ranjit Rajan, Ria Cheruvu, Rima Shahbazyan, Ritika Borkar, Ritu Gala, Roger Waleffe, Ruoxi Zhang, Russell J. Hewett, Ryan Prenger, Sahil Jain, Samuel Kriman, Sanjeev Satheesh, Saori Kaji, Sarah Yurick, Saurav Muralidharan, Sean Narenthiran, Seonmyeong Bak, Sepehr Sameni, Seungju Han, Shanmugam Ramasamy, Shaona Ghosh, Sharath Turuvekere Sreenivas, Shelby Thomas, Shizhe Diao, Shreya Gopal, Shrimai Prabhumoye, Shubham Toshniwal, Shuoyang Ding, Siddharth Singh, Siddhartha Jain, Somshubra Majumdar, Soumye Singhal, Stefania Alborghetti, Syeda Nahida Akter, Terry Kong, Tim Moon, Tomasz Hliwiak, Tomer Asida, Tony Wang, Tugrul Konuk, Twinkle Vashishth, Tyler Poon, Udi Karpas, Vahid Noroozi, Venkat Srinivasan, Vijay Korthikanti, Vikram Fugro, Vineeth Kalluru, Vitaly Kurin, Vitaly Lavrukhin, Wasi Uddin Ahmad, Wei Du, Wonmin Byeon, Ximing Lu, Xin Dong, Yashaswi Karnati, Yejin Choi, Yian Zhang, Ying Lin, Yonggan Fu, Yoshi Suhara, Zhen Dong, Zhiyu Li, Zhongbo Zhu, Zijia Chen
Abstract: We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.
Authors: Ailing Yu, Lan Yao, Jingnan Liu, Zhe Chen, Jiajun Yin, Yuan Wang, Xinhao Liao, Zhiling Ye, Ji Li, Yun Yue, Hansong Xiao, Hualei Zhou, Chunxiao Guo, Peng Wei, Junwei Liu, Jinjie Gu
Abstract: Recent developments in Large Language Model (LLM)-based agents have shown impressive capabilities spanning multiple domains, exemplified by deep research systems that demonstrate superior performance on complex information-seeking and synthesis tasks. While general-purpose deep research agents have shown impressive capabilities, they struggle significantly with medical domain challenges, as evidenced by leading proprietary systems achieving limited accuracy on complex medical benchmarks. The key limitations are: (1) the model lacks sufficient dense medical knowledge for clinical reasoning, and (2) the framework is constrained by the absence of specialized retrieval tools tailored for medical contexts. We present a medical deep research agent that addresses these challenges through two core innovations. First, we develop a novel data synthesis framework using medical knowledge graphs, extracting the longest chains from subgraphs around rare medical entities to generate complex multi-hop question-answer pairs. Second, we integrate a custom-built private medical retrieval engine alongside general-purpose tools, enabling accurate medical information synthesis. Our approach generates 2100+ diverse trajectories across 12 medical specialties, each averaging 4.2 tool interactions. Through a two-stage training paradigm combining supervised fine-tuning and online reinforcement learning with composite rewards, our MedResearcher-R1-32B model demonstrates exceptional performance, establishing new state-of-the-art results on medical benchmarks while maintaining competitive performance on general deep research tasks. Our work demonstrates that strategic domain-specific innovations in architecture, tool design, and training data construction can enable smaller open-source models to outperform much larger proprietary systems in specialized domains.
Authors: Bolian Li, Yanran Wu, Xinyu Luo, Ruqi Zhang
Abstract: Aligning large language models (LLMs) with human preferences has become a critical step in their development. Recent research has increasingly focused on test-time alignment, where additional compute is allocated during inference to enhance LLM safety and reasoning capabilities. However, these test-time alignment techniques often incur substantial inference costs, limiting their practical application. We are inspired by the speculative sampling acceleration, which leverages a small draft model to efficiently predict future tokens, to address the efficiency bottleneck of test-time alignment. We introduce the reward-shifted speculative sampling (SSS) algorithm, in which the draft model is aligned with human preferences, while the target model remains unchanged. We theoretically demonstrate that the distributional shift between the aligned draft model and the unaligned target model can be exploited to recover the RLHF optimal solution without actually obtaining it, by modifying the acceptance criterion and bonus token distribution. Our algorithm achieves superior gold reward scores at a significantly reduced inference cost in test-time weak-to-strong alignment experiments, thereby validating both its effectiveness and efficiency.
Authors: Huanxuan Liao, Yixing Xu, Shizhu He, Guanchen Li, Xuanwu Yin, Dong Li, Emad Barsoum, Jun Zhao, Kang Liu
Abstract: Long-context inference in large language models (LLMs) is increasingly constrained by the KV cache bottleneck: memory usage grows linearly with sequence length, while attention computation scales quadratically. Existing approaches address this issue by compressing the KV cache along the temporal axis through strategies such as token eviction or merging to reduce memory and computational overhead. However, these methods often neglect fine-grained importance variations across feature dimensions (i.e., the channel axis), thereby limiting their ability to effectively balance efficiency and model accuracy. In reality, we observe that channel saliency varies dramatically across both queries and positions: certain feature channels carry near-zero information for a given query, while others spike in relevance. To address this oversight, we propose SPARK, a training-free plug-and-play method that applies unstructured sparsity by pruning KV at the channel level, while dynamically restoring the pruned entries during attention score computation. Notably, our approach is orthogonal to existing KV compression and quantization techniques, making it compatible for integration with them to achieve further acceleration. By reducing channel-level redundancy, SPARK enables processing of longer sequences within the same memory budget. For sequences of equal length, SPARK not only preserves or improves model accuracy but also reduces KV cache storage by over 30% compared to eviction-based methods. Furthermore, even with an aggressive pruning ratio of 80%, SPARK maintains performance with less degradation than 5% compared to the baseline eviction method, demonstrating its robustness and effectiveness. Our code will be available at https://github.com/Xnhyacinth/SparK.
Authors: Chenghao Liu, Aniket Mahanti, Ranesh Naha, Guanghao Wang, Erwann Sbai
Abstract: As cryptocurrencies gain popularity, the digital asset marketplace becomes increasingly significant. Understanding social media signals offers valuable insights into investor sentiment and market dynamics. Prior research has predominantly focused on text-based platforms such as Twitter. However, video content remains underexplored, despite potentially containing richer emotional and contextual sentiment that is not fully captured by text alone. In this study, we present a multimodal analysis comparing TikTok and Twitter sentiment, using large language models to extract insights from both video and text data. We investigate the dynamic dependencies and spillover effects between social media sentiment and cryptocurrency market indicators. Our results reveal that TikTok's video-based sentiment significantly influences speculative assets and short-term market trends, while Twitter's text-based sentiment aligns more closely with long-term dynamics. Notably, the integration of cross-platform sentiment signals improves forecasting accuracy by up to 20%.
Authors: Hyunjun Kim, Junwoo Ha, Sangyoon Yu, Haon Park
Abstract: LLM-as-a-Judge (LLMaaJ) now underpins scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover a conversation's latent objective and know when that inference is trustworthy? LLMs degrade under irrelevant or long context; multi-turn jailbreaks further hide goals across turns. We introduce ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must return a one-sentence base objective and a self-reported confidence. Accuracy is computed via LLM-judge semantic similarity to gold objectives, converted to binary correctness by a single human-aligned threshold calibrated once on N = 100 items ($\tau^*=0.61$). Metacognition is evaluated with ECE, Brier, Wrong-at-High-Conf, and risk-coverage. Across gpt-4.1, claude-sonnet-4, and Qwen3-235B-A22B-FP8 on SafeMTData_Attack600, SafeMTData_1K, MHJ, and CoSafe, claude-sonnet-4 attains the best objective-extraction accuracy (0.515) and calibration (ECE 0.296; Brier 0.324); gpt-4.1 and Qwen3-235B-A22B-FP8 tie at 0.441 but are overconfident (mean confidence $\approx$0.88 vs. accuracy $\approx$0.44; Wrong-at-0.90 $\approx$48-52%). Performance varies by dataset ($\approx$0.167-0.865). ObjexMT thus supplies an actionable test for LLM judges: when objectives are not explicit, judges often misinfer them with high confidence. We recommend exposing objectives when feasible and gating decisions by confidence otherwise. Code and data at https://github.com/hyunjun1121/ObjexMT_dataset.
Authors: Danny Wang, Ruihong Qiu, Guangdong Bai, Zi Huang
Abstract: Out-of-distribution (OOD) detection remains challenging in text-rich networks, where textual features intertwine with topological structures. Existing methods primarily address label shifts or rudimentary domain-based splits, overlooking the intricate textual-structural diversity. For example, in social networks, where users represent nodes with textual features (name, bio) while edges indicate friendship status, OOD may stem from the distinct language patterns between bot and normal users. To address this gap, we introduce the TextTopoOOD framework for evaluating detection across diverse OOD scenarios: (1) attribute-level shifts via text augmentations and embedding perturbations; (2) structural shifts through edge rewiring and semantic connections; (3) thematically-guided label shifts; and (4) domain-based divisions. Furthermore, we propose TNT-OOD to model the complex interplay between Text aNd Topology using: 1) a novel cross-attention module to fuse local structure into node-level text representations, and 2) a HyperNetwork to generate node-specific transformation parameters. This aligns topological and semantic features of ID nodes, enhancing ID/OOD distinction across structural and textual shifts. Experiments on 11 datasets across four OOD scenarios demonstrate the nuanced challenge of TextTopoOOD for evaluating OOD detection in text-rich networks.
Authors: Changsong Liu, Yizhou Peng, Eng Siong Chng
Abstract: Contextual automatic speech recognition (ASR) systems allow for recognizing out-of-vocabulary (OOV) words, such as named entities or rare words. However, it remains challenging due to limited training data and ambiguous or inconsistent pronunciations. In this paper, we propose a synthesis-driven multi-pronunciation contextual biasing method that performs zero-shot contextual ASR on a pretrained Whisper model. Specifically, we leverage text-to-speech (TTS) systems to synthesize diverse speech samples containing each target rare word, and then use the pretrained Whisper model to extract multiple predicted pronunciation variants. These variant token sequences are compiled into a prefix-trie, which assigns rewards to beam hypotheses in a shallow-fusion manner during beam-search decoding. Subsequently, any recognized variant is mapped back to the original rare word in the final transcription. The evaluation results on the LibriSpeech dataset show that our method reduces biased-word error rate (B-WER) by 43% on test-clean and 44% on test-other while maintaining unbiased-WER (U-WER) essentially unchanged.
Authors: Bojan Evkoski, Igor Mozeti\v{c}, Nikola Ljube\v{s}i\'c, Petra Kralj Novak
Abstract: Affective polarization, characterized by increased negativity and hostility towards opposing groups, has become a prominent feature of political discourse worldwide. Our study examines the presence of this type of polarization in a selection of European parliaments in a fully automated manner. Utilizing a comprehensive corpus of parliamentary speeches from the parliaments of six European countries, we employ natural language processing techniques to estimate parliamentarian sentiment. By comparing the levels of negativity conveyed in references to individuals from opposing groups versus one's own, we discover patterns of affectively polarized interactions. The findings demonstrate the existence of consistent affective polarization across all six European parliaments. Although activity correlates with negativity, there is no observed difference in affective polarization between less active and more active members of parliament. Finally, we show that reciprocity is a contributing mechanism in affective polarization between parliamentarians across all six parliaments.
Authors: Yilin Wang, Heng Wang, Yuyang Bai, Minnan Luo
Abstract: In Large Language Models (LLMs) generation, there exist knowledge conflicts and scenarios where parametric knowledge contradicts knowledge provided in the context. Previous works studied tuning, decoding algorithms, or locating and editing context-aware neurons to adapt LLMs to be faithful to new contextual knowledge. However, they are usually inefficient or ineffective for large models, not workable for black-box models, or unable to continuously adjust LLMs' sensitivity to the knowledge provided in the context. To mitigate these problems, we propose CSKS (Continuously Steering Knowledge Sensitivity), a simple framework that can steer LLMs' sensitivity to contextual knowledge continuously at a lightweight cost. Specifically, we tune two small LMs (i.e. proxy models) and use the difference in their output distributions to shift the original distribution of an LLM without modifying the LLM weights. In the evaluation process, we not only design synthetic data and fine-grained metrics to measure models' sensitivity to contextual knowledge but also use a real conflict dataset to validate CSKS's practical efficacy. Extensive experiments demonstrate that our framework achieves continuous and precise control over LLMs' sensitivity to contextual knowledge, enabling both increased sensitivity and reduced sensitivity, thereby allowing LLMs to prioritize either contextual or parametric knowledge as needed flexibly. Our data and code are available at https://github.com/OliveJuiceLin/CSKS.
Authors: Venkatesh Mishra, Amir Saeidi, Satyam Raj, Mutsumi Nakamura, Jayanth Srinivasa, Gaowen Liu, Ali Payani, Chitta Baral
Abstract: Recent advances in reasoning and planning capabilities of large language models (LLMs) have enabled their potential as autonomous agents capable of tool use in dynamic environments. However, in multi-turn conversational environments like $\tau$-bench, these agents often struggle with consistent reasoning, adherence to domain-specific policies, and extracting correct information over a long horizon of tool-calls and conversation. To capture and mitigate these failures, we conduct a comprehensive manual analysis of the common errors occurring in the conversation trajectories. We then experiment with reformulations of inputs to the tool-calling agent for improvement in agent decision making. Finally, we propose the Input-Reformulation Multi-Agent (IRMA) framework, which automatically reformulates user queries augmented with relevant domain rules and tool suggestions for the tool-calling agent to focus on. The results show that IRMA significantly outperforms ReAct, Function Calling, and Self-Reflection by 16.1%, 12.7%, and 19.1%, respectively, in overall pass^5 scores. These findings highlight the superior reliability and consistency of IRMA compared to other methods in dynamic environments.
Authors: Ramazan Ali Bahrami, Ramin Yahyapour
Abstract: Sentential relation extraction (RE) is an important task in natural language processing (NLP). In this paper we propose to do sentential RE with dynamic routing in capsules. We first show that the proposed approach outperform state of the art on common sentential relation extraction datasets Tacred, Tacredrev, Retacred, and Conll04. We then investigate potential reasons for its good performance on the mentioned datasets, and yet low performance on another similar, yet larger sentential RE dataset, Wikidata. As such, we identify noise in Wikidata labels as one of the reasons that can hinder performance. Additionally, we show associativity of better performance with better re-representation, a term from neuroscience referred to change of representation in human brain to improve the match at comparison time. As example, in the given analogous terms King:Queen::Man:Woman, at comparison time, and as a result of re-representation, the similarity between related head terms (King,Man), and tail terms (Queen,Woman) increases. As such, our observation show that our proposed model can do re-representation better than the vanilla model compared with. To that end, beside noise in the labels of the distantly supervised RE datasets, we propose re-representation as a challenge in sentential RE.
Authors: William Jurayj, Nils Holzenberger, Benjamin Van Durme
Abstract: According to the United States Internal Revenue Service, ''the average American spends $\$270$ and 13 hours filing their taxes''. Even beyond the U.S., tax filing requires complex reasoning, combining application of overlapping rules with numerical calculations. Because errors can incur costly penalties, any automated system must deliver high accuracy and auditability, making modern large language models (LLMs) poorly suited for this task. We propose an approach that integrates LLMs with a symbolic solver to calculate tax obligations. We evaluate variants of this system on the challenging StAtutory Reasoning Assessment (SARA) dataset, and include a novel method for estimating the cost of deploying such a system based on real-world penalties for tax errors. We further show how combining up-front translation of plain-text rules into formal logic programs, combined with intelligently retrieved exemplars for formal case representations, can dramatically improve performance on this task and reduce costs to well below real-world averages. Our results demonstrate the promise and economic feasibility of neuro-symbolic architectures for increasing equitable access to reliable tax assistance.
Authors: Muskan Saraf, Sajjad Rezvani Boroujeni, Justin Beaudry, Hossein Abedi, Tom Bush
Abstract: Large language models (LLMs) are increasingly used to evaluate outputs, yet their judgments may be influenced. This study examines bias in self- and cross-model evaluations by ChatGPT, Gemini, and Claude under four conditions: no labels, true labels, and two false-label scenarios. Blog posts authored by each model were evaluated by all three using both overall preference voting and quality ratings for Coherence, Informativeness, and Conciseness, with all scores expressed as percentages for direct comparison. Results reveal striking asymmetries: the "Claude" label consistently boosts scores, while the "Gemini" label consistently depresses them, regardless of actual content. False labels frequently reversed rankings, producing shifts of up to 50 percentage points in preference votes and up to 12 percentage points in converted quality ratings. Gemini's self-scores collapsed under true labels, while Claude's self-preference intensified. These findings show that perceived model identity can heavily distort high-level judgments and subtly influence detailed quality ratings, underscoring the need for blind or multimodel evaluation protocols to ensure fairness in LLM benchmarking.
Authors: Khalid M. Saqr
Abstract: Despite rapid gains in scale, research evaluation still relies on opaque, lagging proxies. To serve the scientific community, we pursue transparency: reproducible, auditable epistemic classification useful for funding and policy. Here we formalize KGX3 as a scenario-based model for mapping Kuhnian stages from research papers, prove determinism of the classification pipeline, and define the epistemic manifold that yields paradigm maps. We report validation across recent corpora, operational complexity at global scale, and governance that preserves interpretability while protecting core IP. The system delivers early, actionable signals of drift, crisis, and shift unavailable to citation metrics or citations-anchored NLP. KGX3 is the latest iteration of a deterministic epistemic engine developed since 2019, originating as Soph.io (2020), advanced as iKuhn (2024), and field-tested through Preprint Watch in 2025.
Authors: Haojun Jiang, Jianke Zhang, Rui Huang, Chunjiang Ge, Zanlin Ni, Shiji Song, Gao Huang
Abstract: Vision-language retrieval is an important multi-modal learning topic, where the goal is to retrieve the most relevant visual candidate for a given text query. Recently, pre-trained models, e.g., CLIP, show great potential on retrieval tasks. However, as pre-trained models are scaling up, fully fine-tuning them on donwstream retrieval datasets has a high risk of overfitting. Moreover, in practice, it would be costly to train and store a large model for each task. To overcome the above issues, we present a novel Cross-Modal Adapter for parameter-efficient transfer learning. Inspired by adapter-based methods, we adjust the pre-trained model with a few parameterization layers. However, there are two notable differences. First, our method is designed for the multi-modal domain. Secondly, it allows encoder-level implicit cross-modal interactions between vision and language encoders. Although surprisingly simple, our approach has three notable benefits: (1) reduces the vast majority of fine-tuned parameters, (2) saves training time, and (3) allows all the pre-trained parameters to be fixed, enabling the pre-trained model to be shared across datasets. Extensive experiments demonstrate that, without bells and whistles, our approach outperforms adapter-based methods on image-text retrieval datasets (MSCOCO, Flickr30K) and video-text retrieval datasets (MSR-VTT, DiDeMo, and ActivityNet).
Authors: Xueying Du, Mingwei Liu, Hanlin Wang, Juntao Li, Xin Peng, Yiling Lou
Abstract: Software crash bugs cause unexpected program behaviors or even abrupt termination, thus demanding immediate resolution. However, resolving crash bugs can be challenging due to their complex root causes, which can originate from issues in the source code or external factors like third-party library dependencies. Large language models (LLMs) have shown promise in software engineering tasks. However, existing research predominantly focuses on the capability of LLMs to localize and repair code-related crash bugs, leaving their effectiveness in resolving environment-related crash bugs in real-world software unexplored. To fill this gap, we conducted the first comprehensive study to assess the capability of LLMs in resolving real-world environment-related crash bugs. We first systematically compare LLMs' performance in resolving code-related and environment-related crash bugs with varying levels of crash contextual information. Our findings reveal that localization is the primary challenge for resolving code-related crashes, while repair poses a greater challenge for environment-related crashes. Furthermore, we investigate the impact of different prompt strategies on improving the resolution of environment-related crash bugs, incorporating different prompt templates and multi-round interactions. Building on this, we further explore an advanced active inquiry prompting strategy leveraging the self-planning capabilities of LLMs. Based on these explorations, we propose IntDiagSolver, an interactive methodology designed to enable precise crash bug resolution through ongoing engagement with LLMs. Extensive evaluations of IntDiagSolver across multiple LLMs (including GPT-3.5, GPT-4, Claude, CodeLlama, DeepSeek-R1, and Qwen-3-Coder) demonstrate consistent improvements in resolution accuracy, with substantial enhancements ranging from 9.1% to 43.3% in localization and 9.1% to 53.3% in repair.
Authors: Mladjan Jovanovic, Peter Voss
Abstract: Incremental learning is the ability of systems to acquire knowledge over time, enabling their adaptation and generalization to novel tasks. It is a critical ability for intelligent, real-world systems, especially when data changes frequently or is limited. This review provides a comprehensive analysis of incremental learning in Large Language Models. It synthesizes the state-of-the-art incremental learning paradigms, including continual learning, meta-learning, parameter-efficient learning, and mixture-of-experts learning. We demonstrate their utility for incremental learning by describing specific achievements from these related topics and their critical factors. An important finding is that many of these approaches do not update the core model, and none of them update incrementally in real-time. The paper highlights current problems and challenges for future research in the field. By consolidating the latest relevant research developments, this review offers a comprehensive understanding of incremental learning and its implications for designing and developing LLM-based learning systems.
Authors: Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, Irwin King
Abstract: Embodied AI is widely recognized as a key element of artificial general intelligence because it involves controlling embodied agents to perform tasks in the physical world. Building on the success of large language models and vision-language models, a new category of multimodal models -- referred to as vision-language-action models (VLAs) -- has emerged to address language-conditioned robotic tasks in embodied AI by leveraging their distinct ability to generate actions. In recent years, a myriad of VLAs have been developed, making it imperative to capture the rapidly evolving landscape through a comprehensive survey. To this end, we present the first survey on VLAs for embodied AI. This work provides a detailed taxonomy of VLAs, organized into three major lines of research. The first line focuses on individual components of VLAs. The second line is dedicated to developing control policies adept at predicting low-level actions. The third line comprises high-level task planners capable of decomposing long-horizon tasks into a sequence of subtasks, thereby guiding VLAs to follow more general user instructions. Furthermore, we provide an extensive summary of relevant resources, including datasets, simulators, and benchmarks. Finally, we discuss the challenges faced by VLAs and outline promising future directions in embodied AI. We have created a project associated with this survey, which is available at https://github.com/yueen-ma/Awesome-VLA.
Authors: Yang Zhang, Shixin Yang, Chenjia Bai, Fei Wu, Xiu Li, Zhen Wang, Xuelong Li
Abstract: Grounding the reasoning ability of large language models (LLMs) for embodied tasks is challenging due to the complexity of the physical world. Especially, LLM planning for multi-agent collaboration requires communication of agents or credit assignment as the feedback to re-adjust the proposed plans and achieve effective coordination. However, existing methods that overly rely on physical verification or self-reflection suffer from excessive and inefficient querying of LLMs. In this paper, we propose a novel framework for multi-agent collaboration that introduces Reinforced Advantage feedback (ReAd) for efficient self-refinement of plans. Specifically, we perform critic regression to learn a sequential advantage function from LLM-planned data, and then treat the LLM planner as an optimizer to generate actions that maximize the advantage function. It endows the LLM with the foresight to discern whether the action contributes to accomplishing the final task. We provide theoretical analysis by extending advantage-weighted regression in reinforcement learning to multi-agent systems. Experiments on Overcooked-AI and a difficult variant of RoCoBench show that ReAd surpasses baselines in success rate, and also significantly decreases the interaction steps of agents and query rounds of LLMs, demonstrating its high efficiency for grounding LLMs. More results are given at https://read-llm.github.io.
Authors: Jinyuan Li, Ziyan Li, Han Li, Jianfei Yu, Rui Xia, Di Sun, Gang Pan
Abstract: Grounded Multimodal Named Entity Recognition (GMNER) task aims to identify named entities, entity types and their corresponding visual regions. GMNER task exhibits two challenging attributes: 1) The tenuous correlation between images and text on social media contributes to a notable proportion of named entities being ungroundable. 2) There exists a distinction between coarse-grained noun phrases used in similar tasks (e.g., phrase localization) and fine-grained named entities. In this paper, we propose RiVEG, a unified framework that reformulates GMNER into a joint MNER-VE-VG task by leveraging large language models (LLMs) as connecting bridges. This reformulation brings two benefits: 1) It enables us to optimize the MNER module for optimal MNER performance and eliminates the need to pre-extract region features using object detection methods, thus naturally addressing the two major limitations of existing GMNER methods. 2) The introduction of Entity Expansion Expression module and Visual Entailment (VE) module unifies Visual Grounding (VG) and Entity Grounding (EG). This endows the proposed framework with unlimited data and model scalability. Furthermore, to address the potential ambiguity stemming from the coarse-grained bounding box output in GMNER, we further construct the new Segmented Multimodal Named Entity Recognition (SMNER) task and corresponding Twitter-SMNER dataset aimed at generating fine-grained segmentation masks, and experimentally demonstrate the feasibility and effectiveness of using box prompt-based Segment Anything Model (SAM) to empower any GMNER model with the ability to accomplish the SMNER task. Extensive experiments demonstrate that RiVEG significantly outperforms SoTA methods on four datasets across the MNER, GMNER, and SMNER tasks.
Authors: Zhengyu Hu, Linxin Song, Jieyu Zhang, Zheyuan Xiao, Tianfu Wang, Zhengyu Chen, Nicholas Jing Yuan, Jianxun Lian, Kaize Ding, Hui Xiong
Abstract: The use of large language models (LLMs) as judges, particularly in preference comparisons, has become widespread, but this reveals a notable bias towards longer responses, undermining the reliability of such evaluations. To better understand such bias, we propose to decompose the preference evaluation metric, specifically the win rate, into two key components: desirability and information mass, where the former is length-independent and related to trustworthiness such as correctness, toxicity, and consistency, and the latter is length-dependent and represents the amount of information in the response. We empirically demonstrated the decomposition through controlled experiments and found that response length impacts evaluations by influencing information mass. To derive a reliable evaluation metric that assesses content quality without being confounded by response length, we propose AdapAlpaca, a simple yet effective adjustment to win rate measurement. Specifically, AdapAlpaca ensures a fair comparison of response quality by aligning the lengths of reference and test model responses under equivalent length intervals.
Authors: Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, Carlos Guestrin
Abstract: Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden states. We present a practical framework for instantiating sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Similar to Transformer, TTT-Linear and TTT-MLP can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.
Authors: Haoyu Tang, Ye Liu, Xi Zhao, Xukai Liu, Yanghai Zhang, Kai Zhang, Xiaofang Zhou, Enhong Chen
Abstract: Recent advances in machine learning, particularly in Natural Language Processing (NLP), have produced powerful models trained on vast datasets. However, these models risk leaking sensitive information, raising privacy concerns. In response, regulatory measures such as the European Union's General Data Protection Regulation (GDPR) have driven increasing interest in Machine Unlearning techniques, which enable models to selectively forget specific data entries. Early unlearning approaches primarily relied on pre-processing methods, while more recent research has shifted towards training-based solutions. Despite their effectiveness, a key limitation persists: most methods require access to original training data, which is often unavailable. Additionally, directly applying unlearning techniques bears the cost of undermining the model's expressive capabilities. To address these challenges, we introduce the Iterative Contrastive Unlearning (ICU) framework, which consists of three core components: A Knowledge Unlearning Induction module designed to target specific knowledge for removal using an unlearning loss; A Contrastive Learning Enhancement module to preserve the model's expressive capabilities against the pure unlearning goal; And an Iterative Unlearning Refinement module that dynamically adjusts the unlearning process through ongoing evaluation and updates. Experimental results demonstrate the efficacy of our ICU method in unlearning sensitive information while maintaining the model's overall performance, offering a promising solution for privacy-conscious machine learning applications.
Authors: Yi Liu, Junzhe Yu, Huijia Sun, Ling Shi, Gelei Deng, Yuqi Chen, Yang Liu
Abstract: Large language models (LLMs) like ChatGPT and Gemini have significantly advanced natural language processing, enabling various applications such as chatbots and automated content generation. However, these models can be exploited by malicious individuals who craft toxic prompts to elicit harmful or unethical responses. These individuals often employ jailbreaking techniques to bypass safety mechanisms, highlighting the need for robust toxic prompt detection methods. Existing detection techniques, both blackbox and whitebox, face challenges related to the diversity of toxic prompts, scalability, and computational efficiency. In response, we propose ToxicDetector, a lightweight greybox method designed to efficiently detect toxic prompts in LLMs. ToxicDetector leverages LLMs to create toxic concept prompts, uses embedding vectors to form feature vectors, and employs a Multi-Layer Perceptron (MLP) classifier for prompt classification. Our evaluation on various versions of the LLama models, Gemma-2, and multiple datasets demonstrates that ToxicDetector achieves a high accuracy of 96.39\% and a low false positive rate of 2.00\%, outperforming state-of-the-art methods. Additionally, ToxicDetector's processing time of 0.0780 seconds per prompt makes it highly suitable for real-time applications. ToxicDetector achieves high accuracy, efficiency, and scalability, making it a practical method for toxic prompt detection in LLMs.
Authors: Hangfeng He, Weijie J. Su
Abstract: Large language models (LLMs) have been widely employed across various application domains, yet their black-box nature poses significant challenges to understanding how these models process input data internally to make predictions. In this paper, we introduce a precise and quantitative law that governs the learning of contextualized token embeddings through intermediate layers in pre-trained LLMs for next-token prediction. Our findings reveal that each layer contributes equally to enhancing prediction accuracy, from the lowest to the highest layer -- a universal phenomenon observed across a diverse array of open-source LLMs, irrespective of their architectures or pre-training data. We demonstrate that this law offers new perspectives and actionable insights to inform and guide practices in LLM development and applications, including model scaling, pre-training tasks, and interpretation.
Authors: Zhaowei Wang, Hongming Zhang, Tianqing Fang, Ye Tian, Yue Yang, Kaixin Ma, Xiaoman Pan, Yangqiu Song, Dong Yu
Abstract: Large Vision-Language Models (LVLMs) have achieved significant progress in tasks like visual question answering and document understanding. However, their potential to comprehend embodied environments and navigate within them remains underexplored. In this work, we first study the challenge of open-vocabulary object navigation by introducing DivScene, a large-scale dataset with 4,614 houses across 81 scene types and 5,707 kinds of target objects. Our dataset provides a much greater diversity of target objects and scene types than existing datasets, enabling a comprehensive task evaluation. We evaluated various methods with LVLMs and LLMs on our dataset and found that current models still fall short of open-vocab object navigation ability. Then, we fine-tuned LVLMs to predict the next action with CoT explanations. We observe that LVLM's navigation ability can be improved substantially with only BFS-generated shortest paths without any human supervision, surpassing GPT-4o by over 20% in success rates.
Authors: Yunkai Dang, Mengxi Gao, Yibo Yan, Xin Zou, Yanggan Gu, Jungang Li, Jingyu Wang, Peijie Jiang, Aiwei Liu, Jia Liu, Xuming Hu
Abstract: Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual-textual misalignment, leaving largely unexplored the MLLMs' ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model's original response on unperturbed inputs; (2) inject explicit (false-answer hints) and implicit (contextual contradictions) misleading instructions, and compute the misleading rate - the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image-question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks. Our code is available at https://github.com/Yunkaidang/uncertainty
Authors: Omar Naim, Guilhem Fouilh\'e, Nicholas Asher
Abstract: In-context learning (ICL) has emerged as a powerful paradigm for easily adapting Large Language Models (LLMs) to various tasks. However, our understanding of how ICL works remains limited. We explore a simple model of ICL in a controlled setup with synthetic training data to investigate ICL of univariate linear functions. We experiment with a range of GPT-2-like transformer models trained from scratch. Our findings challenge the prevailing narrative that transformers adopt algorithmic approaches like linear regression to learn a linear function in-context. These models fail to generalize beyond their training distribution, highlighting fundamental limitations in their capacity to infer abstract task structures. Our experiments lead us to propose a mathematically precise hypothesis of what the model might be learning.
Authors: Aritra Kumar Lahiri, Qinmin Vivian Hu
Abstract: Recent advancements in generative AI have fostered the development of highly adept Large Language Models (LLMs) that integrate diverse data types to empower decision-making. Among these, multimodal retrieval-augmented generation (RAG) applications are promising because they combine the strengths of information retrieval and generative models, enhancing their utility across various domains, including clinical use cases. This paper introduces AlzheimerRAG, a Multimodal RAG application for clinical use cases, primarily focusing on Alzheimer's Disease case studies from PubMed articles. This application incorporates cross-modal attention fusion techniques to integrate textual and visual data processing by efficiently indexing and accessing vast amounts of biomedical literature. Our experimental results, compared to benchmarks such as BioASQ and PubMedQA, have yielded improved performance in the retrieval and synthesis of domain-specific information. We also present a case study using our multimodal RAG in various Alzheimer's clinical scenarios. We infer that AlzheimerRAG can generate responses with accuracy non-inferior to humans and with low rates of hallucination.
Authors: Gabrielle Littlefair, Niladri Shekhar Dutt, Niloy J. Mitra
Abstract: Interior design involves the careful selection and arrangement of objects to create an aesthetically pleasing, functional, and harmonized space that aligns with the client's design brief. This task is particularly challenging, as a successful design must not only incorporate all the necessary objects in a cohesive style, but also ensure they are arranged in a way that maximizes accessibility, while adhering to a variety of affordability and usage considerations. Data-driven solutions have been proposed, but these are typically room- or domain-specific and lack explainability in their design design considerations used in producing the final layout. In this paper, we investigate if large language models (LLMs) can be directly utilized for interior design. While we find that LLMs are not yet capable of generating complete layouts, they can be effectively leveraged in a structured manner, inspired by the workflow of interior designers. By systematically probing LLMs, we can reliably generate a list of objects along with relevant constraints that guide their placement. We translate this information into a design layout graph, which is then solved using an off-the-shelf constrained optimization setup to generate the final layouts. We benchmark our algorithm in various design configurations against existing LLM-based methods and human designs, and evaluate the results using a variety of quantitative and qualitative metrics along with user studies. In summary, we demonstrate that LLMs, when used in a structured manner, can effectively generate diverse high-quality layouts, making them a viable solution for creating large-scale virtual scenes. Project webpage at https://flairgpt.github.io/
Authors: Ruixiang Jiang, Changwen Chen
Abstract: The rapid technical progress of generative art (GenArt) has democratized the creation of visually appealing imagery. However, achieving genuine artistic impact - the kind that resonates with viewers on a deeper, more meaningful level - remains formidable as it requires a sophisticated aesthetic sensibility. This sensibility involves a multifaceted cognitive process extending beyond mere visual appeal, which is often overlooked by current computational methods. This paper pioneers an approach to capture this complex process by investigating how the reasoning capabilities of Multimodal LLMs (MLLMs) can be effectively elicited to perform aesthetic judgment. Our analysis reveals a critical challenge: MLLMs exhibit a tendency towards hallucinations during aesthetic reasoning, characterized by subjective opinions and unsubstantiated artistic interpretations. We further demonstrate that these hallucinations can be suppressed by employing an evidence-based and objective reasoning process, as substantiated by our proposed baseline, ArtCoT. MLLMs prompted by this principle produce multifaceted, in-depth aesthetic reasoning that aligns significantly better with human judgment. These findings have direct applications in areas such as AI art tutoring and as reward models for image generation. Ultimately, we hope this work paves the way for AI systems that can truly understand, appreciate, and contribute to art that aligns with human aesthetic values. Project homepage: https://github.com/songrise/MLLM4Art.
Authors: Rui Li, Xiaohan Wang, Yuhui Zhang, Orr Zohar, Zeyu Wang, Serena Yeung-Levy
Abstract: Despite significant advancements in video large multimodal models (video-LMMs), achieving effective temporal grounding in long-form videos remains a challenge for existing models. To address this limitation, we propose Temporal Preference Optimization (TPO), a novel post-training framework designed to enhance the temporal grounding capabilities of video-LMMs through preference learning. TPO adopts a self-training approach that enables models to differentiate between well-grounded and less accurate temporal responses by leveraging curated preference datasets at two granularities: localized temporal grounding, which focuses on specific video segments, and comprehensive temporal grounding, which captures extended temporal dependencies across entire video sequences. By optimizing on these preference datasets, TPO significantly enhances temporal understanding while reducing reliance on manually annotated data. Extensive experiments on three long-form video understanding benchmarks--LongVideoBench, MLVU, and Video-MME--demonstrate the effectiveness of TPO across two state-of-the-art video-LMMs. Notably, LLaVA-Video-TPO establishes itself as the leading 7B model on the Video-MME benchmark, underscoring the potential of TPO as a scalable and efficient solution for advancing temporal reasoning in long-form video understanding. Project page: https://ruili33.github.io/tpo_website.
Authors: Jiefeng Chen, Jie Ren, Xinyun Chen, Chengrun Yang, Ruoxi Sun, Jinsung Yoon, Sercan \"O Ar{\i}k
Abstract: Recent advancements in Large Language Models (LLMs) have created new opportunities to enhance performance on complex reasoning tasks by leveraging test-time computation. However, existing scaling methods have key limitations: parallel methods like repeated sampling are often inefficient and quickly saturate, while sequential methods like SELF-REFINE struggle to improve after a few rounds. Although combining these approaches shows promise, current methods require fine-tuned reward and revision models. This paper proposes Self-Enhanced Test-Time Scaling (SETS), a simple yet effective approach that overcomes these limitations by strategically combining parallel and sequential techniques and fully leveraging LLMs' self-improvement abilities. SETS exploits the inherent self-verification and self-correction capabilities of LLMs, unifying sampling, verification, and correction within a single framework. This facilitates efficient and scalable test-time computation for enhanced performance on complex tasks without any model training. Our comprehensive experimental results on challenging benchmarks spanning planning, reasoning, math, and coding demonstrate that SETS achieves significant performance improvements and more advantageous test-time scaling behavior than the alternatives.
Authors: Kemal Kurniawan, Meladel Mistica, Timothy Baldwin, Jey Han Lau
Abstract: Human label variation (HLV) challenges the standard assumption that a labelled instance has a single ground truth, instead embracing the natural variation in human annotation to train and evaluate models. While various training methods and metrics for HLV have been proposed, it is still unclear which methods and metrics perform best in what settings. We propose new evaluation metrics for HLV leveraging fuzzy set theory. Since these new proposed metrics are differentiable, we then in turn experiment with employing these metrics as training objectives. We conduct an extensive study over 6 HLV datasets testing 14 training methods and 6 evaluation metrics. We find that training on either disaggregated annotations or soft labels performs best across metrics, outperforming training using the proposed training objectives with differentiable metrics. We also show that our proposed soft micro F1 score is one of the best metrics for HLV data.
Authors: Euntae Choi, Sumin Song, Woosang Lim, Sungjoo Yoo
Abstract: We propose Rotate, Clip, and Partition (RCP), a quantization-aware training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4(2-bit weight, 4-bit activation, and 4-bit KV cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design, by quantitatively analyzing the impact of random rotation on 2-bit weight quantization. Our weight quantizer features Learnable Direct Partitioning (LDP), which introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a specialized GPU kernel that supports GEMV on non-uniform W2A4. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 ppl and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code is available at https://github.com/ songsm921/RCP.
URLs: https://github.com/
Authors: Tianmi Ma, Jiawei Du, Wenxin Huang, Wenjie Wang, Liang Xie, Xian Zhong, Joey Tianyi Zhou
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in natural language tasks, yet their performance in dynamic, real-world financial environments remains underexplored. Existing approaches are limited to historical backtesting, where trading actions cannot influence market prices and agents train only on static data. To address this limitation, we present the Agent Trading Arena, a virtual zero-sum stock market in which LLM-based agents engage in competitive multi-agent trading and directly impact price dynamics. By simulating realistic bid-ask interactions, our platform enables training in scenarios that closely mirror live markets, thereby narrowing the gap between training and evaluation. Experiments reveal that LLMs struggle with numerical reasoning when given plain-text data, often overfitting to local patterns and recent values. In contrast, chart-based visualizations significantly enhance both numerical reasoning and trading performance. Furthermore, incorporating a reflection module yields additional improvements, especially with visual inputs. Evaluations on NASDAQ and CSI datasets demonstrate the superiority of our method, particularly under high volatility. All code and data are available at https://github.com/wekjsdvnm/Agent-Trading-Arena.
Authors: Yizheng Sun, Hao Li, Chang Xu, Hongpeng Zhou, Chenghua Lin, Riza Batista-Navarro, Jingyuan Sun
Abstract: Vision-Language Models (VLMs) are powerful yet computationally intensive for widespread practical deployments. To address such challenge without costly re-training, post-training acceleration techniques like quantization and token reduction are extensively explored. However, current acceleration evaluations primarily target minimal overall performance degradation, overlooking a crucial question: does the accelerated model still give the same answers to the same questions as it did before acceleration? This is vital for stability-centered industrial applications where consistently correct answers for specific, known situations are paramount, such as in AI-based disease diagnosis. We systematically investigate this for accelerated VLMs, testing four leading models (LLaVA-1.5, LLaVA-Next, Qwen2-VL, Qwen2.5-VL) with eight acceleration methods on ten multi-modal benchmarks. Our findings are stark: despite minimal aggregate performance drops, accelerated models changed original answers up to 20% of the time. Critically, up to 6.5% of these changes converted correct answers to incorrect. Input perturbations magnified these inconsistencies, and the trend is confirmed by case studies with the medical VLM LLaVA-Med. This research reveals a significant oversight in VLM acceleration, stressing an urgent need for instance-level stability checks to ensure trustworthy real-world deployment.
Authors: Jiale Cheng, Ruiliang Lyu, Xiaotao Gu, Xiao Liu, Jiazheng Xu, Yida Lu, Jiayan Teng, Zhuoyi Yang, Yuxiao Dong, Jie Tang, Hongning Wang, Minlie Huang
Abstract: Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO.
Authors: Lang Cao, Renhong Chen, Yingtian Zou, Chao Peng, Wu Ning, Huacong Xu, Qian Chen, Yuxian Wang, Peishuo Su, Mofan Peng, Zijie Chen, Yitong Li
Abstract: We introduce the Entropy Driven Uncertainty Process Reward Model (EDU-PRM), a novel entropy-driven training framework for process reward modeling that enables dynamic, uncertainty-aligned segmentation of complex reasoning steps, eliminating the need for costly manual step annotations. Unlike previous Process Reward Models (PRMs) that rely on static partitioning and human labeling, EDU-PRM automatically anchors step boundaries at tokens with high predictive entropy. On the MATH test set, EDU-PRM achieves 65.5% accuracy, surpassing strong public PRM baselines such as Math-Shepherd PRM (61.7%) and Omega PRM (62.4%) under the High Temperature (HT) Sample + BON setting. Furthermore, when replacing HT sampling with EDU sampling, EDU-PRM further improves both accuracy and efficiency: at N=64, accuracy increases from 64.7% (HT Sample + BON) to 67.3% (EDU Sample + BON), while the number of generated tokens is reduced by 47%, demonstrating a superior accuracy-cost balance. On the ProcessBench test set, EDU-PRM achieves a new state-of-the-art accuracy of 88.4% using less than 1.5% of the Qwen2.5-Math-PRM-72B training data, surpassing the previous best of 87.8%. In summary, EDU-PRM provides a scalable and annotation-efficient paradigm for process supervision in mathematical reasoning, opening new avenues for efficient complex reasoning on math.
Authors: Manuel Scheibl, Birte Richter, Alissa M\"uller, Michael Beetz, Britta Wrede
Abstract: This research addresses the question, which characteristics a cognitive architecture must have to leverage the benefits of natural language in Co-Constructive Task Learning (CCTL). To provide context, we first discuss Interactive Task Learning (ITL), the mechanisms of the human memory system, and the significance of natural language and multi-modality. Next, we examine the current state of cognitive architectures, analyzing their capabilities to inform a concept of CCTL grounded in multiple sources. We then integrate insights from various research domains to develop a unified framework. Finally, we conclude by identifying the remaining challenges and requirements necessary to achieve CCTL in Human-Robot Interaction (HRI).
Authors: Antoun Yaacoub, J\'er\^ome Da-Rugna, Zainab Assaghir
Abstract: This study evaluates the integration of Bloom's Taxonomy into OneClickQuiz, an Artificial Intelligence (AI) driven plugin for automating Multiple-Choice Question (MCQ) generation in Moodle. Bloom's Taxonomy provides a structured framework for categorizing educational objectives into hierarchical cognitive levels. Our research investigates whether incorporating this taxonomy can improve the alignment of AI-generated questions with specific cognitive objectives. We developed a dataset of 3691 questions categorized according to Bloom's levels and employed various classification models-Multinomial Logistic Regression, Naive Bayes, Linear Support Vector Classification (SVC), and a Transformer-based model (DistilBERT)-to evaluate their effectiveness in categorizing questions. Our results indicate that higher Bloom's levels generally correlate with increased question length, Flesch-Kincaid Grade Level (FKGL), and Lexical Density (LD), reflecting the increased complexity of higher cognitive demands. Multinomial Logistic Regression showed varying accuracy across Bloom's levels, performing best for "Knowledge" and less accurately for higher-order levels. Merging higher-level categories improved accuracy for complex cognitive tasks. Naive Bayes and Linear SVC also demonstrated effective classification for lower levels but struggled with higher-order tasks. DistilBERT achieved the highest performance, significantly improving classification of both lower and higher-order cognitive levels, achieving an overall validation accuracy of 91%. This study highlights the potential of integrating Bloom's Taxonomy into AI-driven assessment tools and underscores the advantages of advanced models like DistilBERT for enhancing educational content generation.
Authors: Daniel A. P. Oliveira, David Martins de Matos
Abstract: Visual storytelling systems struggle to maintain character identity across frames and link actions to appropriate subjects, frequently leading to referential hallucinations. These issues can be addressed through grounding of characters, objects, and other entities on the visual elements. We propose StoryReasoning, a dataset containing 4,178 stories derived from 52,016 movie images, with both structured scene analyses and grounded stories. Each story maintains character and object consistency across frames while explicitly modeling multi-frame relationships through structured tabular representations. Our approach features cross-frame object re-identification using visual similarity and face recognition, chain-of-thought reasoning for explicit narrative modeling, and a grounding scheme that links textual elements to visual entities across multiple frames. We establish baseline performance by fine-tuning Qwen2.5-VL 7B, creating Qwen Storyteller, which performs end-to-end object detection, re-identification, and landmark detection while maintaining consistent object references throughout the story. Evaluation demonstrates a reduction from 4.06 to 3.56 (-12.3%) hallucinations on average per story and an improvement in creativity from 2.58 to 3.38 (+31.0%) when compared to a non-fine-tuned model.
Authors: Sifeng Shang, Jiayi Zhou, Chenyu Lin, Minxian Li, Kaiyang Zhou
Abstract: As the size of large language models grows exponentially, GPU memory has become a bottleneck for adapting these models to downstream tasks. In this paper, we aim to push the limits of memory-efficient training by minimizing memory usage on model weights, gradients, and optimizer states, within a unified framework. Our idea is to eliminate both gradients and optimizer states using zeroth-order optimization, which approximates gradients by perturbing weights during forward passes to identify gradient directions. To minimize memory usage on weights, we employ model quantization, e.g., converting from bfloat16 to int4. However, directly applying zeroth-order optimization to quantized weights is infeasible due to the precision gap between discrete weights and continuous gradients, which would otherwise require de-quantization and re-quantization. To overcome this challenge, we propose Quantized Zeroth-order Optimization (QZO), a simple yet effective approach that perturbs the continuous quantization scale for gradient estimation and uses a directional derivative clipping method to stabilize training. QZO is orthogonal to both scalar-based and codebook-based post-training quantization methods. Compared to full-parameter fine-tuning in 16 bits, QZO can reduce the total memory cost by more than 18$\times$ for 4-bit LLMs, and enables fine-tuning Llama-2-13B within a single 24GB GPU. Code will be released publicly.
Authors: Takashi Ishida, Thanawat Lodkaew, Ikko Yamane
Abstract: Publishing a large language model (LLM) benchmark on the Internet risks contaminating future LLMs: the benchmark may be unintentionally (or intentionally) used to train or select a model. A common mitigation is to keep the benchmark private and let participants submit their models or predictions to the organizers. However, this strategy will require trust in a single organization and still permits test-set overfitting through repeated queries. To overcome this issue, we propose a way to publish benchmarks without completely disclosing the ground-truth answers to the questions, while still maintaining the ability to openly evaluate LLMs. Our main idea is to inject randomness to the answers by preparing several logically correct answers, and only include one of them as the solution in the benchmark. This reduces the best possible accuracy, i.e., Bayes accuracy, of the benchmark. Not only is this helpful to keep us from disclosing the ground truth, but this approach also offers a test for detecting data contamination. In principle, even fully capable models should not surpass the Bayes accuracy. If a model surpasses this ceiling despite this expectation, this is a strong signal of data contamination. We present experimental evidence that our method can detect data contamination accurately on a wide range of benchmarks, models, and training methodologies.
Authors: Jan Carreras Boada, Rao Muhammad Umer, Carsten Marr
Abstract: Biomedical datasets are often constrained by stringent privacy requirements and frequently suffer from severe class imbalance. These two aspects hinder the development of accurate machine learning models. While generative AI offers a promising solution, producing synthetic images of sufficient quality for training robust classifiers remains challenging. This work addresses the classification of individual white blood cells, a critical task in diagnosing hematological malignancies such as acute myeloid leukemia (AML). We introduce CytoDiff, a stable diffusion model fine-tuned with LoRA weights and guided by few-shot samples that generates high-fidelity synthetic white blood cell images. Our approach demonstrates substantial improvements in classifier performance when training data is limited. Using a small, highly imbalanced real dataset, the addition of 5,000 synthetic images per class improved ResNet classifier accuracy from 27\% to 78\% (+51\%). Similarly, CLIP-based classification accuracy increased from 62\% to 77\% (+15\%). These results establish synthetic image generation as a valuable tool for biomedical machine learning, enhancing data coverage and facilitating secure data sharing while preserving patient privacy. Paper code is publicly available at https://github.com/JanCarreras24/CytoDiff.
Authors: Siting Wang, Minnan Pei, Luoyang Sun, Cheng Deng, Kun Shao, Zheng Tian, Haifeng Zhang, Jun Wang
Abstract: Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models show difficulty perception misaligned with human intuition, exhibit dramatic 2Dto-3D performance cliffs, default to formulaic derivation over visualization, and paradoxically suffer performance degradation from Chain-of-Thought prompting in open-source models. Through statistical and qualitative analysis of error types, SpatialViz-Bench demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark data and evaluation code are publicly available.
Authors: Mingda Zhang, Na Zhao, Jianglong Qin, Guoyu Ye, Ruixiang Tang
Abstract: Despite advances from medical large language models in healthcare, rare-disease diagnosis remains hampered by insufficient knowledge-representation depth, limited concept understanding, and constrained clinical reasoning. We propose a framework that couples multi-granularity sparse activation of medical concepts with a hierarchical knowledge graph. Four complementary matching algorithms, diversity control, and a five-level fallback strategy enable precise concept activation, while a three-layer knowledge graph (taxonomy, clinical features, instances) provides structured, up-to-date context. Experiments on the BioASQ rare-disease QA set show BLEU gains of 0.09, ROUGE gains of 0.05, and accuracy gains of 0.12, with peak accuracy of 0.89 approaching the 0.90 clinical threshold. Expert evaluation confirms improvements in information quality, reasoning, and professional expression, suggesting our approach shortens the "diagnostic odyssey" for rare-disease patients.
Authors: Yiran Wu, Mauricio Velazco, Andrew Zhao, Manuel Ra\'ul Mel\'endez Luj\'an, Srisuma Movva, Yogesh K Roy, Quang Nguyen, Roberto Rodriguez, Qingyun Wu, Michael Albada, Julia Kiseleva, Anand Mudgerikar
Abstract: We present ExCyTIn-Bench, the first benchmark to Evaluate an LLM agent x on the task of Cyber Threat Investigation through security questions derived from investigation graphs. Real-world security analysts must sift through a large number of heterogeneous alert signals and security logs, follow multi-hop chains of evidence, and compile an incident report. With the developments of LLMs, building LLM-based agents for automatic thread investigation is a promising direction. To assist the development and evaluation of LLM agents, we construct a dataset from a controlled Azure tenant that covers 8 simulated real-world multi-step attacks, 57 log tables from Microsoft Sentinel and related services, and 589 automatically generated questions. We leverage security logs extracted with expert-crafted detection logic to build threat investigation graphs, and then generate questions with LLMs using paired nodes on the graph, taking the start node as background context and the end node as answer. Anchoring each question to these explicit nodes and edges not only provides automatic, explainable ground truth answers but also makes the pipeline reusable and readily extensible to new logs. This also enables the automatic generation of procedural tasks with verifiable rewards, which can be naturally extended to training agents via reinforcement learning. Our comprehensive experiments with different models confirm the difficulty of the task: with the base setting, the average reward across all evaluated models is 0.249, and the best achieved is 0.368, leaving substantial headroom for future research. Code and data are coming soon!
Authors: Yu Zhang, Baotong Tian, Zhiyao Duan
Abstract: Zero-shot online voice conversion (VC) holds significant promise for real-time communications and entertainment. However, current VC models struggle to preserve semantic fidelity under real-time constraints, deliver natural-sounding conversions, and adapt effectively to unseen speaker characteristics. To address these challenges, we introduce Conan, a chunkwise online zero-shot voice conversion model that preserves the content of the source while matching the voice timbre and styles of reference speech. Conan comprises three core components: 1) a Stream Content Extractor that leverages Emformer for low-latency streaming content encoding; 2) an Adaptive Style Encoder that extracts fine-grained stylistic features from reference speech for enhanced style adaptation; 3) a Causal Shuffle Vocoder that implements a fully causal HiFiGAN using a pixel-shuffle mechanism. Experimental evaluations demonstrate that Conan outperforms baseline models in subjective and objective metrics. Audio samples can be found at https://aaronz345.github.io/ConanDemo.
Authors: Yifan Zhang
Abstract: Autoregressive language models achieve remarkable performance, yet a unified theory explaining their internal mechanisms--how training shapes their representations and enables complex behaviors--remains elusive. We introduce a new analytical framework that models the single-step generation process as a composition of information-processing stages using the language of Markov categories. This compositional perspective provides a unified mathematical language to connect three critical aspects of language modeling that are typically studied in isolation: the training objective, the geometry of the learned representation space, and practical model capabilities. First, our framework provides a precise information-theoretic rationale for the success of multi-token prediction methods like speculative decoding, quantifying the "information surplus" a model's hidden state contains about tokens beyond the immediate next one. Second, we clarify how the standard negative log-likelihood (NLL) objective compels the model to learn not just the next word, but also the data's intrinsic conditional uncertainty, a process we formalize using categorical entropy. Our central result reveals that NLL training functions as an implicit form of spectral contrastive learning. We prove that, for common model architectures, this simple predictive objective forces the model to sculpt a geometrically structured representation space, implicitly aligning representations with the eigenspectrum of a "predictive similarity" operator. This work offers a powerful new lens to understand how information flows through a model and how the training objective shapes its internal geometry, thereby bridging the gap between learning theory and the practical success of large language models.
Authors: Hongjin Qian, Zheng Liu
Abstract: In this work, we propose MetaAgent, an agentic paradigm inspired by the principle of learning-by-doing, where expertise is developed through hands-on practice and continual self-improvement. MetaAgent starts with a minimal workflow, equipped only with basic reasoning and adaptive help-seeking abilities. When a knowledge gap is encountered, MetaAgent generates natural language help requests, which are routed to the most suitable external tool by a dedicated tool router. As MetaAgent solves tasks, it continually conducts self-reflection and answer verification, distilling actionable experience into concise texts that are dynamically incorporated into future task contexts. Besides, MetaAgent autonomously builds in-house tools and a persistent knowledge base by organizing its tool-use history, further enhancing its ability to retrieve and integrate relevant information We term this continual, data-driven process as \textit{meta tool learning}, through which MetaAgent incrementally refines its reasoning and tool-use strategies, without changing model parameters or requiring further post-training. Evaluated on challenging knowledge discovery benchmarks, including GAIA, WebWalkerQA, and BrowseCamp, MetaAgent consistently outperforms workflow-based baselines and matches or exceeds end-to-end trained agents, demonstrating the promise of self-evolving agentic systems for robust, general-purpose knowledge discovery. We provide our source codes in https://github.com/qhjqhj00/MetaAgent.
Authors: Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, Zhaochun Ren, Nikos Aletras, Xi Wang, Han Zhou, Zaiqiao Meng
Abstract: Recent advances in large language models have sparked growing interest in AI agents capable of solving complex, real-world tasks. However, most existing agent systems rely on manually crafted configurations that remain static after deployment, limiting their ability to adapt to dynamic and evolving environments. To this end, recent research has explored agent evolution techniques that aim to automatically enhance agent systems based on interaction data and environmental feedback. This emerging direction lays the foundation for self-evolving AI agents, which bridge the static capabilities of foundation models with the continuous adaptability required by lifelong agentic systems. In this survey, we provide a comprehensive review of existing techniques for self-evolving agentic systems. Specifically, we first introduce a unified conceptual framework that abstracts the feedback loop underlying the design of self-evolving agentic systems. The framework highlights four key components: System Inputs, Agent System, Environment, and Optimisers, serving as a foundation for understanding and comparing different strategies. Based on this framework, we systematically review a wide range of self-evolving techniques that target different components of the agent system. We also investigate domain-specific evolution strategies developed for specialised fields such as biomedicine, programming, and finance, where optimisation objectives are tightly coupled with domain constraints. In addition, we provide a dedicated discussion on the evaluation, safety, and ethical considerations for self-evolving agentic systems, which are critical to ensuring their effectiveness and reliability. This survey aims to provide researchers and practitioners with a systematic understanding of self-evolving AI agents, laying the foundation for the development of more adaptive, autonomous, and lifelong agentic systems.
Authors: Kangyu Wang, Hongliang He, Lin Liu, Ruiqi Liang, Zhenzhong Lan, Jianguo Li
Abstract: Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) have ushered in a new era of AI capabilities, demonstrating near-human-level performance across diverse scenarios. While numerous benchmarks (e.g., MMLU) and leaderboards (e.g., Chatbot Arena) have been proposed to help evolve the development of LLMs and MLLMs, most rely on static datasets or crowdsourced general-domain prompts, often falling short of reflecting performance in real-world applications. To bridge this critical gap, we present Inclusion Arena, a live leaderboard that ranks models based on human feedback collected directly from AI-powered applications. Our platform integrates pairwise model comparisons into natural user interactions, ensuring evaluations reflect practical usage scenarios. For robust model ranking, we employ the Bradley-Terry model augmented with two key innovations: (1) Placement Matches, a cold-start mechanism to quickly estimate initial ratings for newly integrated models, and (2) Proximity Sampling, an intelligent comparison strategy that prioritizes battles between models of similar capabilities to maximize information gain and enhance rating stability. Extensive empirical analyses and simulations demonstrate that Inclusion Arena yields reliable and stable rankings, exhibits higher data transitivity compared to general crowdsourced datasets, and significantly mitigates the risk of malicious manipulation. By fostering an open alliance between foundation models and real-world applications, Inclusion Arena aims to accelerate the development of LLMs and MLLMs truly optimized for practical, user-centric deployments. The platform is publicly accessible at https://www.tbox.cn/about/model-ranking.
Authors: Hikaru Tsujimura, Arush Tagade
Abstract: Large Language Models (LLMs) often display overconfidence, presenting information with unwarranted certainty in high-stakes contexts. We investigate the internal basis of this behavior via mechanistic interpretability. Using open-sourced Llama 3.2 models fine-tuned on human annotated assertiveness datasets, we extract residual activations across all layers, and compute similarity metrics to localize assertive representations. Our analysis identifies layers most sensitive to assertiveness contrasts and reveals that high-assertive representations decompose into two orthogonal sub-components of emotional and logical clusters-paralleling the dual-route Elaboration Likelihood Model in Psychology. Steering vectors derived from these sub-components show distinct causal effects: emotional vectors broadly influence prediction accuracy, while logical vectors exert more localized effects. These findings provide mechanistic evidence for the multi-component structure of LLM assertiveness and highlight avenues for mitigating overconfident behavior.
Authors: Zicong Tang, Ziyang Ma, Suqing Wang, Zuchao Li, Lefei Zhang, Hai Zhao, Yun Li, Qianren Wang
Abstract: Large Vision-Language Models (LVLMs) process multimodal inputs consisting of text tokens and vision tokens extracted from images or videos. Due to the rich visual information, a single image can generate thousands of vision tokens, leading to high computational costs during the prefilling stage and significant memory overhead during decoding. Existing methods attempt to prune redundant vision tokens, revealing substantial redundancy in visual representations. However, these methods often struggle in shallow layers due to the lack of sufficient contextual information. We argue that many visual tokens are inherently redundant even in shallow layers and can be safely and effectively pruned with appropriate contextual signals. In this work, we propose CoViPAL, a layer-wise contextualized visual token pruning method that employs a Plug-and-Play Pruning Module (PPM) to predict and remove redundant vision tokens before they are processed by the LVLM. The PPM is lightweight, model-agnostic, and operates independently of the LVLM architecture, ensuring seamless integration with various models. Extensive experiments on multiple benchmarks demonstrate that CoViPAL outperforms training-free pruning methods under equal token budgets and surpasses training-based methods with comparable supervision. CoViPAL offers a scalable and efficient solution to improve inference efficiency in LVLMs without compromising accuracy.
Authors: Antony C Chan
Abstract: This article presents a practitioner's reflection on applying declarative, 5th generation, problem formulation language (5GL) to de novo imaging system design, informed by experiences across the interdisciplinary research in academia and cross-functional product development within the private sector. Using the 96-Eyes project: 96-camera parallel multi-modal imager for high-throughput drug discovery as a representative case, I illustrate how project requirements, ranging from hardware constraints to life sciences needs, can be formalized into machine-readable problem statements to preserve mission-critical input from diverse domain stakeholders. This declarative approach enhances transparency, ensures design traceability, and minimizes costly misalignment across optical, algorithmic, hardware-accelerated compute, and life sciences teams. Alongside the technical discussion of 5GL with real-world code examples, I reflect on the practical barriers to adopting 5GL in environments where imperative, 3rd-generation languages (3GL) remain the default medium for inter-team collaboration. Rather than offering an one-size-fits-all solution, these learned lessons highlight how programming paradigms implicitly shapes research workflows through existing domain hierarchies. The discussion aims to invite further explorations into how declarative problem formulations can facilitate innovation in settings where concurrent R\&{}D workflows are gaining traction, as opposed to environments where sequential, phase-driven workflows remain the norm.
Authors: Yuxuan Cai, Yipeng Hao, Jie Zhou, Hang Yan, Zhikai Lei, Rui Zhen, Zhenhua Han, Yutao Yang, Junsong Li, Qianjun Pan, Tianyu Huai, Qin Chen, Xin Li, Kai Chen, Bo Zhang, Xipeng Qiu, Liang He
Abstract: As AI advances toward general intelligence, the focus is shifting from systems optimized for static tasks to creating open-ended agents that learn continuously. In this paper, we introduce Experience-driven Lifelong Learning (ELL), a framework for building self-evolving agents capable of continuous growth through real-world interaction. The framework is built on four core principles: (1) Experience Exploration: Agents learn through continuous, self-motivated interaction with dynamic environments, navigating interdependent tasks and generating rich experiential trajectories. (2) Long-term Memory: Agents preserve and structure historical knowledge, including personal experiences, domain expertise, and commonsense reasoning, into a persistent memory system. (3) Skill Learning: Agents autonomously improve by abstracting recurring patterns from experience into reusable skills, which are actively refined and validated for application in new tasks. (4) Knowledge Internalization: Agents internalize explicit and discrete experiences into implicit and intuitive capabilities as "second nature". We also introduce StuLife, a benchmark dataset for ELL that simulates a student's holistic college journey, from enrollment to academic and personal development, across three core phases and ten detailed sub-scenarios. StuLife is designed around three key paradigm
Authors: Huaiyuan Yao, Wanpeng Xu, Justin Turnau, Nadia Kellam, Hua Wei
Abstract: Preparing high-quality instructional materials remains a labor-intensive process that often requires extensive coordination among teaching faculty, instructional designers, and teaching assistants. In this work, we present Instructional Agents, a multi-agent large language model (LLM) framework designed to automate end-to-end course material generation, including syllabus creation, lecture scripts, LaTeX-based slides, and assessments. Unlike existing AI-assisted educational tools that focus on isolated tasks, Instructional Agents simulates role-based collaboration among educational agents to produce cohesive and pedagogically aligned content. The system operates in four modes: Autonomous, Catalog-Guided, Feedback-Guided, and Full Co-Pilot mode, enabling flexible control over the degree of human involvement. We evaluate Instructional Agents across five university-level computer science courses and show that it produces high-quality instructional materials while significantly reducing development time and human workload. By supporting institutions with limited instructional design capacity, Instructional Agents provides a scalable and cost-effective framework to democratize access to high-quality education, particularly in underserved or resource-constrained settings.
Authors: Taebaek Hwang, Minseo Kim, Gisang Lee, Seonuk Kim, Hyunjun Eun
Abstract: Understanding and reasoning over text within visual contexts poses a significant challenge for Vision-Language Models (VLMs), given the complexity and diversity of real-world scenarios. To address this challenge, text-rich Visual Question Answering (VQA) datasets and benchmarks have emerged for high-resource languages like English. However, a critical gap persists for low-resource languages such as Korean, where the lack of comprehensive benchmarks hinders robust model evaluation and comparison. To bridge this gap, we introduce KRETA, a benchmark for Korean Reading and rEasoning in Text-rich VQA Attuned to diverse visual contexts. KRETA facilitates an in-depth evaluation of both visual text understanding and reasoning capabilities, while also supporting a multifaceted assessment across 15 domains and 26 image types. Additionally, we introduce a semi-automated VQA generation pipeline specifically optimized for text-rich settings, leveraging refined stepwise image decomposition and a rigorous seven-metric evaluation protocol to ensure data quality. While KRETA is tailored for Korean, we hope our adaptable and extensible pipeline will facilitate the development of similar benchmarks in other languages, thereby accelerating multilingual VLM research. The code and dataset for KRETA are available at https://github.com/tabtoyou/KRETA.
Authors: Haoze Wu, Cheng Wang, Wenshuo Zhao, Junxian He
Abstract: Recent advances in applying reinforcement learning (RL) to large language models (LLMs) have led to substantial progress. In particular, a series of remarkable yet often counterintuitive phenomena have been reported in LLMs, exhibiting patterns not typically observed in traditional RL settings. For example, notable claims include that a single training example can match the performance achieved with an entire dataset, that the reward signal does not need to be very accurate, and that training solely with negative samples can match or even surpass sophisticated reward-based methods. However, the precise conditions under which these observations hold - and, critically, when they fail - remain unclear. In this work, we identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong Model-Task Alignment, as measured by pass@k accuracy on the evaluated task. Through a systematic and comprehensive examination of a series of counterintuitive claims, supported by rigorous experimental validation across different model architectures and task domains, our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment. In contrast, these techniques fail to drive substantial learning in more challenging regimes, where standard RL methods remain effective.