Authors: Wannan Yang, Xinchi Qiu, Lei Yu, Yuchen Zhang, Oliver Aobo Yang, Narine Kokhlikyan, Nicola Cancedda, Diego Garcia-Olano
Abstract: Large Language Models (LLMs) exhibit impressive capabilities but often hallucinate, confidently providing incorrect answers instead of admitting ignorance. Prior work has shown that models encode linear representations of their own knowledge and that activation steering can reduce hallucinations. These approaches, however, require real-time monitoring and intervention during inference. We introduce Contrastive Activation Steering for Amortized Learning (CASAL), an efficient algorithm that connects interpretability with amortized optimization. CASAL directly bakes the benefits of activation steering into model's weights. Once trained, LLMs answer questions they know while abstaining from answering those they do not. CASAL's light-weight design requires training only a submodule of a single transformer layer and yet reduces hallucination by 30%-40% across multiple short-form QA benchmarks. CASAL is 30x more compute-efficient and 20x more data-efficient than strong LoRA-based baselines such as SFT and DPO, boosting its practical applicability in data scarce domains. Importantly, CASAL also generalizes effectively to out-of-distribution (OOD) domains. We showcase CASAL's flexibility in mitigating hallucinations in both text-only and vision-language models. To our knowledge, CASAL is the first steering-based training method that has been shown to be effective for both dense and Mixture-of-Experts (MoE) models. CASAL represents a promising step forward for applying interpretability-inspired method for practical deployment in production systems.
Authors: Vivek Bhavsar, Joseph Ereifej, Aravanan Gurusami
Abstract: Large language models accelerate literature synthesis but can hallucinate and mis-cite, limiting their usefulness in expert workflows. We present RA-FSM (Research Assistant - Finite State Machine), a modular GPT-based research assistant that wraps generation in a finite-state control loop: Relevance -> Confidence -> Knowledge. The system is grounded in vector retrieval and a deterministic citation pipeline. The controller filters out-of-scope queries, scores answerability, decomposes questions, and triggers retrieval only when needed, and emits answers with confidence labels and in-corpus, de-duplicated references. A ranked-tier ingestion workflow constructs a domain knowledge base from journals, conferences, indices, preprints, and patents, writing both to a dense vector index and to a relational store of normalized metrics. We implement the system for photonics and evaluate it on six task categories: analytical reasoning, numerical analysis, methodological critique, comparative synthesis, factual extraction, and application design. In blinded A/B reviews, domain experts prefer RA-FSM to both a strong Notebook LM (NLM) and a vanilla Default GPT API call single-pass baseline, citing stronger boundary-condition handling and more defensible evidence use. Coverage and novelty analyses indicate that RA-FSM explores beyond the NLM while incurring tunable latency and cost overheads. The design emphasizes transparent, well-cited answers for high-stakes technical work and is generalizable to other scientific domains.
Authors: So Kuroki, Yotaro Kubo, Takuya Akiba, Yujin Tang
Abstract: Real-time speech-to-speech (S2S) models excel at generating natural, low-latency conversational responses but often lack deep knowledge and semantic understanding. Conversely, cascaded systems combining automatic speech recognition, a text-based Large Language Model (LLM), and text-to-speech synthesis offer superior knowledge representation at the cost of high latency, which disrupts the flow of natural interaction. This paper introduces a novel hybrid architecture that bridges the gap between these two paradigms. Our framework processes user speech through an S2S transformer for immediate responsiveness while concurrently relaying the query to a powerful back-end LLM. The LLM's text-based response is then injected in real time to guide the S2S model's speech generation, effectively infusing its output with rich knowledge without the full latency penalty of a cascaded system. We evaluated our method using a speech-synthesized variant of the MT-Bench benchmark that consists of multi-turn question-answering sessions. The results demonstrate that our system substantially outperforms a baseline S2S model in response correctness, approaching that of a cascaded system, while maintaining a latency on par with the baseline.
Authors: Ziqing Wang, Chengsheng Mao, Xiaole Wen, Yuan Luo, Kaize Ding
Abstract: Medical Multimodal Large Language Models (Med-MLLMs) have shown great promise in medical visual question answering (Med-VQA). However, when deployed in low-resource settings where abundant labeled data are unavailable, existing Med-MLLMs commonly fail due to their medical reasoning capability bottlenecks: (i) the intrinsic reasoning bottleneck that ignores the details from the medical image; (ii) the extrinsic reasoning bottleneck that fails to incorporate specialized medical knowledge. To address those limitations, we propose AMANDA, a training-free agentic framework that performs medical knowledge augmentation via LLM agents. Specifically, our intrinsic medical knowledge augmentation focuses on coarse-to-fine question decomposition for comprehensive diagnosis, while extrinsic medical knowledge augmentation grounds the reasoning process via biomedical knowledge graph retrieval. Extensive experiments across eight Med-VQA benchmarks demonstrate substantial improvements in both zero-shot and few-shot Med-VQA settings. The code is available at https://github.com/REAL-Lab-NU/AMANDA.
Authors: Kanghoon Yoon, Minsub Kim, Sungjae Lee, Joonhyung Lee, Sunghyeon Woo, Yeonjun In, Se Jung Kwon, Chanyoung Park, Dongsoo Lee
Abstract: Speculative decoding accelerates LLM inference by verifying candidate tokens from a draft model against a larger target model. Recent judge decoding boosts this process by relaxing verification criteria by accepting draft tokens that may exhibit minor discrepancies from target model output, but existing methods are restricted by their reliance on human annotations or tasks with verifiable ground truths, limiting generalizability across diverse NLP tasks. We propose SelfJudge, which trains judge verifiers via self-supervision of the target model. Our method measures semantic preservation by assessing whether token-substituted responses preserve the meaning of original responses, enabling automatic verifier training across diverse NLP tasks. Our experiments show SelfJudge achieves superior inference-accuracy trade-offs than judge decoding baselines, offering a broadly applicable solution for faster LLM inference.
Authors: Junlong Jia, Ziyang Chen, Xing Wu, Chaochen Gao, Zijia Lin, Debing Zhang, Songlin Hu, Binghui Guo
Abstract: Training long-context language models to capture long-range dependencies requires specialized data construction. Current approaches, such as generic text concatenation or heuristic-based variants, frequently fail to guarantee genuine long-range dependencies. We propose EntropyLong, a novel data construction method that leverages predictive uncertainty to verify dependency quality. Our approach identifies high-entropy positions in documents, retrieves semantically relevant contexts from large corpora, and verifies their utility by assessing whether they reduce prediction entropy. This model-in-the-loop verification ensures each dependency represents measurable information gain rather than spurious correlation. We construct training samples with long-range dependencies by combining original documents with these verified contextual supplements. Using FineWebEdu and Cosmopedia, we generate a dataset of 128K-length sequences with verified dependencies. Models trained on this data demonstrate significant improvements on RULER benchmarks, particularly in tasks requiring distant information. Following instruction fine-tuning, our models also achieve substantial gains on LongBenchv2, demonstrating enhanced long-context understanding. Extensive ablation studies further validate the necessity and effectiveness of entropybased verification for long-context training.
Authors: Moonkyung Ryu, Chih-Wei Hsu, Yinlam Chow, Mohammad Ghavamzadeh, Craig Boutilier
Abstract: While language models (LMs) offer great potential for conversational recommender systems (CRSs), the paucity of public CRS data makes fine-tuning LMs for CRSs challenging. In response, LMs as user simulators qua data generators can be used to train LM-based CRSs, but often lack behavioral consistency, generating utterance sequences inconsistent with those of any real user. To address this, we develop a methodology for generating natural dialogues that are consistent with a user's underlying state using behavior simulators together with LM-prompting. We illustrate our approach by generating a large, open-source CRS data set with both preference elicitation and example critiquing. Rater evaluation on some of these dialogues shows them to exhibit considerable consistency, factuality and naturalness.
Authors: Yapei Feng, Feng Jiang, Shanhao Wu, Hua Zhong
Abstract: Neural linguistic steganography aims to embed information into natural text while preserving statistical undetectability. A fundamental challenge in this ffeld stems from tokenization ambiguity in modern tokenizers, which can lead to catastrophic decoding failures. The recent method, SyncPool, addresses this ambiguity by employing a coarse-grained synchronization mechanism over groups of ambiguous candidates. However, SyncPool sacriffces embedding capacity, as it utilizes the entire Shannon entropy of an ambiguous group solely for synchronization rather than for payload embedding. We propose a method named look-ahead Sync, which overcomes the capacity limitation of SyncPool while retaining its provable security guarantees. Our approach performs minimal synchronized sampling only on truly indistinguishable token sequences, while strategically preserving all other discernible paths to maximize embedding capacity. We provide theoretical proofs for the security of our method and analyze the gap between its achievable embedding capacity and the theoretical upper bound. Experiments on English (using Llama 3) and Chinese (using Qwen 2.5) benchmarks show that our method consistently approaches the theoretical capacity upper bound and signiffcantly outperforms SyncPool. The improvement in embedding rate exceeds 160% in English and 25% in Chinese, particularly in settings with larger candidate pools. This work represents a signiffcant step toward practical high-capacity provably secure linguistic steganography.
Authors: Chiara Pugliese, Francesco Lettich, Guido Rocchietti, Chiara Renso, Fabio Pinelli
Abstract: In this resource paper, we present two publicly available datasets of semantically enriched human trajectories, together with the pipeline to build them. The trajectories are publicly available GPS traces retrieved from OpenStreetMap. Each dataset includes contextual layers such as stops, moves, points of interest (POIs), inferred transportation modes, and weather data. A novel semantic feature is the inclusion of synthetic, realistic social media posts generated by Large Language Models (LLMs), enabling multimodal and semantic mobility analysis. The datasets are available in both tabular and Resource Description Framework (RDF) formats, supporting semantic reasoning and FAIR data practices. They cover two structurally distinct, large cities: Paris and New York. Our open source reproducible pipeline allows for dataset customization, while the datasets support research tasks such as behavior modeling, mobility prediction, knowledge graph construction, and LLM-based applications. To our knowledge, our resource is the first to combine real-world movement, structured semantic enrichment, LLM-generated text, and semantic web compatibility in a reusable framework.
Authors: Zhe Li, Wei Zhao, Yige Li, Jun Sun
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their deployment is frequently undermined by undesirable behaviors such as generating harmful content, factual inaccuracies, and societal biases. Diagnosing the root causes of these failures poses a critical challenge for AI safety. Existing attribution methods, particularly those based on parameter gradients, often fall short due to prohibitive noisy signals and computational complexity. In this work, we introduce a novel and efficient framework that diagnoses a range of undesirable LLM behaviors by analyzing representation and its gradients, which operates directly in the model's activation space to provide a semantically meaningful signal linking outputs to their training data. We systematically evaluate our method for tasks that include tracking harmful content, detecting backdoor poisoning, and identifying knowledge contamination. The results demonstrate that our approach not only excels at sample-level attribution but also enables fine-grained token-level analysis, precisely identifying the specific samples and phrases that causally influence model behavior. This work provides a powerful diagnostic tool to understand, audit, and ultimately mitigate the risks associated with LLMs. The code is available at https://github.com/plumprc/RepT.
Authors: Xiao-Wen Yang, Zihao Zhang, Jianuo Cao, Zhi Zhou, Zenan Li, Lan-Zhe Guo, Yuan Yao, Taolue Chen, Yu-Feng Li, Xiaoxing Ma
Abstract: Large language models (LLMs) have recently demonstrated remarkable progress in formal theorem proving. Yet their ability to serve as practical assistants for mathematicians, filling in missing steps within complex proofs, remains underexplored. We identify this challenge as the task of subgoal completion, where an LLM must discharge short but nontrivial proof obligations left unresolved in a human-provided sketch. To study this problem, we introduce FormalML, a Lean 4 benchmark built from foundational theories of machine learning. Using a translation tactic that converts procedural proofs into declarative form, we extract 4937 problems spanning optimization and probability inequalities, with varying levels of difficulty. FormalML is the first subgoal completion benchmark to combine premise retrieval and complex research-level contexts. Evaluation of state-of-the-art provers highlights persistent limitations in accuracy and efficiency, underscoring the need for more capable LLM-based theorem provers for effective subgoal completion,
Authors: Abdulhady Abas Abdullah, Hadi Veisi, Hussein M. Al
Abstract: Semantic Textual Similarity (STS) measures the degree of meaning overlap between two texts and underpins many NLP tasks. While extensive resources exist for high-resource languages, low-resource languages such as Kurdish remain underserved. We present, to our knowledge, the first Kurdish STS dataset: 10,000 sentence pairs spanning formal and informal registers, each annotated for similarity. We benchmark Sentence-BERT, multilingual BERT, and other strong baselines, obtaining competitive results while highlighting challenges arising from Kurdish morphology, orthographic variation, and code-mixing. The dataset and baselines establish a reproducible evaluation suite and provide a strong starting point for future research on Kurdish semantics and low-resource NLP.
Authors: Ishak Soltani, Francisco Belo, Bernardo Tavares
Abstract: This paper presents CRACQ, a multi-dimensional evaluation framework tailored to evaluate documents across f i v e specific traits: Coherence, Rigor, Appropriateness, Completeness, and Quality. Building on insights from traitbased Automated Essay Scoring (AES), CRACQ expands its fo-cus beyond essays to encompass diverse forms of machine-generated text, providing a rubricdriven and interpretable methodology for automated evaluation. Unlike singlescore approaches, CRACQ integrates linguistic, semantic, and structural signals into a cumulative assessment, enabling both holistic and trait-level analysis. Trained on 500 synthetic grant pro-posals, CRACQ was benchmarked against an LLM-as-a-judge and further tested on both strong and weak real applications. Preliminary results in-dicate that CRACQ produces more stable and interpretable trait-level judgments than direct LLM evaluation, though challenges in reliability and domain scope remain
Authors: Samyak Jhaveri, Praphul Singh, Jangwon Kim, Tara Taghavi, Krishnaram Kenthapadi
Abstract: Automating clinical documentation with large language models requires precise alignment with priorities such as completeness and factual grounding. We present an evaluation-integrated reinforcement learning framework for long-form clinical text generation that couples Group Relative Policy Optimization (GRPO) with DocLens, a claim-level evaluator that provides deterministic, dialogue-grounded rewards. Our method directly optimizes factual grounding and completeness without training a separate reward model or relying on human-authored references. Empirically, the approach improves clinical note quality and reduces training cost via a simple reward-gating strategy. An independent GPT-5 qualitative evaluation further supports these gains, showing higher preference for GRPO outputs in factuality, completeness, and brevity, with fewer omissions and hallucinations. Because the benchmarks are relatively clean and the base model already well aligned, these improvements likely represent a conservative lower bound. The framework is scalable to real-world settings and can incorporate custom objectives such as guideline adherence or billing preferences.
Authors: Kevin Zhou, Adam Dejl, Gabriel Freedman, Lihu Chen, Antonio Rago, Francesca Toni
Abstract: Research in uncertainty quantification (UQ) for large language models (LLMs) is increasingly important towards guaranteeing the reliability of this groundbreaking technology. We explore the integration of LLM UQ methods in argumentative LLMs (ArgLLMs), an explainable LLM framework for decision-making based on computational argumentation in which UQ plays a critical role. We conduct experiments to evaluate ArgLLMs' performance on claim verification tasks when using different LLM UQ methods, inherently performing an assessment of the UQ methods' effectiveness. Moreover, the experimental procedure itself is a novel way of evaluating the effectiveness of UQ methods, especially when intricate and potentially contentious statements are present. Our results demonstrate that, despite its simplicity, direct prompting is an effective UQ strategy in ArgLLMs, outperforming considerably more complex approaches.
Authors: Xin Gao, Ruiyi Zhang, Daniel Du, Saurabh Mahindre, Sai Ashish Somayajula, Pengtao Xie
Abstract: Large Language Models (LLMs) are widely used for temporal prediction, but their reliance on pretraining data raises contamination concerns, as accurate predictions on pre-cutoff test data may reflect memorization rather than reasoning, leading to an overestimation of their generalization capability. With the recent emergence of prompting-based unlearning techniques, a natural question arises: Can LLMs be prompted to simulate an earlier knowledge cutoff? In this work, we investigate the capability of prompting to simulate earlier knowledge cutoff in LLMs. We construct three evaluation datasets to assess the extent to which LLMs can forget (1) direct factual knowledge, (2) semantic shifts, and (3) causally related knowledge. Results demonstrate that while prompt-based simulated knowledge cutoffs show effectiveness when directly queried with the information after that date, they struggle to induce forgetting when the forgotten content is not directly asked but causally related to the query. These findings highlight the need for more rigorous evaluation settings when applying LLMs for temporal prediction tasks. The full dataset and evaluation code are available at https://github.com/gxx27/time_unlearn.
Authors: Yifan Wang, Bolian Li, Junlin Wu, Zhaoxuan Tan, Zheli Liu, Ruqi Zhang, Ananth Grama, Qingkai Zeng
Abstract: Real-world large language model deployments (e.g., conversational AI systems, code generation assistants) naturally generate abundant implicit user dissatisfaction (DSAT) signals, as users iterate toward better answers through refinements, corrections, and expressed preferences, while explicit satisfaction (SAT) feedback is scarce. Existing preference learning approaches are poorly aligned with this data profile, as they rely on costly human annotations or assume plentiful positive responses. In this paper, we introduce \textbf{DRIFT} (\textbf{D}issatisfaction-\textbf{R}efined \textbf{I}terative pre\textbf{F}erence \textbf{T}raining), which anchors training on real-world DSAT signals and samples positives dynamically from the evolving policy. Empirically, DRIFT models trained on real-world \textit{WildFeedback} datasets and synthetic \textit{UltraFeedback} datasets achieve up to +6.23\% (7B) / +7.61\% (14B) on WildBench Task Score and up to +8.95\% (7B) / +12.29\% (14B) on AlpacaEval2 win rate over base models, outperforming strong baseline methods such as iterative DPO and SPIN. At larger scales, the improvements are particularly pronounced: 14B models trained with DRIFT surpass GPT-4o-mini on WildBench. Further analysis shows that DRIFT also preserves exploratory capacity, yielding more diverse high-reward solutions rather than collapsing to narrow subsets. Theoretically, we demonstrate that this design preserves preference margins and avoids the gradient degeneration. These results show that DRIFT is an effective and scalable recipe for real-world post-training that leverages the most abundant and informative signal. The code and data are available at https://github.com/cacayaya/DRIFT.git.
Authors: Aur\'elien B\"uck-Kaeffer, Je Qin Chooi, Dan Zhao, Maximilian Puelma Touzel, Kellin Pelrine, Jean-Fran\c{c}ois Godbout, Reihaneh Rabbany, Zachary Yang
Abstract: Large language models (LLMs) offer promising capabilities for simulating social media dynamics at scale, enabling studies that would be ethically or logistically challenging with human subjects. However, the field lacks standardized data resources for fine-tuning and evaluating LLMs as realistic social media agents. We address this gap by introducing SIMPACT, the SIMulation-oriented Persona and Action Capture Toolkit, a privacy respecting framework for constructing behaviorally-grounded social media datasets suitable for training agent models. We formulate next-action prediction as a task for training and evaluating LLM-based agents and introduce metrics at both the cluster and population levels to assess behavioral fidelity and stylistic realism. As a concrete implementation, we release BluePrint, a large-scale dataset built from public Bluesky data focused on political discourse. BluePrint clusters anonymized users into personas of aggregated behaviours, capturing authentic engagement patterns while safeguarding privacy through pseudonymization and removal of personally identifiable information. The dataset includes a sizable action set of 12 social media interaction types (likes, replies, reposts, etc.), each instance tied to the posting activity preceding it. This supports the development of agents that use context-dependence, not only in the language, but also in the interaction behaviours of social media to model social media users. By standardizing data and evaluation protocols, SIMPACT provides a foundation for advancing rigorous, ethically responsible social media simulations. BluePrint serves as both an evaluation benchmark for political discourse modeling and a template for building domain specific datasets to study challenges such as misinformation and polarization.
Authors: Peijun Zhu, Ning Yang, Jiayu Wei, Jinghang Wu, Haijun Zhang
Abstract: Mixture-of-Experts (MoE) Large Language Models (LLMs) face a trilemma of load imbalance, parameter redundancy, and communication overhead. We introduce a unified framework based on dynamic expert clustering and structured compression to address these issues cohesively. Our method employs an online clustering procedure that periodically regroups experts using a fused metric of parameter and activation similarity, which stabilizes expert utilization. To our knowledge, this is one of the first frameworks to leverage the semantic embedding capability of the router to dynamically reconfigure the model's architecture during training for substantial efficiency gains. Within each cluster, we decompose expert weights into a shared base matrix and extremely low-rank residual adapters, achieving up to fivefold parameter reduction per group while preserving specialization. This structure enables a two-stage hierarchical routing strategy: tokens are first assigned to a cluster, then to specific experts within it, drastically reducing the routing search space and the volume of all-to-all communication. Furthermore, a heterogeneous precision scheme, which stores shared bases in FP16 and residual factors in INT4, coupled with dynamic offloading of inactive clusters, reduces peak memory consumption to levels comparable to dense models. Evaluated on GLUE and WikiText-103, our framework matches the quality of standard MoE models while reducing total parameters by approximately 80%, improving throughput by 10% to 20%, and lowering expert load variance by a factor of over three. Our work demonstrates that structural reorganization is a principled path toward scalable, efficient, and memory-effective MoE LLMs.
Authors: Konstantinos Katharakis, Sippo Rossi, Raghava Rao Mukkamala
Abstract: The adoption of generative AI and large language models (LLMs) in education is still emerging. In this study, we explore the development and evaluation of AI teaching assistants that provide curriculum-based guidance using a retrieval-augmented generation (RAG) pipeline applied to selected open-source small language models (SLMs). We benchmarked eight SLMs, including LLaMA 3.1, IBM Granite 3.3, and Gemma 3 (7-17B parameters), against GPT-4o. Our findings show that with proper prompting and targeted retrieval, SLMs can match LLMs in delivering accurate, pedagogically aligned responses. Importantly, SLMs offer significant sustainability benefits due to their lower computational and energy requirements, enabling real-time use on consumer-grade hardware without depending on cloud infrastructure. This makes them not only cost-effective and privacy-preserving but also environmentally responsible, positioning them as viable AI teaching assistants for educational institutions aiming to scale personalized learning in a sustainable and energy-efficient manner.
Authors: Guy Dar
Abstract: We build upon vec2vec, a procedure designed to align text embedding spaces without parallel data. vec2vec finds a near-perfect alignment, but it is expensive and unstable. We present mini-vec2vec, a simple and efficient alternative that requires substantially lower computational cost and is highly robust. Moreover, the learned mapping is a linear transformation. Our method consists of three main stages: a tentative matching of pseudo-parallel embedding vectors, transformation fitting, and iterative refinement. Our linear alternative exceeds the original instantiation of vec2vec by orders of magnitude in efficiency, while matching or exceeding their results. The method's stability and interpretable algorithmic steps facilitate scaling and unlock new opportunities for adoption in new domains and fields.
Authors: Dzmitry Pihulski, Karol Charchut, Viktoria Novogrodskaia, Jan Koco\'n
Abstract: Converting natural language questions into SQL queries (Text-to-SQL) enables non-expert users to interact with relational databases and has long been a central task for natural language interfaces to data. While the WikiSQL dataset played a key role in early NL2SQL research, its usage has declined due to structural and annotation issues, including case sensitivity inconsistencies, data type mismatches, syntax errors, and unanswered questions. We present LLMSQL, a systematic revision and transformation of WikiSQL designed for the LLM era. We classify these errors and implement automated methods for cleaning and re-annotation. To assess the impact of these improvements, we evaluated multiple large language models (LLMs), including Gemma 3, LLaMA 3.2, Mistral 7B, gpt-oss 20B, Phi-3.5 Mini, Qwen 2.5, OpenAI o4-mini, DeepSeek R1 and others. Rather than serving as an update, LLMSQL is introduced as an LLM-ready benchmark: unlike the original WikiSQL, tailored for pointer-network models selecting tokens from input, LLMSQL provides clean natural language questions and full SQL queries as plain text, enabling straightforward generation and evaluation for modern natural language-to-SQL models.
Authors: Dzmitry Pihulski, Jan Koco\'n
Abstract: We explore how large language models (LLMs) assess offensiveness in political discourse when prompted to adopt specific political and cultural perspectives. Using a multilingual subset of the MD-Agreement dataset centered on tweets from the 2020 US elections, we evaluate several recent LLMs - including DeepSeek-R1, o4-mini, GPT-4.1-mini, Qwen3, Gemma, and Mistral - tasked with judging tweets as offensive or non-offensive from the viewpoints of varied political personas (far-right, conservative, centrist, progressive) across English, Polish, and Russian contexts. Our results show that larger models with explicit reasoning abilities (e.g., DeepSeek-R1, o4-mini) are more consistent and sensitive to ideological and cultural variation, while smaller models often fail to capture subtle distinctions. We find that reasoning capabilities significantly improve both the personalization and interpretability of offensiveness judgments, suggesting that such mechanisms are key to adapting LLMs for nuanced sociopolitical text classification across languages and ideologies.
Authors: Yihao Wu, Tianrui Wang, Yizhou Peng, Yi-Wen Chao, Xuyi Zhuang, Xinsheng Wang, Shunshun Yin, Ziyang Ma
Abstract: While biases in large language models (LLMs), such as stereotypes and cultural tendencies in outputs, have been examined and identified, their presence and characteristics in spoken dialogue models (SDMs) with audio input and output remain largely unexplored. Paralinguistic features, such as age, gender, and accent, can affect model outputs; when compounded by multi-turn conversations, these effects may exacerbate biases, with potential implications for fairness in decision-making and recommendation tasks. In this paper, we systematically evaluate biases in speech LLMs and study the impact of multi-turn dialogues with repeated negative feedback. Bias is measured using Group Unfairness Score (GUS) for decisions and similarity-based normalized statistics rate (SNSR) for recommendations, across both open-source models like Qwen2.5-Omni and GLM-4-Voice, as well as closed-source APIs such as GPT-4o Audio and Gemini-2.5-Flash. Our analysis reveals that closed-source models generally exhibit lower bias, while open-source models are more sensitive to age and gender, and recommendation tasks tend to amplify cross-group disparities. We found that biased decisions may persist in multi-turn conversations. This work provides the first systematic study of biases in end-to-end spoken dialogue models, offering insights towards fair and reliable audio-based interactive systems. To facilitate further research, we release the FairDialogue dataset and evaluation code.
Authors: Oumar Kane, Mouhamad M. Allaya, Dame Samb, Mamadou Bousso
Abstract: This study examines the application of artificial intelligence (AI) and large language models (LLM) to improve access to legal texts in Senegal's judicial system. The emphasis is on the difficulties of extracting and organizing legal documents, highlighting the need for better access to judicial information. The research successfully extracted 7,967 articles from various legal documents, particularly focusing on the Land and Public Domain Code. A detailed graph database was developed, which contains 2,872 nodes and 10,774 relationships, aiding in the visualization of interconnections within legal texts. In addition, advanced triple extraction techniques were utilized for knowledge, demonstrating the effectiveness of models such as GPT-4o, GPT-4, and Mistral-Large in identifying relationships and relevant metadata. Through these technologies, the aim is to create a solid framework that allows Senegalese citizens and legal professionals to more effectively understand their rights and responsibilities.
Authors: Shreya Saha, Shurui Li, Greta Tuckute, Yuanning Li, Ru-Yuan Zhang, Leila Wehbe, Evelina Fedorenko, Meenakshi Khosla
Abstract: The human language system represents both linguistic forms and meanings, but the abstractness of the meaning representations remains debated. Here, we searched for abstract representations of meaning in the language cortex by modeling neural responses to sentences using representations from vision and language models. When we generate images corresponding to sentences and extract vision model embeddings, we find that aggregating across multiple generated images yields increasingly accurate predictions of language cortex responses, sometimes rivaling large language models. Similarly, averaging embeddings across multiple paraphrases of a sentence improves prediction accuracy compared to any single paraphrase. Enriching paraphrases with contextual details that may be implicit (e.g., augmenting "I had a pancake" to include details like "maple syrup") further increases prediction accuracy, even surpassing predictions based on the embedding of the original sentence, suggesting that the language system maintains richer and broader semantic representations than language models. Together, these results demonstrate the existence of highly abstract, form-independent meaning representations within the language cortex.
Authors: Guanghao Li, Zhihui Fu, Min Fang, Qibin Zhao, Ming Tang, Chun Yuan, Jun Wang
Abstract: As large language models (LLMs) scale up, accuracy improves, but the autoregressive (AR) nature of decoding increases latency since each token requires a serial forward pass. Speculative decoding addresses this by employing a fast drafter to propose multi-token drafts, which are then verified in parallel by the target model. However, many deployments still rely on AR drafters, where sequential passes limit wall-clock gains. We revisit the drafting stage and present DiffuSpec, a training-free drop-in framework that uses a pretrained diffusion language model (DLM) to produce multi-token drafts in a single forward pass, while remaining compatible with standard AR verifiers. Because DLM drafts are generated under bidirectional conditioning, parallel per-position candidates form a token lattice in which the locally highest-probability token at each position need not form a causal left-to-right path. Moreover, DLM drafting requires pre-specifying a draft length, inducing a speed-quality trade-off. To address these challenges, we introduce two practical components: (i) a causal-consistency path search (CPS) over this lattice that extracts a left-to-right path aligned with AR verification; and (ii) an adaptive draft-length (ADL) controller that adjusts next proposal size based on recent acceptance feedback and realized generated length. Across benchmarks, DiffuSpec yields up to 3x wall-clock speedup, establishing diffusion-based drafting as a robust alternative to autoregressive drafters for speculative decoding.
Authors: Jiashu Ye, Tong Wu, Weiwen Chen, Hao Zhang, Zeteng Lin, Xingxing Li, Shujuan Weng, Manni Zhu, Xin Yuan, Xinlong Hong, Jingjie Li, Junyu Zheng, Zhijiong Huang, Jing Tang
Abstract: Improving air quality and addressing climate change relies on accurate understanding and analysis of air pollutant and greenhouse gas emissions. However, emission-related knowledge is often fragmented and highly specialized, while existing methods for accessing and compiling emissions data remain inefficient. These issues hinder the ability of non-experts to interpret emissions information, posing challenges to research and management. To address this, we present Emission-GPT, a knowledge-enhanced large language model agent tailored for the atmospheric emissions domain. Built on a curated knowledge base of over 10,000 documents (including standards, reports, guidebooks, and peer-reviewed literature), Emission-GPT integrates prompt engineering and question completion to support accurate domain-specific question answering. Emission-GPT also enables users to interactively analyze emissions data via natural language, such as querying and visualizing inventories, analyzing source contributions, and recommending emission factors for user-defined scenarios. A case study in Guangdong Province demonstrates that Emission-GPT can extract key insights--such as point source distributions and sectoral trends--directly from raw data with simple prompts. Its modular and extensible architecture facilitates automation of traditionally manual workflows, positioning Emission-GPT as a foundational tool for next-generation emission inventory development and scenario-based assessment.
Authors: Mingze Zhong, Meng Fang, Zijing Shi, Yuxuan Huang, Shunfeng Zheng, Yali Du, Ling Chen, Jun Wang
Abstract: The Spiral of Silence (SoS) theory holds that individuals with minority views often refrain from speaking out for fear of social isolation, enabling majority positions to dominate public discourse. When the 'agents' are large language models (LLMs), however, the classical psychological explanation is not directly applicable, since SoS was developed for human societies. This raises a central question: can SoS-like dynamics nevertheless emerge from purely statistical language generation in LLM collectives? We propose an evaluation framework for examining SoS in LLM agents. Specifically, we consider four controlled conditions that systematically vary the availability of 'History' and 'Persona' signals. Opinion dynamics are assessed using trend tests such as Mann-Kendall and Spearman's rank, along with concentration measures including kurtosis and interquartile range. Experiments across open-source and closed-source models show that history and persona together produce strong majority dominance and replicate SoS patterns; history signals alone induce strong anchoring; and persona signals alone foster diverse but uncorrelated opinions, indicating that without historical anchoring, SoS dynamics cannot emerge. The work bridges computational sociology and responsible AI design, highlighting the need to monitor and mitigate emergent conformity in LLM-agent systems.
Authors: Haojie Ouyang, Jianwei Lv, Lei Ren, Chen Wei, Xiaojie Wang, Fangxiang Feng
Abstract: Transformer-based large models excel in natural language processing and computer vision, but face severe computational inefficiencies due to the self-attention's quadratic complexity with input tokens. Recently, researchers have proposed a series of methods based on block selection and compression to alleviate this problem, but they either have issues with semantic incompleteness or poor training-inference efficiency. To comprehensively address these challenges, we propose ChunkLLM, a lightweight and pluggable training framework. Specifically, we introduce two components: QK Adapter (Q-Adapter and K-Adapter) and Chunk Adapter. The former is attached to each Transformer layer, serving dual purposes of feature compression and chunk attention acquisition. The latter operates at the bottommost layer of the model, functioning to detect chunk boundaries by leveraging contextual semantic information. During the training phase, the parameters of the backbone remain frozen, with only the QK Adapter and Chunk Adapter undergoing training. Notably, we design an attention distillation method for training the QK Adapter, which enhances the recall rate of key chunks. During the inference phase, chunk selection is triggered exclusively when the current token is detected as a chunk boundary, thereby accelerating model inference. Experimental evaluations are conducted on a diverse set of long-text and short-text benchmark datasets spanning multiple tasks. ChunkLLM not only attains comparable performance on short-text benchmarks but also maintains 98.64% of the performance on long-context benchmarks while preserving a 48.58% key-value cache retention rate. Particularly, ChunkLLM attains a maximum speedup of 4.48x in comparison to the vanilla Transformer in the processing of 120K long texts.
Authors: Matei-Iulian Cocu, R\u{a}zvan-Cosmin Cristia, Adrian Marius Dumitran
Abstract: In this case study, we select a set of controversial Romanian historical questions and ask multiple Large Language Models to answer them across languages and contexts, in order to assess their biases. Besides being a study mainly performed for educational purposes, the motivation also lies in the recognition that history is often presented through altered perspectives, primarily influenced by the culture and ideals of a state, even through large language models. Since they are often trained on certain data sets that may present certain ambiguities, the lack of neutrality is subsequently instilled in users. The research process was carried out in three stages, to confirm the idea that the type of response expected can influence, to a certain extent, the response itself; after providing an affirmative answer to some given question, an LLM could shift its way of thinking after being asked the same question again, but being told to respond with a numerical value of a scale. Results show that binary response stability is relatively high but far from perfect and varies by language. Models often flip stance across languages or between formats; numeric ratings frequently diverge from the initial binary choice, and the most consistent models are not always those judged most accurate or neutral. Our research brings to light the predisposition of models to such inconsistencies, within a specific contextualization of the language for the question asked.
Authors: Kuntai Cai, Juncheng Liu, Xianglin Yang, Zhaojie Niu, Xiaokui Xiao, Xing Chen
Abstract: Large language model (LLM) agents typically receive two kinds of context: (i) environment-level manuals that define interaction interfaces and global rules, and (ii) task-level guidance or demonstrations tied to specific goals. In this work, we identify a crucial but overlooked third type of context, instance-level context, which consists of verifiable and reusable facts tied to a specific environment instance, such as object locations, crafting recipes, and local rules. We argue that the absence of instance-level context is a common source of failure for LLM agents in complex tasks, as success often depends not only on reasoning over global rules or task prompts but also on making decisions based on precise and persistent facts. Acquiring such context requires more than memorization: the challenge lies in efficiently exploring, validating, and formatting these facts under tight interaction budgets. We formalize this problem as Instance-Level Context Learning (ILCL) and introduce our task-agnostic method to solve it. Our method performs a guided exploration, using a compact TODO forest to intelligently prioritize its next actions and a lightweight plan-act-extract loop to execute them. This process automatically produces a high-precision context document that is reusable across many downstream tasks and agents, thereby amortizing the initial exploration cost. Experiments across TextWorld, ALFWorld, and Crafter demonstrate consistent gains in both success and efficiency: for instance, ReAct's mean success rate in TextWorld rises from 37% to 95%, while IGE improves from 81% to 95%. By transforming one-off exploration into persistent, reusable knowledge, our method complements existing contexts to enable more reliable and efficient LLM agents.
Authors: Minsung Kim, Dong-Kyum Kim, Jea Kwon, Nakyeong Yang, Kyomin Jung, Meeyoung Cha
Abstract: Large language models often encounter conflicts between in-context knowledge retrieved at inference time and parametric knowledge acquired during pretraining. Models that accept external knowledge uncritically are vulnerable to misinformation, whereas models that adhere rigidly to parametric knowledge fail to benefit from retrieval. Despite the widespread adoption of retrieval-augmented generation, we still lack a systematic understanding of what shapes knowledge-arbitration strategies during training. This gap risks producing pretrained models with undesirable arbitration behaviors and, consequently, wasting substantial computational resources after the pretraining budget has already been spent. To address this problem, we present the first controlled study of how training conditions influence models' use of in-context and parametric knowledge, and how they arbitrate between them. We train transformer-based language models on a synthetic biographies corpus while systematically controlling various conditions. Our experiments reveal that intra-document repetition of facts fosters the development of both parametric and in-context capabilities. Moreover, training on a corpus that contains inconsistent information or distributional skew encourages models to develop robust strategies for leveraging parametric and in-context knowledge. Rather than viewing these non-ideal properties as artifacts to remove, our results indicate that they are important for learning robust arbitration. These insights offer concrete, empirical guidance for pretraining models that harmoniously integrate parametric and in-context knowledge.
Authors: Hadi Pouransari, David Grangier, C Thomas, Michael Kirchhof, Oncel Tuzel
Abstract: The impressive performance gains of modern language models currently rely on scaling parameters: larger models store more world knowledge and reason better. Yet compressing all world knowledge into parameters is unnecessary, as only a fraction is used per prompt, and impractical for edge devices with limited inference-time memory and compute. We address this shortcoming by a memory-augmented architecture and a pretraining strategy aligned with existing hardware paradigms. We introduce small language models that access large hierarchical parametric memory banks encoding world knowledge. During pretraining and inference, we fetch a small, context-dependent memory block and add it to the model. Our pretraining learns to store long-tail world knowledge in the memory parameters, while the small language model acts as an anchor capturing common knowledge and general reasoning abilities. Through trillion-token-scale experiments, we show significant gains: a 160M-parameters model augmented with an 18M-parameters memory fetched from a 4.6B memory bank obtains comparable performance to a regular model with more than 2x the parameters. Through extensive experiments, we study the optimal type and size of parametric memories in transformers, scaling them to over 21B parameters. We find that our proposed hierarchical feed-forward memories work robustly across transformer architectures, whether added during pretraining or post-hoc.
Authors: Aakriti Agrawal, Rohith Aralikatti, Anirudh Satheesh, Souradip Chakraborty, Amrit Singh Bedi, Furong Huang
Abstract: Large Language Models (LLMs) have demonstrated exceptional capabilities, yet selecting the most reliable response from multiple LLMs remains a challenge, particularly in resource-constrained settings. Existing approaches often depend on costly external verifiers, human evaluators, or self-consistency techniques that require multiple samples from a single model. While multi-LLM systems produce more diverse responses than single models and thus have greater potential, they often underperform compared to single LLM self-consistency. We propose a principled, novel and computationally efficient method to select the best response from multiple different LLMs using a calibrated log-likelihood score, implicitly leveraging the inherent knowledge and confidence of these models. Our method demonstrates improvements of approx. 4%, 3%, and 5% across both debate (multi-round LLM discussions) and non-debate (Best-of-N with multiple LLMs) settings on GSM8K, MMLU (6 subsets), and ARC datasets respectively.
Authors: Haoyue Bai, Haoyu Wang, Shengyu Chen, Zhengzhang Chen, Lu-An Tang, Wei Cheng, Haifeng Chen, Yanjie Fu
Abstract: Large Language Models (LLMs) have shown remarkable performance on general Question Answering (QA), yet they often struggle in domain-specific scenarios where accurate and up-to-date information is required. Retrieval-Augmented Generation (RAG) addresses this limitation by enriching LLMs with external knowledge, but existing systems primarily rely on unstructured documents, while largely overlooking relational databases, which provide precise, timely, and efficiently queryable factual information, serving as indispensable infrastructure in domains such as finance, healthcare, and scientific research. Motivated by this gap, we conduct a systematic analysis that reveals three central observations: (i) databases and documents offer complementary strengths across queries, (ii) naively combining both sources introduces noise and cost without consistent accuracy gains, and (iii) selecting the most suitable source for each query is crucial to balance effectiveness and efficiency. We further observe that query types show consistent regularities in their alignment with retrieval paths, suggesting that routing decisions can be effectively guided by systematic rules that capture these patterns. Building on these insights, we propose a rule-driven routing framework. A routing agent scores candidate augmentation paths based on explicit rules and selects the most suitable one; a rule-making expert agent refines the rules over time using QA feedback to maintain adaptability; and a path-level meta-cache reuses past routing decisions for semantically similar queries to reduce latency and cost. Experiments on three QA benchmarks demonstrate that our framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.
Authors: Yinyi Luo, Zhexian Zhou, Hao Chen, Kai Qiu, Marios Savvides, Yixuan Li, Jindong Wang
Abstract: Knowledge editing and machine unlearning are two popular approaches for large language models (LLMs) to stay up-to-date. However, the knowledge updating mechanism of LLMs remains largely unexplored due to insufficient, isolated, and small-scale evaluation. For instance, are LLMs similar to humans in modifying certain knowledge? What differs editing and unlearning as training data increases? This paper proposes KnowledgeSmith, a unified framework to systematically understand the updating mechanism of LLMs. We first cast editing and unlearning as instances of one constrained optimization problem. Then, we propose an automatic dataset generator that provides structured interventions across multiple graph levels and data scales, enabling controlled studies of how different modification strategies propagate through model knowledge. Extensive experiments demonstrate nuanced insights over knowledge propagation, plasticity scaling, consistency, and robustness. For instance, our results show that LLMs do not exhibit similar updating as humans for different levels of knowledge, and there exists consistency-capacity trade-off. We hope our findings can offer suggestions to the design of more reliable and scalable strategies. Code: https://github.com/AIFrontierLab/KnowledgeSmith.git
Authors: Manasi Patwardhan, Ayush Agarwal, Shabbirhussain Bhaisaheb, Aseem Arora, Lovekesh Vig, Sunita Sarawagi
Abstract: The performance of Large Language Models (LLMs) for translating Natural Language (NL) queries into SQL varies significantly across databases (DBs). NL queries are often expressed using a domain specific vocabulary, and mapping these to the correct SQL requires an understanding of the embedded domain expressions, their relationship to the DB schema structure. Existing benchmarks rely on unrealistic, ad-hoc query specific textual hints for expressing domain knowledge. In this paper, we propose a systematic framework for associating structured domain statements at the database level. We present retrieval of relevant structured domain statements given a user query using sub-string level match. We evaluate on eleven realistic DB schemas covering diverse domains across five open-source and proprietary LLMs and demonstrate that (1) DB level structured domain statements are more practical and accurate than existing ad-hoc query specific textual domain statements, and (2) Our sub-string match based retrieval of relevant domain statements provides significantly higher accuracy than other retrieval approaches.
Authors: Sophie L. Wang, Phillip Isola, Brian Cheung
Abstract: Large language models (LLMs) trained purely on text ostensibly lack any direct perceptual experience, yet their internal representations are implicitly shaped by multimodal regularities encoded in language. We test the hypothesis that explicit sensory prompting can surface this latent structure, bringing a text-only LLM into closer representational alignment with specialist vision and audio encoders. When a sensory prompt tells the model to 'see' or 'hear', it cues the model to resolve its next-token predictions as if they were conditioned on latent visual or auditory evidence that is never actually supplied. Our findings reveal that lightweight prompt engineering can reliably activate modality-appropriate representations in purely text-trained LLMs.
Authors: Vladimir Shaposhnikov, Aleksandr Nesterov, Ilia Kopanichuk, Ivan Bakulin, Egor Zhelvakov, Ruslan Abramov, Ekaterina Tsapieva, Dmitry V. Dylov, Ivan Oseledets
Abstract: We present CLARITY (Clinical Assistant for Routing, Inference, and Triage), an AI-driven platform designed to facilitate patient-to-specialist routing, clinical consultations, and severity assessment of patients' conditions. Its hybrid architecture combines a Finite State Machine (FSM) for structured dialogue flows with collaborative agents that employ Large Language Model (LLM) to analyze symptoms and prioritize referrals to appropriate specialists. Built on a modular microservices framework, CLARITY ensures safe, efficient, and robust performance, flexible and readily scalable to meet the demands of existing workflows and IT solutions in healthcare. We report integration of our clinical assistant into a large-scale nation-wide inter-hospital IT platform, with over 55,000 content-rich user dialogues completed within the two months of deployment, 2,500 of which were expert-annotated for a consequent validation. The validation results show that CLARITY surpasses human-level performance in terms of the first-attempt routing precision, naturally requiring up to 3 times shorter duration of the consultation than with a human.
Authors: Laura Ying Schulz, Daniel Mitropolsky, Tomaso Poggio
Abstract: We introduce a new framework for understanding how language models acquire syntax. While large models achieve impressive results, little is known about their learning dynamics. Our approach starts with the observation that most domains of interest, such as natural language syntax, coding languages, arithmetic problems, are captured by probabilistic context-free grammars (PCFGs). We study the learning dynamics of small models trained on synthetic languages generated from PCFGs, enabling precise control over grammar complexity, recursion depth, and subgrammar structure. We prove several general, recursive formulae for the training loss and Kullback-Leibler divergence over the subgrammar structure of a PCFG. Empirically, we find that unlike children, who first master simple substructures before progressing to more complex constructions, transformers reduce loss across all subgrammars in parallel. We further show that subgrammar pretraining can improve the final loss for smaller models, and that pretrained models develop internal representations more aligned with the grammar's substructure. Finally, we demonstrate that models struggle with deeper recursive structures (a limitation even of large language models), revealing fundamental challenges in how neural networks represent hierarchical syntax. Overall, our work initiates the study of the learning dynamics of transformers on PCFGs as a versatile testbed for probing learning in language models, opening a research direction with many open questions.
Authors: Anant Gupta, Karthik Singaravadivelan, Zekun Wang
Abstract: Neural document retrieval often treats a corpus as a flat cloud of vectors scored at a single granularity, leaving corpus structure underused and explanations opaque. We use Cobweb--a hierarchy-aware framework--to organize sentence embeddings into a prototype tree and rank documents via coarse-to-fine traversal. Internal nodes act as concept prototypes, providing multi-granular relevance signals and a transparent rationale through retrieval paths. We instantiate two inference approaches: a generalized best-first search and a lightweight path-sum ranker. We evaluate our approaches on MS MARCO and QQP with encoder (e.g., BERT/T5) and decoder (GPT-2) representations. Our results show that our retrieval approaches match the dot product search on strong encoder embeddings while remaining robust when kNN degrades: with GPT-2 vectors, dot product performance collapses whereas our approaches still retrieve relevant results. Overall, our experiments suggest that Cobweb provides competitive effectiveness, improved robustness to embedding quality, scalability, and interpretable retrieval via hierarchical prototypes.
Authors: Sicheng Dong, Vahid Zolfaghari, Nenad Petrovic, Alois Knoll
Abstract: Large language models (LLMs) has become a significant research focus and is utilized in various fields, such as text generation and dialog systems. One of the most essential applications of LLM is Retrieval Augmented Generation (RAG), which greatly enhances generated content's reliability and relevance. However, evaluating RAG systems remains a challenging task. Traditional evaluation metrics struggle to effectively capture the key features of modern LLM-generated content that often exhibits high fluency and naturalness. Inspired by the RAGAS tool, a well-known RAG evaluation framework, we extended this framework into a KG-based evaluation paradigm, enabling multi-hop reasoning and semantic community clustering to derive more comprehensive scoring metrics. By incorporating these comprehensive evaluation criteria, we gain a deeper understanding of RAG systems and a more nuanced perspective on their performance. To validate the effectiveness of our approach, we compare its performance with RAGAS scores and construct a human-annotated subset to assess the correlation between human judgments and automated metrics. In addition, we conduct targeted experiments to demonstrate that our KG-based evaluation method is more sensitive to subtle semantic differences in generated outputs. Finally, we discuss the key challenges in evaluating RAG systems and highlight potential directions for future research.
Authors: Tol\'ul\d{o}p\'e \`Og\'unr\`em\'i, Christopher D. Manning, Dan Jurafsky, Karen Livescu
Abstract: Spoken language models (SLMs) that integrate speech with large language models (LMs) rely on modality adapters (MAs) to map the output of speech encoders to a representation that is understandable to the decoder LM. Yet we know very little about how these crucial MAs transform representations. Here we examine the MA output representation in three SLMs (SALMONN, Qwen2-Audio and Phi-4-Multimodal-Instruct). By finding the nearest decoder LM token to an MA representation, we uncover two strategies for MA representations. For models using a Whisper encoder, MAs appear to represent the meaning of the input using an English-based interlingua, allowing them to handle languages unseen in instruction tuning. For models that don't, like Phi-4-Multimodal-Instruct, MAs instead represent the phonetics of the input, but expressed with English words. We hypothesise that which arises depends on whether the speech encoder is trained only for speech recognition or also for translation.
Authors: Jingyi Sun, Pepa Atanasova, Sagnik Ray Choudhury, Sekh Mainul Islam, Isabelle Augenstein
Abstract: Context utilisation, the ability of Language Models (LMs) to incorporate relevant information from the provided context when generating responses, remains largely opaque to users, who cannot determine whether models draw from parametric memory or provided context, nor identify which specific context pieces inform the response. Highlight explanations (HEs) offer a natural solution as they can point the exact context pieces and tokens that influenced model outputs. However, no existing work evaluates their effectiveness in accurately explaining context utilisation. We address this gap by introducing the first gold standard HE evaluation framework for context attribution, using controlled test cases with known ground-truth context usage, which avoids the limitations of existing indirect proxy evaluations. To demonstrate the framework's broad applicability, we evaluate four HE methods -- three established techniques and MechLight, a mechanistic interpretability approach we adapt for this task -- across four context scenarios, four datasets, and five LMs. Overall, we find that MechLight performs best across all context scenarios. However, all methods struggle with longer contexts and exhibit positional biases, pointing to fundamental challenges in explanation accuracy that require new approaches to deliver reliable context utilisation explanations at scale.
Authors: Fulei Zhang, Zhou Yu
Abstract: As Large Language Models (LLMs) are increasingly deployed in customer-facing applications, a critical yet underexplored question is how users communicate differently with LLM chatbots compared to human agent. In this study, we present empirical evidence that users adopt distinct communication styles when users interact with chatbots versus human agents. Our analysis reveals significant differences in grammatical fluency, politeness, and lexical diversity in user language between the two settings. These findings suggest that models trained exclusively on human-human interaction data may not adequately accommodate the communication style shift that occurs once an LLM chatbot is deployed. To enhance LLM robustness to post-launch communication style changes, we experimented with two strategies: (1) data augmentation during the post-training phase and (2) inference-time user message reformulation. Our results indicate that models trained on stylistically diverse datasets significantly outperform those trained exclusively on original or stylistically uniform datasets, while inference-time reformulation proved less effective. These insights help us to better adapt our models for improved LLM-user interaction experiences.
Authors: Rui Qi, Zhibo Man, Yufeng Chen, Fengran Mo, Jinan Xu, Kaiyu Huang
Abstract: Recent developments have enabled Large Language Models (LLMs) to engage in complex reasoning tasks through deep thinking. However, the capacity of reasoning has not been successfully transferred to non-high-resource languages due to resource constraints, which struggles with multilingual reasoning tasks. To this end, we propose Structured-of-Thought (SoT), a training-free method that improves the performance on multilingual reasoning through a multi-step transformation: Language Thinking Transformation and Structured Knowledge Transformation. The SoT method converts language-specific semantic information into language-agnostic structured representations, enabling the models to understand the query in different languages more sophisticated. Besides, SoT effectively guides LLMs toward more concentrated reasoning to maintain consistent underlying reasoning pathways when handling cross-lingual variations in expression. Experimental results demonstrate that SoT outperforms several strong baselines on multiple multilingual reasoning benchmarks when adapting to various backbones of LLMs. It can also be integrated with other training-free strategies for further improvements. Our code is available at https://github.com/Cherry-qwq/SoT.
Authors: Shijian Deng, Kai Wang, Tianyu Yang, Harsh Singh, Yapeng Tian
Abstract: Recent advancements in self-improvement for Large Language Models (LLMs) have efficiently enhanced model capabilities without significantly increasing costs, particularly in terms of human effort. While this area is still relatively young, its extension to the multimodal domain holds immense potential for leveraging diverse data sources and developing more general self-improving models. This survey is the first to provide a comprehensive overview of self-improvement in Multimodal LLMs (MLLMs). We provide a structured overview of the current literature and discuss methods from three perspectives: 1) data collection, 2) data organization, and 3) model optimization, to facilitate the further development of self-improvement in MLLMs. We also include commonly used evaluations and downstream applications. Finally, we conclude by outlining open challenges and future research directions.
Authors: Yavuz Bakman, Sungmin Kang, Zhiqi Huang, Duygu Nur Yaldiz, Catarina G. Bel\'em, Chenyang Zhu, Anoop Kumar, Alfy Samuel, Salman Avestimehr, Daben Liu, Sai Praneeth Karimireddy
Abstract: Uncertainty Quantification (UQ) research has primarily focused on closed-book factual question answering (QA), while contextual QA remains unexplored, despite its importance in real-world applications. In this work, we focus on UQ for the contextual QA task and propose a theoretically grounded approach to quantify epistemic uncertainty. We begin by introducing a task-agnostic, token-level uncertainty measure defined as the cross-entropy between the predictive distribution of the given model and the unknown true distribution. By decomposing this measure, we isolate the epistemic component and approximate the true distribution by a perfectly prompted, idealized model. We then derive an upper bound for epistemic uncertainty and show that it can be interpreted as semantic feature gaps in the given model's hidden representations relative to the ideal model. We further apply this generic framework to the contextual QA task and hypothesize that three features approximate this gap: context-reliance (using the provided context rather than parametric knowledge), context comprehension (extracting relevant information from context), and honesty (avoiding intentional lies). Using a top-down interpretability approach, we extract these features by using only a small number of labeled samples and ensemble them to form a robust uncertainty score. Experiments on multiple QA benchmarks in both in-distribution and out-of-distribution settings show that our method substantially outperforms state-of-the-art unsupervised (sampling-free and sampling-based) and supervised UQ methods, achieving up to a 13-point PRR improvement while incurring a negligible inference overhead.
Authors: Yubo Li, Ramayya Krishnan, Rema Padman
Abstract: Large Language Models (LLMs) have revolutionized conversational AI, yet their robustness in extended multi-turn dialogues remains poorly understood. Existing evaluation frameworks focus on static benchmarks and single-turn assessments, failing to capture the temporal dynamics of conversational degradation that characterize real-world interactions. In this work, we present the first comprehensive survival analysis of conversational AI robustness, analyzing 36,951 conversation turns across 9 state-of-the-art LLMs to model failure as a time-to-event process. Our survival modeling framework-employing Cox proportional hazards, Accelerated Failure Time, and Random Survival Forest approaches-reveals extraordinary temporal dynamics. We find that abrupt, prompt-to-prompt(P2P) semantic drift is catastrophic, dramatically increasing the hazard of conversational failure. In stark contrast, gradual, cumulative drift is highly protective, vastly reducing the failure hazard and enabling significantly longer dialogues. AFT models with interactions demonstrate superior performance, achieving excellent discrimination and exceptional calibration. These findings establish survival analysis as a powerful paradigm for evaluating LLM robustness, offer concrete insights for designing resilient conversational agents, and challenge prevailing assumptions about the necessity of semantic consistency in conversational AI Systems.
Authors: Srinivas Billa, Xiaonan Jing
Abstract: Results on existing LLM benchmarks capture little information over the model capabilities in low-resource tasks, making it difficult to develop effective solutions in these domains. To address these challenges, we curated 14 travel-domain datasets spanning 7 common NLP tasks using anonymised data from real-world scenarios, and analysed the performance across LLMs. We report on the accuracy, scaling behaviour, and reasoning capabilities of LLMs in a variety of tasks. Our results confirm that general benchmarking results are insufficient for understanding model performance in low-resource tasks. Despite the amount of training FLOPs, out-of-the-box LLMs hit performance bottlenecks in complex, domain-specific scenarios. Furthermore, reasoning provides a more significant boost for smaller LLMs by making the model a better judge on certain tasks.
Authors: KM Pooja, Cheng Long, Aixin Sun
Abstract: The task of entity linking, which involves associating mentions with their respective entities in a knowledge graph, has received significant attention due to its numerous potential applications. Recently, various multimodal entity linking (MEL) techniques have been proposed, targeted to learn comprehensive embeddings by leveraging both text and vision modalities. The selection of high-quality negative samples can potentially play a crucial role in metric/representation learning. However, to the best of our knowledge, this possibility remains unexplored in existing literature within the framework of MEL. To fill this gap, we address the multimodal entity linking problem in a generative adversarial setting where the generator is responsible for generating high-quality negative samples, and the discriminator is assigned the responsibility for the metric learning tasks. Since the generator is involved in generating samples, which is a discrete process, we optimize it using policy gradient techniques and propose a policy gradient-based generative adversarial network for multimodal entity linking (PGMEL). Experimental results based on Wiki-MEL, Richpedia-MEL and WikiDiverse datasets demonstrate that PGMEL learns meaningful representation by selecting challenging negative samples and outperforms state-of-the-art methods.
Authors: Santhosh G S, Akshay Govind S, Gokul S Krishnan, Balaraman Ravindran, Sriraam Natarajan
Abstract: Large Language Models (LLMs) have gained significant traction across critical domains owing to their impressive contextual understanding and generative capabilities. However, their increasing deployment in high stakes applications necessitates rigorous evaluation of embedded biases, particularly in culturally diverse contexts like India where existing embedding-based bias assessment methods often fall short in capturing nuanced stereotypes. We propose an evaluation framework based on a encoder trained using contrastive learning that captures fine-grained bias through embedding similarity. We also introduce a novel dataset - IndiCASA (IndiBias-based Contextually Aligned Stereotypes and Anti-stereotypes) comprising 2,575 human-validated sentences spanning five demographic axes: caste, gender, religion, disability, and socioeconomic status. Our evaluation of multiple open-weight LLMs reveals that all models exhibit some degree of stereotypical bias, with disability related biases being notably persistent, and religion bias generally lower likely due to global debiasing efforts demonstrating the need for fairer model development.
Authors: Hangfan Zhang, Siyuan Xu, Zhimeng Guo, Huaisheng Zhu, Shicheng Liu, Xinrun Wang, Qiaosheng Zhang, Yang Chen, Peng Ye, Lei Bai, Shuyue Hu
Abstract: Reinforcement learning (RL) has demonstrated potential in enhancing the reasoning capabilities of large language models (LLMs), but such training typically demands substantial efforts in creating and annotating data. In this work, we explore improving LLMs through RL with minimal data. Our approach alternates between the LLM proposing a task and then attempting to solve it. To minimize data dependency, we introduce two novel mechanisms grounded in self-awareness: (1) self-aware difficulty prediction, where the model learns to assess task difficulty relative to its own abilities and prioritize challenging yet solvable tasks, and (2) self-aware limit breaking, where the model recognizes when a task is beyond its capability boundary and proactively requests external data to break through that limit. Extensive experiments on nine benchmarks showing a 53.8% relative improvement with less than 1.2% extra data demonstrate the efficacy of self-aware RL and underscore the promise of self-evolving agent training.
Authors: Tien Phat Nguyen, Vu Minh Ngo, Tung Nguyen, Linh Van Ngo, Duc Anh Nguyen, Sang Dinh, Trung Le
Abstract: Cross-lingual topic modeling aims to uncover shared semantic themes across languages. Several methods have been proposed to address this problem, leveraging both traditional and neural approaches. While previous methods have achieved some improvements in topic diversity, they often struggle to ensure high topic coherence and consistent alignment across languages. We propose XTRA (Cross-Lingual Topic Modeling with Topic and Representation Alignments), a novel framework that unifies Bag-of-Words modeling with multilingual embeddings. XTRA introduces two core components: (1) representation alignment, aligning document-topic distributions via contrastive learning in a shared semantic space; and (2) topic alignment, projecting topic-word distributions into the same space to enforce crosslingual consistency. This dual mechanism enables XTRA to learn topics that are interpretable (coherent and diverse) and well-aligned across languages. Experiments on multilingual corpora confirm that XTRA significantly outperforms strong baselines in topic coherence, diversity, and alignment quality. Code and reproducible scripts are available at https: //github.com/tienphat140205/XTRA.
Authors: Matej Gjurkovi\'c
Abstract: Personality refers to individual differences in behavior, thinking, and feeling. With the growing availability of digital footprints, especially from social media, automated methods for personality assessment have become increasingly important. Natural language processing (NLP) enables the analysis of unstructured text data to identify personality indicators. However, two main challenges remain central to this thesis: the scarcity of large, personality-labeled datasets and the disconnect between personality psychology and NLP, which restricts model validity and interpretability. To address these challenges, this thesis presents two datasets -- MBTI9k and PANDORA -- collected from Reddit, a platform known for user anonymity and diverse discussions. The PANDORA dataset contains 17 million comments from over 10,000 users and integrates the MBTI and Big Five personality models with demographic information, overcoming limitations in data size, quality, and label coverage. Experiments on these datasets show that demographic variables influence model validity. In response, the SIMPA (Statement-to-Item Matching Personality Assessment) framework was developed - a computational framework for interpretable personality assessment that matches user-generated statements with validated questionnaire items. By using machine learning and semantic similarity, SIMPA delivers personality assessments comparable to human evaluations while maintaining high interpretability and efficiency. Although focused on personality assessment, SIMPA's versatility extends beyond this domain. Its model-agnostic design, layered cue detection, and scalability make it suitable for various research and practical applications involving complex label taxonomies and variable cue associations with target concepts.
Authors: Tengjun Ni, Xin Yuan, Shenghong Li, Kai Wu, Ren Ping Liu, Wei Ni, Wenjie Zhang
Abstract: Recent progress in retrieval-augmented generation (RAG) has led to more accurate and interpretable multi-hop question answering (QA). Yet, challenges persist in integrating iterative reasoning steps with external knowledge retrieval. To address this, we introduce StepChain GraphRAG, a framework that unites question decomposition with a Breadth-First Search (BFS) Reasoning Flow for enhanced multi-hop QA. Our approach first builds a global index over the corpus; at inference time, only retrieved passages are parsed on-the-fly into a knowledge graph, and the complex query is split into sub-questions. For each sub-question, a BFS-based traversal dynamically expands along relevant edges, assembling explicit evidence chains without overwhelming the language model with superfluous context. Experiments on MuSiQue, 2WikiMultiHopQA, and HotpotQA show that StepChain GraphRAG achieves state-of-the-art Exact Match and F1 scores. StepChain GraphRAG lifts average EM by 2.57% and F1 by 2.13% over the SOTA method, achieving the largest gain on HotpotQA (+4.70% EM, +3.44% F1). StepChain GraphRAG also fosters enhanced explainability by preserving the chain-of-thought across intermediate retrieval steps. We conclude by discussing how future work can mitigate the computational overhead and address potential hallucinations from large language models to refine efficiency and reliability in multi-hop QA.
Authors: Shinya Uryu
Abstract: Large Language Models (LLMs) are rapidly being adopted in conservation to address the biodiversity crisis, yet their reliability for species evaluation is uncertain. This study systematically validates five leading models on 21,955 species across four core IUCN Red List assessment components: taxonomy, conservation status, distribution, and threats. A critical paradox was revealed: models excelled at taxonomic classification (94.9%) but consistently failed at conservation reasoning (27.2% for status assessment). This knowledge-reasoning gap, evident across all models, suggests inherent architectural constraints, not just data limitations. Furthermore, models exhibited systematic biases favoring charismatic vertebrates, potentially amplifying existing conservation inequities. These findings delineate clear boundaries for responsible LLM deployment: they are powerful tools for information retrieval but require human oversight for judgment-based decisions. A hybrid approach is recommended, where LLMs augment expert capacity while human experts retain sole authority over risk assessment and policy.
Authors: Jahidul Arafat, Fariha Tasmin, Sanjaya Poudel, Kamrujjaman, Eftakhar Ahmed Arnob, Ahsan Habib Tareq
Abstract: Wordle presents an algorithmically rich testbed for constraint satisfaction problem (CSP) solving. While existing solvers rely on information-theoretic entropy maximization or frequency-based heuristics without formal constraint treatment, we present the first comprehensive CSP formulation of Wordle with novel constraint-aware solving strategies. We introduce CSP-Aware Entropy, computing information gain after constraint propagation rather than on raw candidate sets, and a Probabilistic CSP framework integrating Bayesian word-frequency priors with logical constraints. Through evaluation on 2,315 English words, CSP-Aware Entropy achieves 3.54 average guesses with 99.9% success rate, a statistically significant 1.7% improvement over Forward Checking (t=-4.82, p<0.001, Cohen's d=0.07) with 46% faster runtime (12.9ms versus 23.7ms per guess). Under 10% noise, CSP-aware approaches maintain 5.3 percentage point advantages (29.0% versus 23.7%, p=0.041), while Probabilistic CSP achieves 100% success across all noise levels (0-20%) through constraint recovery mechanisms. Cross-lexicon validation on 500 Spanish words demonstrates 88% success with zero language-specific tuning, validating that core CSP principles transfer across languages despite an 11.2 percentage point gap from linguistic differences (p<0.001, Fisher's exact test). Our open-source implementation with 34 unit tests achieving 91% code coverage provides reproducible infrastructure for CSP research. The combination of formal CSP treatment, constraint-aware heuristics, probabilistic-logical integration, robustness analysis, and cross-lexicon validation establishes new performance benchmarks demonstrating that principled constraint satisfaction techniques outperform classical information-theoretic and learning-based approaches for structured puzzle-solving domains.
Authors: Jian Mu, Qixin Zhang, Zhiyong Wang, Menglin Yang, Shuang Qiu, Chengwei Qin, Zhongxiang Dai, Yao Shu
Abstract: Large language models (LLMs) increasingly solve complex reasoning tasks via long chain-of-thought, but their forward-only autoregressive generation process is fragile; early token errors can cascade, which creates a clear need for self-reflection mechanisms. However, existing self-reflection either performs revisions over full drafts or learns self-correction via expensive training, both fundamentally reactive and inefficient. To address this, we propose Self-Reflective Generation at Test Time (SRGen), a lightweight test-time framework that reflects before generating at uncertain points. During token generation, SRGen utilizes dynamic entropy thresholding to identify high-uncertainty tokens. For each identified token, it trains a specific corrective vector, which fully exploits the already generated context for a self-reflective generation to correct the token probability distribution. By retrospectively analyzing the partial output, this self-reflection enables more trustworthy decisions, thereby significantly reducing the probability of errors at highly uncertain points. Evaluated on challenging mathematical reasoning benchmarks and a diverse set of LLMs, SRGen can consistently strengthen model reasoning: improvements in single-pass quality also translate into stronger self-consistency voting. Especially, on AIME2024 with DeepSeek-R1-Distill-Qwen-7B, SRGen yields absolute improvements of +12.0% on Pass@1 and +13.3% on Cons@5. Moreover, our findings position SRGen as a plug-and-play method that integrates reflection into the generation process for reliable LLM reasoning, achieving consistent gains with bounded overhead and broad composability with other training-time (e.g., RLHF) and test-time (e.g., SLOT) techniques.
Authors: Yohan Lee, Yongwoo Song, Sangyeop Kim
Abstract: We present the Conversational Data Retrieval (CDR) benchmark, the first comprehensive test set for evaluating systems that retrieve conversation data for product insights. With 1.6k queries across five analytical tasks and 9.1k conversations, our benchmark provides a reliable standard for measuring conversational data retrieval performance. Our evaluation of 16 popular embedding models shows that even the best models reach only around NDCG@10 of 0.51, revealing a substantial gap between document and conversational data retrieval capabilities. Our work identifies unique challenges in conversational data retrieval (implicit state recognition, turn dynamics, contextual references) while providing practical query templates and detailed error analysis across different task categories. The benchmark dataset and code are available at https://github.com/l-yohai/CDR-Benchmark.
Authors: Jingqi Zhang, Ruibo Chen, Yingqing Yang, Peihua Mai, Heng Huang, Yan Pang
Abstract: Large Language Models (LLMs) are increasingly fine-tuned on smaller, domain-specific datasets to improve downstream performance. These datasets often contain proprietary or copyrighted material, raising the need for reliable safeguards against unauthorized use. Existing membership inference attacks (MIAs) and dataset-inference methods typically require access to internal signals such as logits, while current black-box approaches often rely on handcrafted prompts or a clean reference dataset for calibration, both of which limit practical applicability. Watermarking is a promising alternative, but prior techniques can degrade text quality or reduce task performance. We propose TRACE, a practical framework for fully black-box detection of copyrighted dataset usage in LLM fine-tuning. \texttt{TRACE} rewrites datasets with distortion-free watermarks guided by a private key, ensuring both text quality and downstream utility. At detection time, we exploit the radioactivity effect of fine-tuning on watermarked data and introduce an entropy-gated procedure that selectively scores high-uncertainty tokens, substantially amplifying detection power. Across diverse datasets and model families, TRACE consistently achieves significant detections (p<0.05), often with extremely strong statistical evidence. Furthermore, it supports multi-dataset attribution and remains robust even after continued pretraining on large non-watermarked corpora. These results establish TRACE as a practical route to reliable black-box verification of copyrighted dataset usage. We will make our code available at: https://github.com/NusIoraPrivacy/TRACE.
Authors: Matthew Lewis, Samuel Thio, Richard JB Dobson, Spiros Denaxas
Abstract: This paper presents the development and evaluation of a Retrieval-Augmented Generation (RAG) system for querying the United Kingdom's National Institute for Health and Care Excellence (NICE) clinical guidelines using Large Language Models (LLMs). The extensive length and volume of these guidelines can impede their utilisation within a time-constrained healthcare system, a challenge this project addresses through the creation of a system capable of providing users with precisely matched information in response to natural language queries. The system's retrieval architecture, composed of a hybrid embedding mechanism, was evaluated against a database of 10,195 text chunks derived from three hundred guidelines. It demonstrates high performance, with a Mean Reciprocal Rank (MRR) of 0.814, a Recall of 81% at the first chunk and of 99.1% within the top ten retrieved chunks, when evaluated on 7901 queries. The most significant impact of the RAG system was observed during the generation phase. When evaluated on a manually curated dataset of seventy question-answer pairs, RAG-enhanced models showed substantial gains in performance. Faithfulness, the measure of whether an answer is supported by the source text, was increased by 64.7 percentage points to 99.5% for the RAG-enhanced O4-Mini model and significantly outperformed the medical-focused Meditron3-8B LLM, which scored 43%. This, combined with a perfect Context Precision score of 1 for all RAG-enhanced models, confirms the system's ability to prevent information fabrication by grounding its answers in relevant source material. This study thus establishes RAG as an effective, reliable, and scalable approach for applying generative AI in healthcare, enabling cost-effective access to medical guidelines.
Authors: Rongchen Guo, Vincent Francoeur, Isar Nejadgholi, Sylvain Gagnon, Miodrag Bolic
Abstract: Speech Emotion Recognition (SER) is essential for improving human-computer interaction, yet its accuracy remains constrained by the complexity of emotional nuances in speech. In this study, we distinguish between descriptive semantics, which represents the contextual content of speech, and expressive semantics, which reflects the speaker's emotional state. After watching emotionally charged movie segments, we recorded audio clips of participants describing their experiences, along with the intended emotion tags for each clip, participants' self-rated emotional responses, and their valence/arousal scores. Through experiments, we show that descriptive semantics align with intended emotions, while expressive semantics correlate with evoked emotions. Our findings inform SER applications in human-AI interaction and pave the way for more context-aware AI systems.
Authors: Oriol Pareras, Gerard I. G\'allego, Federico Costa, Cristina Espa\~na-Bonet, Javier Hernando
Abstract: Recent work on Speech-to-Text Translation (S2TT) has focused on LLM-based models, introducing the increasingly adopted Chain-of-Thought (CoT) prompting, where the model is guided to first transcribe the speech and then translate it. CoT typically outperforms direct prompting primarily because it can exploit abundant Automatic Speech Recognition (ASR) and Text-to-Text Translation (T2TT) datasets to explicitly model its steps. In this paper, we systematically compare CoT and Direct prompting under increasing amounts of S2TT data. To this end, we pseudo-label an ASR corpus by translating its transcriptions into six European languages, and train LLM-based S2TT systems with both prompting strategies at different data scales. Our results show that Direct improves more consistently as the amount of data increases, suggesting that it may become a more effective approach as larger S2TT resources are created.
Authors: Beth Pearson, Ahmed Adnan, Zahraa Abdallah
Abstract: Radiology report evaluation is a crucial part of radiologists' training and plays a key role in ensuring diagnostic accuracy. As part of the standard reporting workflow, a junior radiologist typically prepares a preliminary report, which is then reviewed and edited by a senior radiologist to produce the final report. Identifying semantic differences between preliminary and final reports is essential for junior doctors, both as a training tool and to help uncover gaps in clinical knowledge. While AI in radiology is a rapidly growing field, the application of large language models (LLMs) remains challenging due to the need for specialised domain knowledge. In this paper, we explore the ability of LLMs to provide explainable and accurate comparisons of reports in the radiology domain. We begin by comparing the performance of several LLMs in comparing radiology reports. We then assess a more traditional approach based on Named-Entity-Recognition (NER). However, both approaches exhibit limitations in delivering accurate feedback on semantic similarity. To address this, we propose Llama-EntScore, a semantic similarity scoring method using a combination of Llama 3.1 and NER with tunable weights to emphasise or de-emphasise specific types of differences. Our approach generates a quantitative similarity score for tracking progress and also gives an interpretation of the score that aims to offer valuable guidance in reviewing and refining their reporting. We find our method achieves 67% exact-match accuracy and 93% accuracy within +/- 1 when compared to radiologist-provided ground truth scores - outperforming both LLMs and NER used independently. Code is available at: \href{https://github.com/otmive/llama_reports}{github.com/otmive/llama\_reports}
Authors: Jacobo Romero-D\'iaz, Gerard I. G\'allego, Oriol Pareras, Federico Costa, Javier Hernando, Cristina Espa\~na-Bonet
Abstract: Speech-to-Text Translation (S2TT) systems built from Automatic Speech Recognition (ASR) and Text-to-Text Translation (T2TT) modules face two major limitations: error propagation and the inability to exploit prosodic or other acoustic cues. Chain-of-Thought (CoT) prompting has recently been introduced, with the expectation that jointly accessing speech and transcription will overcome these issues. Analyzing CoT through attribution methods, robustness evaluations with corrupted transcripts, and prosody-awareness, we find that it largely mirrors cascaded behavior, relying mainly on transcripts while barely leveraging speech. Simple training interventions, such as adding Direct S2TT data or noisy transcript injection, enhance robustness and increase speech attribution. These findings challenge the assumed advantages of CoT and highlight the need for architectures that explicitly integrate acoustic information into translation.
Authors: Zhaojun Sun, Xuzhou Zhu, Xuanhe Zhou, Xin Tong, Shuo Wang, Jie Fu, Guoliang Li, Zhiyuan Liu, Fan Wu
Abstract: Academic survey writing, which distills vast literature into a coherent and insightful narrative, remains a labor-intensive and intellectually demanding task. While recent approaches, such as general DeepResearch agents and survey-specialized methods, can generate surveys automatically (a.k.a. LLM4Survey), their outputs often fall short of human standards and there lacks a rigorous, reader-aligned benchmark for thoroughly revealing their deficiencies. To fill the gap, we propose a fine-grained, quiz-driven evaluation framework SurveyBench, featuring (1) typical survey topics source from recent 11,343 arXiv papers and corresponding 4,947 high-quality surveys; (2) a multifaceted metric hierarchy that assesses the outline quality (e.g., coverage breadth, logical coherence), content quality (e.g., synthesis granularity, clarity of insights), and non-textual richness; and (3) a dual-mode evaluation protocol that includes content-based and quiz-based answerability tests, explicitly aligned with readers' informational needs. Results show SurveyBench effectively challenges existing LLM4Survey approaches (e.g., on average 21% lower than human in content-based evaluation).
Authors: Ej Zhou, Caiqi Zhang, Tiancheng Hu, Chengzu Li, Nigel Collier, Ivan Vuli\'c, Anna Korhonen
Abstract: Confidence calibration, the alignment of a model's predicted confidence with its actual accuracy, is crucial for the reliable deployment of Large Language Models (LLMs). However, this critical property remains largely under-explored in multilingual contexts. In this work, we conduct the first large-scale, systematic studies of multilingual calibration across six model families and over 100 languages, revealing that non-English languages suffer from systematically worse calibration. To diagnose this, we investigate the model's internal representations and find that the final layer, biased by English-centric training, provides a poor signal for multilingual confidence. In contrast, our layer-wise analysis uncovers a key insight that late-intermediate layers consistently offer a more reliable and better-calibrated signal. Building on this, we introduce a suite of training-free methods, including Language-Aware Confidence Ensemble (LACE), which adaptively selects an optimal ensemble of layers for each specific language. Our study highlights the hidden costs of English-centric alignment and offer a new path toward building more globally equitable and trustworthy LLMs by looking beyond the final layer.
Authors: Katherine Thai, Bradley Emi, Elyas Masrour, Mohit Iyyer
Abstract: A significant proportion of queries to large language models ask them to edit user-provided text, rather than generate new text from scratch. While previous work focuses on detecting fully AI-generated text, we demonstrate that AI-edited text is distinguishable from human-written and AI-generated text. First, we propose using lightweight similarity metrics to quantify the magnitude of AI editing present in a text given the original human-written text and validate these metrics with human annotators. Using these similarity metrics as intermediate supervision, we then train EditLens, a regression model that predicts the amount of AI editing present within a text. Our model achieves state-of-the-art performance on both binary (F1=94.7%) and ternary (F1=90.4%) classification tasks in distinguishing human, AI, and mixed writing. Not only do we show that AI-edited text can be detected, but also that the degree of change made by AI to human writing can be detected, which has implications for authorship attribution, education, and policy. Finally, as a case study, we use our model to analyze the effects of AI-edits applied by Grammarly, a popular writing assistance tool. To encourage further research, we commit to publicly releasing our models and dataset.
Authors: I\~nigo Parra
Abstract: Previous work has shown correlations between the hidden states of large language models and fMRI brain responses, on language tasks. These correlations have been taken as evidence of the representational similarity of these models and brain states. This study tests whether these previous results are robust to several possible concerns. Specifically this study shows: (i) that the previous results are still found after dimensionality reduction, and thus are not attributable to the curse of dimensionality; (ii) that previous results are confirmed when using new measures of similarity; (iii) that correlations between brain representations and those from models are specific to models trained on human language; and (iv) that the results are dependent on the presence of positional encoding in the models. These results confirm and strengthen the results of previous research and contribute to the debate on the biological plausibility and interpretability of state-of-the-art large language models.
Authors: Xuan Xu, Haolun Li, Zhongliang Yang, Beilin Chu, Jia Song, Moxuan Xu, Linna Zhou
Abstract: Traditional topic models such as neural topic models rely on inference and generation networks to learn latent topic distributions. This paper explores a new paradigm for topic modeling in the era of large language models, framing TM as a long-form generation task whose definition is updated in this paradigm. We propose a simple but practical approach to implement LLM-based topic model tasks out of the box (sample a data subset, generate topics and representative text with our prompt, text assignment with keyword match). We then investigate whether the long-form generation paradigm can beat NTMs via zero-shot prompting. We conduct a systematic comparison between NTMs and LLMs in terms of topic quality and empirically examine the claim that "a majority of NTMs are outdated."
Authors: Abteen Ebrahimi, Adam Wiemerslage, Katharina von der Wense
Abstract: We present NN-Rank, an algorithm for ranking source languages for cross-lingual transfer, which leverages hidden representations from multilingual models and unlabeled target-language data. We experiment with two pretrained multilingual models and two tasks: part-of-speech tagging (POS) and named entity recognition (NER). We consider 51 source languages and evaluate on 56 and 72 target languages for POS and NER, respectively. When using in-domain data, NN-Rank beats state-of-the-art baselines that leverage lexical and linguistic features, with average improvements of up to 35.56 NDCG for POS and 18.14 NDCG for NER. As prior approaches can fall back to language-level features if target language data is not available, we show that NN-Rank remains competitive using only the Bible, an out-of-domain corpus available for a large number of languages. Ablations on the amount of unlabeled target data show that, for subsets consisting of as few as 25 examples, NN-Rank produces high-quality rankings which achieve 92.8% of the NDCG achieved using all available target data for ranking.
Authors: Imene Kerboua, Sahar Omidi Shayegan, Megh Thakkar, Xing Han L\`u, L\'eo Boisvert, Massimo Caccia, J\'er\'emy Espinas, Alexandre Aussem, V\'eronique Eglin, Alexandre Lacoste
Abstract: Web agents powered by large language models (LLMs) must process lengthy web page observations to complete user goals; these pages often exceed tens of thousands of tokens. This saturates context limits and increases computational cost processing; moreover, processing full pages exposes agents to security risks such as prompt injection. Existing pruning strategies either discard relevant content or retain irrelevant context, leading to suboptimal action prediction. We introduce FocusAgent, a simple yet effective approach that leverages a lightweight LLM retriever to extract the most relevant lines from accessibility tree (AxTree) observations, guided by task goals. By pruning noisy and irrelevant content, FocusAgent enables efficient reasoning while reducing vulnerability to injection attacks. Experiments on WorkArena and WebArena benchmarks show that FocusAgent matches the performance of strong baselines, while reducing observation size by over 50%. Furthermore, a variant of FocusAgent significantly reduces the success rate of prompt-injection attacks, including banner and pop-up attacks, while maintaining task success performance in attack-free settings. Our results highlight that targeted LLM-based retrieval is a practical and robust strategy for building web agents that are efficient, effective, and secure.
Authors: Tianyu Fu, Zihan Min, Hanling Zhang, Jichao Yan, Guohao Dai, Wanli Ouyang, Yu Wang
Abstract: Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains unattainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 8.5-10.5% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.0-5.0%, while delivering an average 2.0x speedup in latency. Our code is available at https://github.com/thu-nics/C2C.
Authors: Hongxiang Zhang, Yuan Tian, Tianyi Zhang
Abstract: To solve complex reasoning tasks for Large Language Models (LLMs), prompting-based methods offer a lightweight alternative to fine-tuning and reinforcement learning. However, as reasoning chains extend, critical intermediate steps and the original prompt will be buried in the context, receiving insufficient attention and leading to errors. In this paper, we propose Self-Anchor, a novel pipeline that leverages the inherent structure of reasoning to steer LLM attention. Self-Anchor decomposes reasoning trajectories into structured plans and automatically aligns the model's attention to the most relevant inference steps, allowing the model to maintain focus throughout generation. Our experiment shows that Self-Anchor outperforms SOTA prompting methods across six benchmarks. Notably, Self-Anchor significantly reduces the performance gap between ``non-reasoning'' models and specialized reasoning models, with the potential to enable most LLMs to tackle complex reasoning tasks without retraining.
Authors: Sebastian Gehrmann
Abstract: The emergence of reinforcement learning in post-training of large language models has sparked significant interest in reward models. Reward models assess the quality of sampled model outputs to generate training signals. This task is also performed by evaluation metrics that monitor the performance of an AI model. We find that the two research areas are mostly separate, leading to redundant terminology and repeated pitfalls. Common challenges include susceptibility to spurious correlations, impact on downstream reward hacking, methods to improve data quality, and approaches to meta-evaluation. Our position paper argues that a closer collaboration between the fields can help overcome these issues. To that end, we show how metrics outperform reward models on specific tasks and provide an extensive survey of the two areas. Grounded in this survey, we point to multiple research topics in which closer alignment can improve reward models and metrics in areas such as preference elicitation methods, avoidance of spurious correlations and reward hacking, and calibration-aware meta-evaluation.
Authors: Lekkala Sai Teja, Annepaka Yadagiri, Sangam Sai Anish, Siva Gopala Krishna Nuthakki, Partha Pakray
Abstract: The growth of highly advanced Large Language Models (LLMs) constitutes a huge dual-use problem, making it necessary to create dependable AI-generated text detection systems. Modern detectors are notoriously vulnerable to adversarial attacks, with paraphrasing standing out as an effective evasion technique that foils statistical detection. This paper presents a comparative study of adversarial robustness, first by quantifying the limitations of standard adversarial training and then by introducing a novel, significantly more resilient detection framework: Perturbation-Invariant Feature Engineering (PIFE), a framework that enhances detection by first transforming input text into a standardized form using a multi-stage normalization pipeline, it then quantifies the transformation's magnitude using metrics like Levenshtein distance and semantic similarity, feeding these signals directly to the classifier. We evaluate both a conventionally hardened Transformer and our PIFE-augmented model against a hierarchical taxonomy of character-, word-, and sentence-level attacks. Our findings first confirm that conventional adversarial training, while resilient to syntactic noise, fails against semantic attacks, an effect we term "semantic evasion threshold", where its True Positive Rate at a strict 1% False Positive Rate plummets to 48.8%. In stark contrast, our PIFE model, which explicitly engineers features from the discrepancy between a text and its canonical form, overcomes this limitation. It maintains a remarkable 82.6% TPR under the same conditions, effectively neutralizing the most sophisticated semantic attacks. This superior performance demonstrates that explicitly modeling perturbation artifacts, rather than merely training on them, is a more promising path toward achieving genuine robustness in the adversarial arms race.
Authors: Yongqi Kang, Yong Zhao
Abstract: The advancement of computational psychology requires AI tools capable of deeply understanding counseling dialogues. Existing audio language models (AudioLLMs) often rely on single speech encoders pre-trained on general data, struggling to capture domain-specific features like complex emotions and professional techniques. To address this, we propose WEE-Therapy, a multi-task AudioLLM incorporating a Weak Encoder Ensemble (WEE) mechanism. This supplements a powerful base encoder with a pool of lightweight, specialized encoders. A novel dual-routing strategy combines stable, data-independent domain knowledge with dynamic, data-dependent expert selection. Evaluated on emotion recognition, technique classification, risk detection, and summarization, WEE-Therapy achieves significant performance gains across all tasks with minimal parameter overhead, demonstrating strong potential for AI-assisted clinical analysis.
Authors: Lukas Buess, Jan Geier, David Bani-Harouni, Chantal Pellegrini, Matthias Keicher, Paula Andrea Perez-Toro, Nassir Navab, Andreas Maier, Tomas Arias-Vergara
Abstract: Spoken communication plays a central role in clinical workflows. In radiology, for example, most reports are created through dictation. Yet, nearly all medical AI systems rely exclusively on written text. In this work, we address this gap by exploring the feasibility of learning visual-language representations directly from spoken radiology reports. Specifically, we synthesize a large-scale dataset (Speech-RATE) of spoken radiology reports and train SpeechCT-CLIP, a contrastive model that aligns speech and 3D CT volumes in a shared representation space. While naive speech-based models underperform compared to text-trained counterparts, we show that knowledge distillation from a pretrained text-image CLIP model effectively transfers semantic alignment capabilities from text to speech, substantially narrowing this gap. Experiments demonstrate improved zero-shot classification F1 from 0.623 to 0.705, recovering 88% of the performance difference, and strong retrieval results without requiring text at inference. These findings highlight speech as a practical alternative to text in multimodal pretraining and open the door to voice-driven diagnostic support tools in clinical practice.
Authors: Yu Zhang, Shuliang Liu, Xu Yang, Xuming Hu
Abstract: Watermarking algorithms for Large Language Models (LLMs) effectively identify machine-generated content by embedding and detecting hidden statistical features in text. However, such embedding leads to a decline in text quality, especially in low-entropy scenarios where performance needs improvement. Existing methods that rely on entropy thresholds often require significant computational resources for tuning and demonstrate poor adaptability to unknown or cross-task generation scenarios. We propose \textbf{C}ontext-\textbf{A}ware \textbf{T}hreshold watermarking ($\myalgo$), a novel framework that dynamically adjusts watermarking intensity based on real-time semantic context. $\myalgo$ partitions text generation into semantic states using logits clustering, establishing context-aware entropy thresholds that preserve fidelity in structured content while embedding robust watermarks. Crucially, it requires no pre-defined thresholds or task-specific tuning. Experiments show $\myalgo$ improves text quality in cross-tasks without sacrificing detection accuracy.
Authors: Parth Asawa, Alan Zhu, Matei Zaharia, Alexandros G. Dimakis, Joseph E. Gonzalez
Abstract: Foundation models are increasingly deployed as black-box services, where model weights cannot be modified and customization is limited to prompting. While static prompt optimization has shown promise, it produces a single fixed prompt that fails to adapt to different inputs, users, or environments. We introduce Advisor Models, lightweight parametric policies trained with reinforcement learning to reactively issue natural language steering instructions in-context to black-box models. The advisor is a second small model that sits between the input and the model, shaping behavior on a per-instance basis using reward signals from the environment. Across multiple domains involving reasoning and personalization, we show that Advisor Models outperform static prompt optimizers, discovering environment dynamics and improving downstream task performance. We also demonstrate the generalizability of advisors by transferring them across black-box models, as well as the framework's ability to achieve specialization while retaining robustness to out-of-distribution inputs. Viewed more broadly, Advisor Models provide a learnable interface to black-box systems where the advisor acts as a parametric, environment-specific memory. We argue that dynamic optimization of black-box models via Advisor Models is a promising direction for enabling personalization and environment-adaptable AI with frontier-level capabilities.
Authors: Sung-Yeon Park, Adam Lee, Juanwu Lu, Can Cui, Luyang Jiang, Rohit Gupta, Kyungtae Han, Ahmadreza Moradipari, Ziran Wang
Abstract: Driving scene manipulation with sensor data is emerging as a promising alternative to traditional virtual driving simulators. However, existing frameworks struggle to generate realistic scenarios efficiently due to limited editing capabilities. To address these challenges, we present SIMSplat, a predictive driving scene editor with language-aligned Gaussian splatting. As a language-controlled editor, SIMSplat enables intuitive manipulation using natural language prompts. By aligning language with Gaussian-reconstructed scenes, it further supports direct querying of road objects, allowing precise and flexible editing. Our method provides detailed object-level editing, including adding new objects and modifying the trajectories of both vehicles and pedestrians, while also incorporating predictive path refinement through multi-agent motion prediction to generate realistic interactions among all agents in the scene. Experiments on the Waymo dataset demonstrate SIMSplat's extensive editing capabilities and adaptability across a wide range of scenarios. Project page: https://sungyeonparkk.github.io/simsplat/
Authors: Nii Osae Osae Dade, Moinul Hossain Rahat
Abstract: Training Large Language Models (LLMs) is plagued by long training times and massive energy consumption, with modern models requiring months of computation and gigawatt-hours of electricity. In light of these challenges,we introduce Litespark, a novel pre-training framework that addresses these inefficiencies through targeted optimizations to transformer attention and MLP layers. Our approach combines architectural improvements with algorithmic enhancements to maximize Model FLOPs Utilization (MFU) while maintaining compatibility with standard transformer implementations. Comprehensive benchmarking on 3B and 30B parameter Llama models using the SlimPajama-627B dataset demonstrates substantial performance gains: 2x-6x training throughput improvement and $55\%-83$% energy consumption reduction across multi-node H200 GPU clusters. These optimizations are model- and hardware-agnostic, enabling broad applicability across transformer architectures and extending to post-training phases including supervised fine-tuning and direct preference optimization.
Authors: Jiangnan Li, Thuy-Trang Vu, Ehsan Abbasnejad, Gholamreza Haffari
Abstract: Conventionally, supervised fine-tuning (SFT) is treated as a simple imitation learning process that only trains a policy to imitate expert behavior on demonstration datasets. In this work, we challenge this view by establishing a fundamental equivalence between SFT and Inverse Reinforcement Learning. We prove that the SFT objective is a special case of Inverse Q-Learning, which implies that the SFT process does not just learn a policy, but also an implicit, dense, token-level reward model that explains the expert demonstrations. We then show how to recover this dense reward signal directly from the SFT model by formulating a baseline-relative reward function. The availability of such a dense reward model offers numerous benefits, providing granular credit assignment for each token generated. We demonstrate one key application by using these recovered rewards to further improve the policy with reinforcement learning. Our method, Dense-Path REINFORCE, consistently outperforms the original SFT models on instruction-following benchmarks. This work reframes SFT not merely as policy imitation but as a powerful reward learning mechanism, opening new possibilities for leveraging expert demonstrations.
Authors: Zhiting Mei, Ola Shorinwa, Anirudha Majumdar
Abstract: Generative video models demonstrate impressive text-to-video capabilities, spurring widespread adoption in many real-world applications. However, like large language models (LLMs), video generation models tend to hallucinate, producing plausible videos even when they are factually wrong. Although uncertainty quantification (UQ) of LLMs has been extensively studied in prior work, no UQ method for video models exists, raising critical safety concerns. To our knowledge, this paper represents the first work towards quantifying the uncertainty of video models. We present a framework for uncertainty quantification of generative video models, consisting of: (i) a metric for evaluating the calibration of video models based on robust rank correlation estimation with no stringent modeling assumptions; (ii) a black-box UQ method for video models (termed S-QUBED), which leverages latent modeling to rigorously decompose predictive uncertainty into its aleatoric and epistemic components; and (iii) a UQ dataset to facilitate benchmarking calibration in video models. By conditioning the generation task in the latent space, we disentangle uncertainty arising due to vague task specifications from that arising from lack of knowledge. Through extensive experiments on benchmark video datasets, we demonstrate that S-QUBED computes calibrated total uncertainty estimates that are negatively correlated with the task accuracy and effectively computes the aleatoric and epistemic constituents.
Authors: Yuheng Wu, Azalia Mirhoseini, Thierry Tambe
Abstract: Large language models (LLMs) can improve reasoning at inference time through test-time scaling (TTS), where multiple reasoning traces are generated and the best one is selected. Prior work shows that increasing the number of samples K steadily improves accuracy. In this paper, we demonstrate that this trend does not hold indefinitely: at large K, further scaling yields no gains, and certain hard questions remain unsolved regardless of the number of traces. Interestingly, we find that different sampling temperatures solve different subsets of problems, implying that single-temperature scaling explores only part of a model's potential. We therefore propose scaling along the temperature dimension, which enlarges the reasoning boundary of LLMs. Averaged over Qwen3 (0.6B, 1.7B, 4B, 8B) and five representative reasoning benchmarks (AIME 2024/2025, MATH500, LiveCodeBench, Hi-ToM), temperature scaling yields an additional 7.3 points over single-temperature TTS. Temperature scaling also enables base models to reach performance comparable to reinforcement learning (RL)-trained counterparts, without additional post-training. We further provide a comprehensive analysis of this phenomenon and design a multi-temperature voting method that reduces the overhead of temperature scaling. Overall, our findings suggest that TTS is more powerful than previously thought, and that temperature scaling offers a simple and effective way to unlock the latent potential of base models.
Authors: Hao Zhang, Zhenjia Li, Runfeng Bao, Yifan Gao, Xi Xiao, Bo Huang, Yuhang Wu, Tianyang Wang, Hao Xu
Abstract: Parameter-Efficient Fine-Tuning (PEFT), especially Low-Rank Adaptation (LoRA), has emerged as a promising approach to fine-tuning large language models(LLMs) while reducing computational and memory overhead. However, LoRA assumes a uniform rank \textit{r} for each incremental matrix, not accounting for the varying significance of weight matrices across different modules and layers. AdaLoRA leverages Singular Value Decomposition (SVD) to parameterize updates and employs pruning of singular values to introduce dynamic rank allocation, thereby enhancing adaptability. However, during the training process, it often encounters issues of slow convergence speed and high computational overhead. To address this issue, we propose HyperAdaLoRA, a novel framework that accelerates the convergence of AdaLoRA by leveraging a hypernetwork. Instead of directly optimizing the components of Singular Value Decomposition $(P, \Lambda, Q)$, HyperAdaLoRA employs a hypernetwork based on attention mechanisms to dynamically generate these parameters. By pruning the outputs of the hypernetwork that generates the singular values, dynamic rank allocation is achieved. Comprehensive experiments on various datasets and models demonstrate that our method achieves faster convergence without sacrificing performance. Additionally, further extension experiments on other LoRA-based approaches validate the broad applicability of our method.
Authors: Jingjie Ning, Yibo Kong, Yunfan Long, Jamie Callan
Abstract: Retrieval-Augmented Generation (RAG) couples document retrieval with large language models (LLMs). While scaling generators improves accuracy, it also raises cost and limits deployability. We explore an orthogonal axis: enlarging the retriever's corpus to reduce reliance on large LLMs. Experimental results show that corpus scaling consistently strengthens RAG and can often serve as a substitute for increasing model size, though with diminishing returns at larger scales. Small- and mid-sized generators paired with larger corpora often rival much larger models with smaller corpora; mid-sized models tend to gain the most, while tiny and large models benefit less. Our analysis shows that improvements arise primarily from increased coverage of answer-bearing passages, while utilization efficiency remains largely unchanged. These findings establish a principled corpus-generator trade-off: investing in larger corpora offers an effective path to stronger RAG, often comparable to enlarging the LLM itself.
Authors: Nicholas Lourie, He He, Kyunghyun Cho
Abstract: Hyperparameters greatly impact models' capabilities; however, modern models are too large for extensive search. Instead, researchers design recipes that train well across scales based on their understanding of the hyperparameters. Despite this importance, few tools exist for understanding the hyperparameter loss surface. We discover novel structure in it and propose a new theory yielding such tools. The loss surface is complex, but as you approach the optimum simple structure emerges. It becomes characterized by a few basic features, like its effective dimension and the best possible loss. To uncover this asymptotic regime, we develop a novel technique based on random search. Within this regime, the best scores from random search take on a new distribution we discover. Its parameters are exactly the features defining the loss surface in the asymptotic regime. From these features, we derive a new asymptotic law for random search that can explain and extrapolate its convergence. These new tools enable new analyses, such as confidence intervals for the best possible performance or determining the effective number of hyperparameters. We make these tools available at https://github.com/nicholaslourie/opda .
Authors: Shashank Agnihotri, Jonas Jakubassa, Priyam Dey, Sachin Goyal, Bernt Schiele, Venkatesh Babu Radhakrishnan, Margret Keuper
Abstract: Open-weight LLMs can be modified at inference time with simple activation edits, which raises a practical question for safety: do common safety interventions like refusal training or metatag training survive such edits? We study model abliteration, a lightweight projection technique designed to remove refusal-sensitive directions, and conduct a controlled evaluation across a granular sequence of Safety Pretraining checkpoints for SmolLM2-1.7B, alongside widely used open baselines. For each of 20 systems, original and abliterated, we issue 100 prompts with balanced harmful and harmless cases, classify responses as **Refusal** or **Non-Refusal** using multiple judges, and validate judge fidelity on a small human-labeled subset. We also probe whether models can identify refusal in their own outputs. Our study produces a checkpoint-level characterization of which data-centric safety components remain robust under abliteration, quantifies how judge selection influences evaluation outcomes, and outlines a practical protocol for integrating inference-time edits into safety assessments. Code: https://github.com/shashankskagnihotri/safety_pretraining.
URLs: https://github.com/shashankskagnihotri/safety_pretraining.
Authors: Jingyuan Deng, Yujiu Yang
Abstract: Large vision-language models (LVLMs) have shown remarkable performance in visual-language understanding for downstream multimodal tasks. While their capabilities are improving, problems emerge simultaneously. Among those problems, the hallucinations have attracted much attention, which stands for the phenomenon where LVLMs generate contradictory content to their input visual and text contents. Many approaches have been proposed to deal with this issue, such as contrastive decoding and attention manipulation. However, contrastive decoding methods struggle in constructing appropriate contrastive samples, and attention manipulation methods are highly sensitive, lacking stability. In this work, we propose image head Masked Contrastive Decoding (MaskCD). Our approach utilizes the "image heads" in LVLMs, masking them to construct contrastive samples for contrastive decoding. We evaluated MaskCD on LLaVA-1.5-7b and Qwen-VL-7b, using various benchmarks such as CHAIR, POPE, AMBER and MME. The results demonstrate that MaskCD effectively alleviates the phenomenon of hallucinations and retains the general capabilities of LVLMs. Corresponding resources could be found at: https://github.com/Deng-Jingyuan/MaskCD .
Authors: Moses Charikar, Chirag Pabbaraju
Abstract: Kleinberg and Mullainathan (2024) recently proposed an interesting model for language generation in the limit: Given a countable collection of languages, and an adversary enumerating the strings of some language $L$ from the collection, the objective is to generate new strings from the target language, such that all strings generated beyond some finite time are valid. Li, Raman and Tewari (2024) and Charikar and Pabbaraju (2024) showed strong non-uniform generation guarantees in this model, giving algorithms that generate new valid strings from $L$ after seeing a number of distinct input strings $t(L)$ that depends only on $L$ (and the collection), but not the enumeration order. However, for both these works, the language-wise generation times $t(L)$ of the algorithm can be strictly sub-optimal. In this work, we study Pareto-optimality of non-uniform language generation in the limit. We propose an algorithm, whose generation times $t^\star(L)$ are (almost) Pareto-optimal: any other algorithm whose generation time for some language $L$ is strictly smaller than $t^\star(L)$, must satisfy that its generation time for some other language $L'$ is strictly worse than $t^\star(L')$. Pareto-optimality is essentially the best that one can achieve for non-uniform generation. Our algorithmic framework conveniently adapts to further give Pareto-optimal non-uniform generation algorithms in the practically motivated settings of noisy as well as representative generation.
Authors: Yulong Zhang, Li Wang, Wei Du, Peilin Li, Yuqin Dai Zhiyuan Zhao, Lingyong Fang, Ziniu Liu, Ru Zhang, Huijia Zhu, Gongshen Liu
Abstract: Verifying multi-step reasoning in large language models is difficult due to imprecise error localization and high token costs. Existing methods either assess entire reasoning chains, suffering attention dilution, or rely on expensive multi-sampling. We introduce Node-wise Consistency Verification (NCV), a training-free framework that recasts verification as lightweight binary consistency checks at the node level. By decomposing the chain of thought into interconnected verification nodes, NCV precisely localizes errors and avoids unnecessary long-form generation. Experiments demonstrate that our approach enhances interpretability and efficiency, presenting a scalable solution for reliable LLM reasoning verification. On public datasets, NCV achieves a 10\% to 25\% improvement in F1 scores over baselines while utilizing $6\times$~$58\times$ fewer tokens than traditional methods like CoT-based verifiers.
Authors: Wonjoong Kim, Sangwu Park, Yeonjun In, Sein Kim, Dongha Lee, Chanyoung Park
Abstract: Although recent tool-augmented benchmarks incorporate complex user requests and diverse tools, the evaluation methods for most of them remain limited to answer matching. However, as the number of steps required to resolve a user request increases, a proper evaluation of an agent's performance must go beyond the final answer to also assess the problem-solving trajectory, including previously ignored aspects such as efficiency, hallucination, and adaptivity. The most straightforward method for evaluating these aspects is to compare an agent's trajectory with the ground-truth trajectory, but this approach is fundamentally limited since annotating all valid ground-truth trajectories is prohibitively expensive. However, a simple LLM-based evaluator struggles to assess trajectories in detail without ground truth. To effectively evaluate the agents in this manner, we introduce TRACE, a framework for the multi-dimensional evaluation of tool-augmented LLM agent performance. By incorporating an evidence bank, which accumulates knowledge gathered from preceding reasoning steps, TRACE enables a multi-faceted analysis and evaluation of an agent's reasoning trajectory effectively. To validate our framework, we develop a new meta-evaluation dataset by augmenting existing benchmarks with diverse and flawed trajectories, each labeled with multi-faceted performance scores. Our results confirm that TRACE accurately evaluates these complex behaviors in a scalable and cost-effective manner, even with small open-source LLMs. Furthermore, we apply our method to evaluate the trajectories that agents produce while solving tool-augmented tasks, presenting previously unreported observations and their corresponding insights.
Authors: Cuong Chi Le, Minh V. T. Pham, Cuong Duc Van, Hoang N. Phan, Huy N. Phan, Tien N. Nguyen
Abstract: Large Language Models (LLMs) achieve strong results on code tasks, but how they derive program meaning remains unclear. We argue that code communicates through two channels: structural semantics, which define formal behavior, and human-interpretable naming, which conveys intent. Removing the naming channel severely degrades intent-level tasks such as summarization, where models regress to line-by-line descriptions. Surprisingly, we also observe consistent reductions on execution tasks that should depend only on structure, revealing that current benchmarks reward memorization of naming patterns rather than genuine semantic reasoning. To disentangle these effects, we introduce a suite of semantics-preserving obfuscations and show that they expose identifier leakage across both summarization and execution. Building on these insights, we release ClassEval-Obf, an obfuscation-enhanced benchmark that systematically suppresses naming cues while preserving behavior. Our results demonstrate that ClassEval-Obf reduces inflated performance gaps, weakens memorization shortcuts, and provides a more reliable basis for assessing LLMs' code understanding and generalization.
Authors: Yilun Hao, Yongchao Chen, Chuchu Fan, Yang Zhang
Abstract: Vision Language Models (VLMs) show strong potential for visual planning but struggle with precise spatial and long-horizon reasoning. In contrast, Planning Domain Definition Language (PDDL) planners excel at long-horizon formal planning, but cannot interpret visual inputs. Recent works combine these complementary advantages by enabling VLMs to turn visual planning problems into PDDL files for formal planning. However, while VLMs can generate PDDL problem files satisfactorily, they struggle to accurately generate the PDDL domain files, which describe all the planning rules. As a result, prior methods rely on human experts to predefine domain files or on constant environment access for refinement. We propose VLMFP, a Dual-VLM-guided framework that can autonomously generate both PDDL problem and domain files for formal visual planning. VLMFP introduces two VLMs to ensure reliable PDDL file generation: A SimVLM that simulates action consequences based on input rule descriptions, and a GenVLM that generates and iteratively refines PDDL files by comparing the PDDL and SimVLM execution results. VLMFP unleashes multiple levels of generalizability: The same generated PDDL domain file works for all the different instances under the same problem, and VLMs generalize to different problems with varied appearances and rules. We evaluate VLMFP with 6 grid-world domains and test its generalization to unseen instances, appearance, and game rules. On average, SimVLM accurately describes 95.5%, 82.6% of scenarios, simulates 85.5%, 87.8% of action sequence, and judges 82.4%, 85.6% goal reaching for seen and unseen appearances, respectively. With the guidance of SimVLM, VLMFP can generate PDDL files to reach 70.0%, 54.1% valid plans for unseen instances in seen and unseen appearances, respectively. Project page: https://sites.google.com/view/vlmfp.
Authors: Cai Zhou, Chenxiao Yang, Yi Hu, Chenyu Wang, Chubin Zhang, Muhan Zhang, Lester Mackey, Tommi Jaakkola, Stephen Bates, Dinghuai Zhang
Abstract: Diffusion language models, especially masked discrete diffusion models, have achieved great success recently. While there are some theoretical and primary empirical results showing the advantages of latent reasoning with looped transformers or continuous chain-of-thoughts, continuous diffusion models typically underperform their discrete counterparts. In this paper, we argue that diffusion language models do not necessarily need to be in the discrete space. In particular, we prove that continuous diffusion models have stronger expressivity than discrete diffusions and looped transformers. We attribute the contradiction between the theoretical expressiveness and empirical performance to their practical trainability: while continuous diffusion provides intermediate supervision that looped transformers lack, they introduce additional difficulty decoding tokens into the discrete token space from the continuous representation space. We therefore propose Coevolutionary Continuous Discrete Diffusion (CCDD), which defines a joint multimodal diffusion process on the union of a continuous representation space and a discrete token space, leveraging a single model to simultaneously denoise in the joint space. By combining two modalities, CCDD is expressive with rich semantics in the latent space, as well as good trainability and sample quality with the help of explicit discrete tokens. We also propose effective architectures and advanced training/sampling techniques for CCDD, which reveals strong empirical performance in extensive language modeling experiments on real-world tasks.
Authors: Guanhua Huang, Tingqiang Xu, Mingze Wang, Qi Yi, Xue Gong, Siheng Li, Ruibin Xiong, Kejiao Li, Yuhao Jiang, Bo Zhou
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has propelled Large Language Models in complex reasoning, yet its scalability is often hindered by a training bottleneck where performance plateaus as policy entropy collapses, signaling a loss of exploration. Previous methods typically address this by maintaining high policy entropy, yet the precise mechanisms that govern meaningful exploration have remained underexplored. Our analysis suggests that an unselective focus on entropy risks amplifying irrelevant tokens and destabilizing training. This paper investigates the exploration dynamics within RLVR and identifies a key issue: the gradual elimination of valuable low-probability exploratory tokens, which we term \textbf{\textit{reasoning sparks}}. We find that while abundant in pre-trained models, these sparks are systematically extinguished during RLVR due to over-penalization, leading to a degeneracy in exploration. To address this, we introduce Low-probability Regularization (Lp-Reg). Its core mechanism regularizes the policy towards a heuristic proxy distribution. This proxy is constructed by filtering out presumed noise tokens and re-normalizing the distribution over the remaining candidates. The result is a less-noisy proxy where the probability of \textit{reasoning sparks} is amplified, which then serves as a soft regularization target to shield these valuable tokens from elimination via KL divergence. Experiments show that Lp-Reg enables stable on-policy training for around 1,000 steps, a regime where baseline entropy-control methods collapse. This sustained exploration leads to state-of-the-art performance, achieving a $60.17\%$ average accuracy on five math benchmarks, an improvement of $2.66\%$ over prior methods. Code is available at https://github.com/CarlanLark/Lp-Reg.
Authors: Ben Peters, Andr\'e F. T. Martins
Abstract: Neural machine translation (MT) models achieve strong results across a variety of settings, but it is widely believed that they are highly sensitive to "noisy" inputs, such as spelling errors, abbreviations, and other formatting issues. In this paper, we revisit this insight in light of recent multilingual MT models and large language models (LLMs) applied to machine translation. Somewhat surprisingly, we show through controlled experiments that these models are far more robust to many kinds of noise than previous models, even when they perform similarly on clean data. This is notable because, even though LLMs have more parameters and more complex training processes than past models, none of the open ones we consider use any techniques specifically designed to encourage robustness. Next, we show that similar trends hold for social media translation experiments -- LLMs are more robust to social media text. We include an analysis of the circumstances in which source correction techniques can be used to mitigate the effects of noise. Altogether, we show that robustness to many types of noise has increased.
Authors: Francesca Lucchetti, Arjun Guha
Abstract: Large Language Models (LLMs) are widely used by software engineers for programming tasks. However, research shows that LLMs often lack a deep understanding of program semantics. Even minor changes to syntax, such as renaming variables, can significantly degrade performance across various tasks. In this work, we examine the task of type prediction: given a partially typed program, can a model predict a missing type annotations such that the resulting program is more typed? We construct a dataset of adversarial examples where models initially predict the correct types, but begin to fail after semantically irrelevant edits. This is problematic, as models should ideally generalize across different syntactic forms of semantically equivalent code. This lack of robustness suggests that models may have a shallow understanding of code semantics. Despite this, we provide evidence that LLMs do, in fact, learn robust mechanisms for type prediction-though these mechanisms often fail to activate in adversarial scenarios. By using activation steering, a method that manipulates a model's internal activations to guide it toward using latent knowledge, we restore accurate predictions on adversarial inputs. We show that steering successfully activates a type prediction mechanism that is shared by both Python and TypeScript, and is more effective than prompting with in-context examples. Across five different models, our comprehensive evaluation demonstrates that LLMs can learn generalizable representations of code semantics that transfer across programming languages.
Authors: Xiang Liu, Zhenheng Tang, Peijie Dong, Zeyu Li, Yue Liu, Bo Li, Xuming Hu, Xiaowen Chu
Abstract: Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. The code is available at \href{https://github.com/NVIDIA/kvpress}{link}.
Authors: Hanxing Ding, Shuchang Tao, Liang Pang, Zihao Wei, Liwei Chen, Kun Xu, Huawei Shen, Xueqi Cheng
Abstract: Retrieval-augmented generation (RAG) systems traditionally employ sophisticated training strategies to enhance robustness against retrieval noise. In this work, we investigate a critical question: does the benefit of these complex robust training methods diminish as language models become more powerful? Through systematic evaluation across multiple model scales and question-answering datasets, our analysis reveals a consistent trend: \emph{the marginal robustness benefit of sophisticated training strategies decreases substantially as model capacity increases.} While smaller models show significant performance improvements from complex document selection and adversarial objectives, more capable models achieve comparable or even superior performance with simpler training approaches. Further investigation demonstrates that stronger models naturally exhibit better confidence calibration, cross-dataset generalization capability, and more effective attention patterns, even under simple training regimes. These findings suggest that as foundation models evolve, the engineering effort invested in complex robust training may yield diminishing returns, indicating that simplified RAG pipelines could suffice for powerful models while maintaining competitive performance.
Authors: Tao Ji, Bin Guo, Yuanbin Wu, Qipeng Guo, Lixing Shen, Zhan Chen, Xipeng Qiu, Qi Zhang, Tao Gui
Abstract: Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.
Authors: EunJeong Hwang, Peter West, Vered Shwartz
Abstract: Humor is prevalent in online communications and it often relies on more than one modality (e.g., cartoons and memes). Interpreting humor in multimodal settings requires drawing on diverse types of knowledge, including metaphorical, sociocultural, and commonsense knowledge. However, identifying the most useful knowledge remains an open question. We introduce \method{}, a method inspired by the information bottleneck principle that elicits relevant world knowledge from vision and language models which is iteratively refined for generating an explanation of the humor in an unsupervised manner. Our experiments on three datasets confirm the advantage of our method over a range of baselines. Our method can further be adapted in the future for additional tasks that can benefit from eliciting and conditioning on relevant world knowledge and open new research avenues in this direction.
Authors: Pranjal Aggarwal, Sean Welleck
Abstract: Reasoning language models have shown an uncanny ability to improve performance at test-time by ``thinking longer''-that is, by generating longer chain-of-thought sequences and hence using more compute. However, the length of their chain-of-thought reasoning is not controllable, making it impossible to allocate test-time compute to achieve a desired level of performance. We introduce Length Controlled Policy Optimization (LCPO), a simple reinforcement learning method that optimizes for accuracy and adherence to user-specified length constraints. We use LCPO to train L1, a reasoning language model that produces outputs satisfying a length constraint given in its prompt. L1's length control allows for smoothly trading off computational cost and accuracy on a wide range of tasks, and outperforms the state-of-the-art S1 method for length control. Furthermore, we uncover an unexpected short chain-of-thought capability in models trained with LCPO. Specifically, using LCPO we derive Short Reasoning Models (SRMs), that exhibit similar reasoning patterns as full-length reasoning models, but can generate CoT lengths comparable to non-reasoning models. They demonstrate significant performance gains, for instance, our 1.5B L1 model surpasses GPT-4o at equal reasoning lengths. Overall, LCPO enables precise control over reasoning length, allowing for fine-grained allocation of test-time compute and accuracy. We release code and models at https://www.cmu-l3.github.io/l1
Authors: Ziming You, Yumiao Zhang, Dexuan Xu, Yiwei Lou, Yandong Yan, Wei Wang, Huaming Zhang, Yu Huang
Abstract: Existing large language model (LLM) agents for automating data science show promise, but they remain constrained by narrow task scopes, limited generalization across tasks and models, and over-reliance on state-of-the-art (SOTA) LLMs. We introduce DatawiseAgent, a notebook-centric LLM agent framework for adaptive and robust data science automation. Inspired by how human data scientists work in computational notebooks, DatawiseAgent introduces a unified interaction representation and a multi-stage architecture based on finite-state transducers (FSTs). This design enables flexible long-horizon planning, progressive solution development, and robust recovery from execution failures. Extensive experiments across diverse data science scenarios and models show that DatawiseAgent consistently achieves SOTA performance by surpassing strong baselines such as AutoGen and TaskWeaver, demonstrating superior effectiveness and adaptability. Further evaluations reveal graceful performance degradation under weaker or smaller models, underscoring the robustness and scalability.
Authors: Kshitij Ambilduke, Ben Peters, Sonal Sannigrahi, Anil Keshwani, Tsz Kin Lam, Bruno Martins, Andr\'e F. T. Martins, Marcely Zanon Boito
Abstract: We introduce Spire, a speech-augmented language model (LM) capable of both translating and transcribing speech input from English into 10 other languages as well as translating text input in both language directions. Spire integrates the speech modality into an existing multilingual LM via speech discretization and continued pre-training using only 42.5K hours of speech. In particular, we adopt the pretraining framework of multilingual LMs and treat discretized speech input as an additional translation language. This approach not only equips the model with speech capabilities, but also preserves its strong text-based performance. We achieve this using significantly less data than existing speech LMs, demonstrating that discretized speech input integration as an additional language is feasible during LM adaptation. We make our code and models available to the community.
Authors: Noah Y. Siegel, Nicolas Heess, Maria Perez-Ortiz, Oana-Maria Camburu
Abstract: When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans. But are these explanations faithful, i.e. do they convey the factors actually responsible for the decision? In this work, we analyse counterfactual faithfulness across 75 models from 13 families. We analyze the tradeoff between conciseness and comprehensiveness, how correlational faithfulness metrics assess this tradeoff, and the extent to which metrics can be gamed. This analysis motivates two new metrics: the phi-CCT, a simplified variant of the Correlational Counterfactual Test (CCT) which avoids the need for token probabilities while explaining most of the variance of the original test; and F-AUROC, which eliminates sensitivity to imbalanced intervention distributions and captures a model's ability to produce explanations with different levels of detail. Our findings reveal a clear scaling trend: larger and more capable models are consistently more faithful on all metrics we consider. Our code is available at https://github.com/google-deepmind/corr_faith.
Authors: Iuri Macocco, Nora Graichen, Gemma Boleda, Marco Baroni
Abstract: We study last-layer outlier dimensions, i.e. dimensions that display extreme activations for the majority of inputs. We show that outlier dimensions arise in many different modern language models, and trace their function back to the heuristic of constantly predicting frequent words. We further show how a model can block this heuristic when it is not contextually appropriate, by assigning a counterbalancing weight mass to the remaining dimensions, and we investigate which model parameters boost outlier dimensions and when they arise during training. We conclude that outlier dimensions are a specialized mechanism discovered by many distinct models to implement a useful token prediction heuristic.
Authors: Jingjin Wang, Jiawei Han
Abstract: Retrieval Augmented Generation (RAG) has become the standard approach for equipping Large Language Models (LLMs) with up-to-date knowledge. However, standard RAG, relying on independent passage retrieval, often fails to capture the interconnected nature of information required for complex, multi-hop reasoning. While structured RAG methods attempt to address this using knowledge graphs built from triples, we argue that the inherent context loss of triples (context collapse) limits the fidelity of the knowledge representation. We introduce PropRAG, a novel RAG framework that shifts from triples to context-rich propositions and introduces an efficient, LLM-free online beam search over proposition paths to discover multi-step reasoning chains. By coupling a higher-fidelity knowledge representation with explicit path discovery, PropRAG achieves state-of-the-art zero-shot Recall@5 and F1 scores on 2Wiki, HotpotQA, and MuSiQue, advancing non-parametric knowledge integration by improving evidence retrieval through richer representation and efficient reasoning path discovery.
Authors: Tobias Domhan, Dawei Zhu
Abstract: Accurately evaluating machine-translated text remains a long-standing challenge, particularly for long documents. Recent work has shown that large language models (LLMs) can serve as reliable and interpretable sentence-level translation evaluators via MQM error span annotations. With modern LLMs supporting larger context windows, a natural question arises: can we feed entire document translations into an LLM for quality assessment? Ideally, evaluation should be invariant to text length, producing consistent error spans regardless of input granularity. However, our analysis shows that text length significantly impacts evaluation: longer texts lead to fewer error spans and reduced system ranking accuracy. To address this limitation, we evaluate several strategies, including granularity-aligned prompting, Focus Sentence Prompting (FSP), and a fine-tuning approach to better align LLMs with the evaluation task. The latter two methods largely mitigate this length bias, making LLMs more reliable for long-form translation evaluation.
Authors: Linxi Zhao, Sofian Zalouk, Christian K. Belardi, Justin Lovelace, Jin Peng Zhou, Ryan Thomas Noonan, Dongyoung Go, Kilian Q. Weinberger, Yoav Artzi, Jennifer J. Sun
Abstract: Neural language models are black-boxes--both linguistic patterns and factual knowledge are distributed across billions of opaque parameters. This entangled encoding makes it difficult to reliably inspect, verify, or update specific facts. We introduce Limited Memory Language Models (LMLM), a new class of language models that externalizes factual knowledge to external database during pre-training rather than memorizing them. Our pre-training approach strategically masks externally retrieved factual values from the training loss, thereby teaching the model to perform targeted lookups rather than relying on memorization in model weights. Our experiments demonstrate that LMLMs achieve competitive performance compared to significantly larger LLMs on standard benchmarks, while offering the advantages of explicit, editable, and verifiable knowledge bases.
Authors: Mehrdad Ghassabi, Pedram Rostami, Hamidreza Baradaran Kashani, Amirhossein Poursina, Zahra Kazemi, Milad Tavakoli
Abstract: The rapid advancement of language models has demonstrated the potential of artificial intelligence in the healthcare industry. However, small language models struggle with specialized domains in low-resource languages like Persian. While numerous medical-domain websites exist in Persian, no curated dataset or corpus has been available making ours the first of its kind. This study introduces a newly curated dataset comprising 20k doctor-patient Q\&A pairs and 60\% of a 90-million-token crawled corpus from medical magazines. Using a parameter-efficient fine-tuning approach, we enhanced the medical knowledge of the baseline model, aya-expanse-8b. Benchmark evaluations demonstrate that the fine-tuned model achieves improved accuracy in medical question answering and successfully passed the Iranian Basic Medical Science Entrance Exam (IBSEE) in September 2023, which the baseline model did not. Additionally, the fine-tuned model improved Persian-translated MMLU accuracy by an average of 2.67\%. This work highlights the potential of leveraging open-access online data to enrich small language models in medical fields, providing a novel solution for Persian medical AI applications suitable for resource-constrained environments. Future research could explore multimodal input to further enhance performance.
Authors: Weiming Wu, Jin Ye, Zi-kang Wang, Zhi Zhou, Yu-Feng Li, Lan-Zhe Guo
Abstract: Obtaining large-scale, high-quality reasoning data is crucial for improving the geometric reasoning capabilities of multi-modal large language models (MLLMs). However, existing data generation methods, whether based on predefined tem plates or constrained symbolic provers, inevitably face diversity and numerical generalization limitations. To address these limitations, we propose NeSyGeo, a novel neuro-symbolic framework for generating geometric reasoning data. First, we propose a domain-specific language grounded in the entity-attributes-relations paradigm to comprehensively represent all components of plane geometry, along with generative actions defined within this symbolic space. We then design a symbolic-visual-text pipeline that synthesizes symbolic sequences, maps them to visual and textual representations and generates reasoning path with reverse search and forward validation. Based on this framework, we construct NeSyGeo CoT and NeSyGeo-Caption datasets, containing 100k samples, and release a new benchmark NeSyGeo-Test for evaluating geometric reasoning abilities in MLLMs. Experiments demonstrate that the proposal significantly and consistently improves the performance of multiple MLLMs under both reinforcement and supervised fine-tuning. With only 4k samples and two epochs of reinforcement fine-tuning, base models achieve improvements of up to +15.8% on MathVision, +8.4% on MathVerse, and +7.3% on GeoQA. Notably, a 4B model can be improved to outperform an 8B model from the same series on geometric reasoning tasks.s
Authors: Zefang Liu, Yinzhu Quan
Abstract: We introduce EconWebArena, a benchmark for evaluating autonomous agents on complex, multimodal economic tasks in realistic web environments. The benchmark comprises 360 curated tasks from 82 authoritative websites spanning domains such as macroeconomics, labor, finance, trade, and public policy. Each task challenges agents to navigate live websites, interpret structured and visual content, interact with real interfaces, and extract precise, time-sensitive data through multi-step workflows. We construct the benchmark by prompting multiple large language models (LLMs) to generate candidate tasks, followed by rigorous human curation to ensure clarity, feasibility, and source reliability. Unlike prior work, EconWebArena emphasizes fidelity to authoritative data sources and the need for grounded web-based economic reasoning. We evaluate a diverse set of state-of-the-art multimodal LLMs as web agents, analyze failure cases, and conduct ablation studies to assess the impact of visual grounding, plan-based reasoning, and interaction design. Our results reveal substantial performance gaps and highlight persistent challenges in grounding, navigation, and multimodal understanding, positioning EconWebArena as a rigorous testbed for economic web intelligence.
Authors: Yaniv Nikankin, Dana Arad, Yossi Gandelsman, Yonatan Belinkov
Abstract: Vision-Language models (VLMs) show impressive abilities to answer questions on visual inputs (e.g., counting objects in an image), yet demonstrate higher accuracies when performing an analogous task on text (e.g., counting words in a text). We investigate this accuracy gap by identifying and comparing the \textit{circuits} - the task-specific computational sub-graphs - in different modalities. We show that while circuits are largely disjoint between modalities, they implement relatively similar functionalities: the differences lie primarily in processing modality-specific data positions (an image or a text sequence). Zooming in on the image data representations, we observe they become aligned with the higher-performing analogous textual representations only towards later layers, too late in processing to effectively influence subsequent positions. To overcome this, we patch the representations of visual data tokens from later layers back into earlier layers. In experiments with multiple tasks and models, this simple intervention closes a third of the performance gap between the modalities, on average. Our analysis sheds light on the multi-modal performance gap in VLMs and suggests a training-free approach for reducing it.
Authors: Milan Bhan, Jean-Noel Vittaut, Nicolas Chesneau, Sarath Chandar, Marie-Jeanne Lesot
Abstract: Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language explanations may not accurately reflect the model's actual reasoning process, indicating a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, a linear faithfulness probe based on NeuroFaith is developed to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.
Authors: Lihu Chen, Gerard de Melo, Fabian M. Suchanek, Ga\"el Varoquaux
Abstract: It is important for Large Language Models (LLMs) to be aware of the boundary of their knowledge, distinguishing queries they can confidently answer from those that lie beyond their capabilities. Such awareness enables models to perform adaptive inference, such as invoking retrieval-augmented generation (RAG), engaging in slow and deep thinking, or abstaining from answering when appropriate. These mechanisms are key to developing efficient and trustworthy AI. In this work, we propose a method to detect knowledge boundaries via Query-Level Uncertainty, which estimates if a model is capable of answering a given query before generating any tokens, thus avoiding the generation cost. To this end, we propose a novel, training-free method called Internal Confidence, which leverages self-evaluations across layers and tokens to provide a reliable signal of uncertainty. Empirical studies on both factual question answering and mathematical reasoning tasks demonstrate that our Internal Confidence outperforms several baselines in quality of confidence while being computationally cheaper. Furthermore, we demonstrate its benefits in adaptive inference settings, showing that for RAG and model cascading it reduces inference costs while preserving overall performance.
Authors: Aakriti Kumar, Nalin Poungpeth, Diyi Yang, Erina Farrell, Bruce Lambert, Matthew Groh
Abstract: Large language models (LLMs) excel at generating empathic responses in text-based conversations. But, how reliably do they judge the nuances of empathic communication? We investigate this question by comparing how experts, crowdworkers, and LLMs annotate empathic communication across four evaluative frameworks drawn from psychology, natural language processing, and communications applied to 200 real-world conversations where one speaker shares a personal problem and the other offers support. Drawing on 3,150 expert annotations, 2,844 crowd annotations, and 3,150 LLM annotations, we assess inter-rater reliability between these three annotator groups. We find that expert agreement is high but varies across the frameworks' sub-components depending on their clarity, complexity, and subjectivity. We show that expert agreement offers a more informative benchmark for contextualizing LLM performance than standard classification metrics. Across all four frameworks, LLMs consistently approach this expert level benchmark and exceed the reliability of crowdworkers. These results demonstrate how LLMs, when validated on specific tasks with appropriate benchmarks, can support transparency and oversight in emotionally sensitive applications including their use as conversational companions.
Authors: Yuchen Su, Yonghua Zhu, Ruofan Wang, Zijian Huang, Diana Benavides-Prado, Michael Witbrock
Abstract: Pun generation seeks to creatively modify linguistic elements in text to produce humour or evoke double meanings. It also aims to preserve coherence and contextual appropriateness, making it useful in creative writing and entertainment across various media and contexts. Although pun generation has received considerable attention in computational linguistics, there is currently no dedicated survey that systematically reviews this specific area. To bridge this gap, this paper provides a comprehensive review of pun generation datasets and methods across different stages, including conventional approaches, deep learning techniques, and pre-trained language models. Additionally, we summarise both automated and human evaluation metrics used to assess the quality of pun generation. Finally, we discuss the research challenges and propose promising directions for future work.
Authors: Mikey Elmers, Koji Inoue, Divesh Lala, Tatsuya Kawahara
Abstract: Turn-taking is a fundamental component of spoken dialogue, however conventional studies mostly involve dyadic settings. This work focuses on applying voice activity projection (VAP) to predict upcoming turn-taking in triadic multi-party scenarios. The goal of VAP models is to predict the future voice activity for each speaker utilizing only acoustic data. This is the first study to extend VAP into triadic conversation. We trained multiple models on a Japanese triadic dataset where participants discussed a variety of topics. We found that the VAP trained on triadic conversation outperformed the baseline for all models but that the type of conversation affected the accuracy. This study establishes that VAP can be used for turn-taking in triadic dialogue scenarios. Future work will incorporate this triadic VAP turn-taking model into spoken dialogue systems.
Authors: Marlene Lutz, Indira Sen, Georg Ahnert, Elisa Rogers, Markus Strohmaier
Abstract: Persona prompting is increasingly used in large language models (LLMs) to simulate views of various sociodemographic groups. However, how a persona prompt is formulated can significantly affect outcomes, raising concerns about the fidelity of such simulations. Using five open-source LLMs, we systematically examine how different persona prompt strategies, specifically role adoption formats and demographic priming strategies, influence LLM simulations across 15 intersectional demographic groups in both open- and closed-ended tasks. Our findings show that LLMs struggle to simulate marginalized groups but that the choice of demographic priming and role adoption strategy significantly impacts their portrayal. Specifically, we find that prompting in an interview-style format and name-based priming can help reduce stereotyping and improve alignment. Surprisingly, smaller models like OLMo-2-7B outperform larger ones such as Llama-3.3-70B. Our findings offer actionable guidance for designing sociodemographic persona prompts in LLM-based simulation studies.
Authors: Jaskaranjeet Singh, Rakesh Thakur
Abstract: Despite rapid advances in large language models (LLMs), low-resource languages remain excluded from NLP, limiting digital access for millions. We present PunGPT2, the first fully open-source Punjabi generative model suite, trained on a 35GB corpus covering literature, religious texts, news, social discourse, etc. PunGPT2 captures Punjabi's syntactic and morphological richness through a tokenizer optimized for Gurmukhi and Shahmukhi scripts. We introduce Pun-RAG, a retrieval-augmented framework integrating PunGPT2 with a FAISS retriever over a curated Punjabi knowledge base, and Pun-Instruct, an instruction-tuned variant using QLoRA for robust zero-shot summarization, translation, and question answering. Our key innovation, Quantum-RAG, fuses sparse, dense, and quantum kernel embeddings for efficient, context-aware retrieval with low memory overhead, marking the first practical quantum-inspired retrieval in a low-resource LLM. Our models outperform multilingual baselines (mBERT, mT5, MuRIL, BLOOM) on FLORES-200, IndicGenBench, and a new PunjabiEval suite. Quantum-RAG yields +7.4 Recall@10 over FAISS and +3.5 BLEU over mT5 on PunjabiEval. We publicly release all training scripts, hyperparameters, evaluation pipelines, the 35GB Punjabi corpus, the PunjabiEval benchmark, and all model weights, establishing new state-of-the-art results for Punjabi language generation and retrieval.
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 prompt-response pairs from the HelpSteer2 dataset, an open source collection where each response is 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 exemplars (both prompts and responses) to refine its own generation through reflective reasoning, and (2) multi-metric contrastive reasoning, where the LLM analyzes the best exemplars 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: Mutsumi Sasaki, Go Kamoda, Ryosuke Takahashi, Kosuke Sato, Kentaro Inui, Keisuke Sakaguchi, Benjamin Heinzerling
Abstract: Temporal reasoning and knowledge are essential capabilities for language models (LMs). While much prior work has analyzed and improved temporal reasoning in LMs, most studies have focused solely on the Gregorian calendar. However, many non-Gregorian systems, such as the Japanese, Hijri, and Hebrew calendars, are in active use and reflect culturally grounded conceptions of time. If and how well current LMs can accurately handle such non-Gregorian calendars has not been evaluated so far. Here, we present a systematic evaluation of how well open-source LMs handle one such non-Gregorian system: the Japanese calendar. For our evaluation, we create datasets for four tasks that require both temporal knowledge and temporal reasoning. Evaluating a range of English-centric and Japanese-centric LMs, we find that some models can perform calendar conversions, but even Japanese-centric models struggle with Japanese-calendar arithmetic and with maintaining consistency across calendars. Our results highlight the importance of developing LMs that are better equipped for culture-specific calendar understanding.
Authors: Weikang Qiu, Tinglin Huang, Ryan Rullo, Yucheng Kuang, Ali Maatouk, S. Raquel Ramos, Rex Ying
Abstract: Large Language Models (LLMs) hold promise in addressing complex medical problems. However, while most prior studies focus on improving accuracy and reasoning abilities, a significant bottleneck in developing effective healthcare agents lies in the readability of LLM-generated responses, specifically, their ability to answer public health problems clearly and simply to people without medical backgrounds. In this work, we introduce RephQA, a benchmark for evaluating the readability of LLMs in public health question answering (QA). It contains 533 expert-reviewed QA pairs from 27 sources across 13 topics, and includes a proxy multiple-choice task to assess informativeness, along with two readability metrics: Flesch-Kincaid grade level and professional score. Evaluation of 25 LLMs reveals that most fail to meet readability standards, highlighting a gap between reasoning and effective communication. To address this, we explore four readability-enhancing strategies-standard prompting, chain-of-thought prompting, Group Relative Policy Optimization (GRPO), and a token-adapted variant. Token-adapted GRPO achieves the best results, advancing the development of more practical and user-friendly public health agents. These results represent a step toward building more practical agents for public health.
Authors: Hieu Tran, Zonghai Yao, Hong Yu
Abstract: Reinforcement learning improves LLM reasoning, yet sparse delayed reward over long sequences makes token-level credit assignment the key bottleneck. We study the verifiable-reward setting, where the final answer is checkable and multiple responses can be drawn per prompt. Reasoning tasks in math and medical QA align with this setup, where only a few decision tokens significantly impact the outcome. PPO offers token-level advantages with a learned value model, but it is complex to train both the actor and critic models simultaneously, and it is not easily generalizable, as the token-level values from the critic model can make training prone to overfitting. GRPO is critic-free and supports verifiable rewards, but spreads a single sequence-level return across tokens and ignores branching. We introduce \textbf{Prefix-to-Tree (P2T)}, a simple procedure that converts a group of responses into a prefix tree and computes \emph{nonparametric} prefix values \(V(s)\) by aggregating descendant outcomes. Built on P2T, we propose \textbf{TEMPO} (\emph{\textbf{T}ree-\textbf{E}stimated \textbf{M}ean Prefix Value for \textbf{P}olicy \textbf{O}ptimization}), a critic-free algorithm that augments the group-relative outcome signal of GRPO with \emph{branch-gated} temporal-difference corrections derived from the tree. At non-branch tokens, the temporal-difference (TD) term is zero, so TEMPO reduces to GRPO; at branching tokens, it supplies precise token-level credit without a learned value network or extra judges/teachers. On Qwen3-1.7B/4B, TEMPO outperforms PPO and GRPO on in-distribution (MATH, MedQA) and out-of-distribution (GSM-HARD, AMC23, MedMCQA, MMLU-Medical) benchmarks, and reaches higher validation accuracy with roughly the same wall-clock time.
Authors: Yingming Zheng, Hanqi Li, Kai Yu, Lu Chen
Abstract: Large language models (LLMs) have achieved impressive performance across natural language processing (NLP) tasks. As real-world applications increasingly demand longer context windows, continued pretraining and supervised fine-tuning (SFT) on long-context data has become a common approach. While the effects of data length in continued pretraining have been extensively studied, their implications for SFT remain unclear. In this work, we systematically investigate how SFT data length influences LLM behavior on short-context tasks. Counterintuitively, we find that long-context SFT improves short-context performance, contrary to the commonly observed degradation from long-context pretraining. To uncover the underlying mechanisms of this phenomenon, we first decouple and analyze two key components, Multi-Head Attention (MHA) and Feed-Forward Network (FFN), and show that both independently benefit from long-context SFT. We further study their interaction and reveal a knowledge preference bias: long-context SFT promotes contextual knowledge, while short-context SFT favors parametric knowledge, making exclusive reliance on long-context SFT suboptimal. Finally, we demonstrate that hybrid training mitigates this bias, offering explainable guidance for fine-tuning LLMs.
Authors: Feng Wang, Yuqing Li, Han Xiao
Abstract: jina-reranker-v3 is a 0.6B-parameter multilingual listwise reranker that introduces a novel "last but not late" interaction. Unlike late interaction models like ColBERT that encode documents separately before multi-vector matching, our approach applies causal attention between the query and all candidate documents in the same context window, enabling rich interactions before extracting contextual embeddings from each document's final token. The new model achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being significantly smaller than other models with comparable performance.
Authors: Wenrui Bao, Zhiben Chen, Dan Xu, Yuzhang Shang
Abstract: Autoregressive decoding in large language models (LLMs) requires $\mathcal{O}(n)$ sequential steps for $n$ tokens, fundamentally limiting inference throughput. Recent diffusion-based LLMs (dLLMs) enable parallel token generation through iterative denoising. However, current parallel decoding strategies rely on fixed, input-agnostic heuristics (e.g., confidence thresholds), which fail to adapt to input-specific characteristics, resulting in suboptimal speed-quality trade-offs across diverse NLP tasks. In this work, we explore a more flexible and dynamic approach to parallel decoding. We propose Learning to Parallel Decode (Learn2PD), a framework that trains a lightweight and adaptive filter model to predict, for each token position, whether the current prediction matches the final output. This learned filter approximates an oracle parallel decoding strategy that unmasks tokens only when correctly predicted. Importantly, the filter model is learned in a post-training manner, requiring only a small amount of computation to optimize it (minute-level GPU time). Additionally, we introduce End-of-Text Prediction (EoTP) to detect decoding completion at the end of sequence, avoiding redundant decoding of padding tokens. Experiments on the LLaDA benchmark demonstrate that our method achieves up to 22.58$\times$ speedup without any performance drop, and up to 57.51$\times$ when combined with KV-Cache.
Authors: Pengzhou Cheng, Lingzhong Dong, Zeng Wu, Zongru Wu, Xiangru Tang, Chengwei Qin, Zhuosheng Zhang, Gongshen Liu
Abstract: Although numerous strategies have recently been proposed to enhance the autonomous interaction capabilities of multimodal agents in graphical user interface (GUI), their reliability remains limited when faced with complex or out-of-domain tasks. This raises a fundamental question: Are existing multimodal agents reasoning spuriously? In this paper, we propose \textbf{Agent-ScanKit}, a systematic probing framework to unravel the memory and reasoning capabilities of multimodal agents under controlled perturbations. Specifically, we introduce three orthogonal probing paradigms: visual-guided, text-guided, and structure-guided, each designed to quantify the contributions of memorization and reasoning without requiring access to model internals. In five publicly available GUI benchmarks involving 18 multimodal agents, the results demonstrate that mechanical memorization often outweighs systematic reasoning. Most of the models function predominantly as retrievers of training-aligned knowledge, exhibiting limited generalization. Our findings underscore the necessity of robust reasoning modeling for multimodal agents in real-world scenarios, offering valuable insights toward the development of reliable multimodal agents.
Authors: Mariam Mahran, Katharina Simbeck
Abstract: As large language models (LLMs) are increasingly trained on massive, uncurated corpora, understanding both model representations and the data they internalize has become a major challenge. In this work, we show that pairing LLMs with sparse autoencoders (SAEs) enables interpretation not only of model behavior but also of the deeper structures, themes, and biases embedded in the training data. We train a GPT-style transformer model exclusively on the novels of Jane Austen, a corpus rich in social constructs and narrative patterns. We then apply SAEs to hidden states across multiple layers, uncovering sparse, interpretable features that reflect the key narratives and concepts present in the corpus, including gender, class, and societal duty. Our findings demonstrate that LLMs combined with SAEs can act as scalable probes into complex datasets, offering a new path for corpus exploration, bias discovery, and model interpretability at scale.
Authors: Qiyuan Liu, Hao Xu, Xuhong Chen, Wei Chen, Yee Whye Teh, Ning Miao
Abstract: Reward models (RMs) play a critical role in enhancing the reasoning performance of LLMs. For example, they can provide training signals to finetune LLMs during reinforcement learning (RL) and help select the best answer from multiple candidates during inference. In this paper, we provide a systematic introduction to RMs, along with a comprehensive survey of their applications in LLM reasoning. We first review fundamental concepts of RMs, including their architectures, training methodologies, and evaluation techniques. Then, we explore their key applications: (1) guiding generation and selecting optimal outputs during LLM inference, (2) facilitating data synthesis and iterative self-improvement for LLMs, and (3) providing training signals in RL-based finetuning. Finally, we discuss critical open questions regarding the selection, generalization, evaluation, and enhancement of RMs, based on existing research and our own empirical findings. Our analysis aims to provide actionable insights for the effective deployment and advancement of RMs for LLM reasoning.
Authors: Jaehong Yoon, Shoubin Yu, Mohit Bansal
Abstract: Recent video generative models primarily rely on carefully written text prompts for specific tasks, like inpainting or style editing. They require labor-intensive textual descriptions for input videos, hindering their flexibility to adapt personal/raw videos to user specifications. This paper proposes RACCooN, a versatile and user-friendly video-to-paragraph-to-video generative framework that supports multiple video editing capabilities such as removal, addition, and modification, through a unified pipeline. RACCooN consists of two principal stages: Video-to-Paragraph (V2P) and Paragraph-to-Video (P2V). In the V2P stage, we automatically describe video scenes in well-structured natural language, capturing both the holistic context and focused object details. Subsequently, in the P2V stage, users can optionally refine these descriptions to guide the video diffusion model, enabling various modifications to the input video, such as removing, changing subjects, and/or adding new objects. The proposed approach stands out from other methods through several significant contributions: (1) RACCooN suggests a multi-granular spatiotemporal pooling strategy to generate well-structured video descriptions, capturing both the broad context and object details without requiring complex human annotations, simplifying precise video content editing based on text for users. (2) Our video generative model incorporates auto-generated narratives or instructions to enhance the quality and accuracy of the generated content. (3) RACCooN also plans to imagine new objects in a given video, so users simply prompt the model to receive a detailed video editing plan for complex video editing. The proposed framework demonstrates impressive versatile capabilities in video-to-paragraph generation, video content editing, and can be incorporated into other SoTA video generative models for further enhancement.
Authors: Zachary Coalson, Jeonghyun Woo, Chris S. Lin, Joyce Qu, Yu Sun, Shiyang Chen, Lishan Yang, Gururaj Saileshwar, Prashant Nair, Bo Fang, Sanghyun Hong
Abstract: We study a new vulnerability in commercial-scale safety-aligned large language models (LLMs): their refusal to generate harmful responses can be broken by flipping only a few bits in model parameters. Our attack jailbreaks billion-parameter language models with just 5 to 25 bit-flips, requiring up to 40$\times$ fewer bit flips than prior attacks on much smaller computer vision models. Unlike prompt-based jailbreaks, our method directly uncensors models in memory at runtime, enabling harmful outputs without requiring input-level modifications. Our key innovation is an efficient bit-selection algorithm that identifies critical bits for language model jailbreaks up to 20$\times$ faster than prior methods. We evaluate our attack on 10 open-source LLMs, achieving high attack success rates (ASRs) of 80-98% with minimal impact on model utility. We further demonstrate an end-to-end exploit via Rowhammer-based fault injection, reliably jailbreaking 5 models (69-91% ASR) on a GDDR6 GPU. Our analyses reveal that: (1) models with weaker post-training alignment require fewer bit-flips to jailbreak; (2) certain model components, e.g., value projection layers, are substantially more vulnerable; and (3) the attack is mechanistically different from existing jailbreak methods. We evaluate potential countermeasures and find that our attack remains effective against defenses at various stages of the LLM pipeline.
Authors: Cristiano Patr\'icio, Isabel Rio-Torto, Jaime S. Cardoso, Lu\'is F. Teixeira, Jo\~ao C. Neves
Abstract: The main challenges limiting the adoption of deep learning-based solutions in medical workflows are the availability of annotated data and the lack of interpretability of such systems. Concept Bottleneck Models (CBMs) tackle the latter by constraining the model output on a set of predefined and human-interpretable concepts. However, the increased interpretability achieved through these concept-based explanations implies a higher annotation burden. Moreover, if a new concept needs to be added, the whole system needs to be retrained. Inspired by the remarkable performance shown by Large Vision-Language Models (LVLMs) in few-shot settings, we propose a simple, yet effective, methodology, CBVLM, which tackles both of the aforementioned challenges. First, for each concept, we prompt the LVLM to answer if the concept is present in the input image. Then, we ask the LVLM to classify the image based on the previous concept predictions. Moreover, in both stages, we incorporate a retrieval module responsible for selecting the best examples for in-context learning. By grounding the final diagnosis on the predicted concepts, we ensure explainability, and by leveraging the few-shot capabilities of LVLMs, we drastically lower the annotation cost. We validate our approach with extensive experiments across four medical datasets and twelve LVLMs (both generic and medical) and show that CBVLM consistently outperforms CBMs and task-specific supervised methods without requiring any training and using just a few annotated examples. More information on our project page: https://cristianopatricio.github.io/CBVLM/.
Authors: George Fatouros, Kostas Metaxas, John Soldatos, Manos Karathanassis
Abstract: MarketSenseAI is a novel framework for holistic stock analysis which leverages Large Language Models (LLMs) to process financial news, historical prices, company fundamentals and the macroeconomic environment to support decision making in stock analysis and selection. In this paper, we present the latest advancements on MarketSenseAI, driven by rapid technological expansion in LLMs. Through a novel architecture combining Retrieval-Augmented Generation and LLM agents, the framework processes SEC filings and earnings calls, while enriching macroeconomic analysis through systematic processing of diverse institutional reports. We demonstrate a significant improvement in fundamental analysis accuracy over the previous version. Empirical evaluation on S\&P 100 stocks over two years (2023-2024) shows MarketSenseAI achieving cumulative returns of 125.9% compared to the index return of 73.5%, while maintaining comparable risk profiles. Further validation on S\&P 500 stocks during 2024 demonstrates the framework's scalability, delivering a 33.8% higher Sortino ratio than the market. This work marks a significant advancement in applying LLM technology to financial analysis, offering insights into the robustness of LLM-driven investment strategies.
Authors: Yao-Ching Yu, Tsun-Han Chiang, Cheng-Wei Tsai, Chien-Ming Huang, Wen-Kwang Tsao
Abstract: Large Language Models (LLMs) have shown remarkable advancements in specialized fields such as finance, law, and medicine. However, in cybersecurity, we have noticed a lack of open-source datasets, with a particular lack of high-quality cybersecurity pretraining corpora, even though much research indicates that LLMs acquire their knowledge during pretraining. To address this, we present a comprehensive suite of datasets covering all major training stages, including pretraining, instruction fine-tuning, and reasoning distillation with cybersecurity-specific self-reflection data. Extensive ablation studies demonstrate their effectiveness on public cybersecurity benchmarks. In particular, continual pre-training on our dataset yields a 15.9% improvement in the aggregate score, while reasoning distillation leads to a 15.8% gain in security certification (CISSP). We will release all datasets and trained cybersecurity LLMs under the ODC-BY and MIT licenses to encourage further research in the community. For access to all datasets and model weights, please refer to https://huggingface.co/collections/trendmicro-ailab/primus-67b1fd27052b802b4af9d243.
URLs: https://huggingface.co/collections/trendmicro-ailab/primus-67b1fd27052b802b4af9d243.
Authors: Ziyuan Zhang, Darcy Wang, Ningyuan Chen, Rodrigo Mansur, Vahid Sarhangian
Abstract: Large language models (LLMs) are increasingly used to simulate or automate human behavior in complex sequential decision-making settings. A natural question is then whether LLMs exhibit similar decision-making behavior to humans, and can achieve comparable (or superior) performance. In this work, we focus on the exploration-exploitation (E&E) tradeoff, a fundamental aspect of dynamic decision-making under uncertainty. We employ canonical multi-armed bandit (MAB) experiments introduced in the cognitive science and psychiatry literature to conduct a comparative study of the E&E strategies of LLMs, humans, and MAB algorithms. We use interpretable choice models to capture the E&E strategies of the agents and investigate how enabling thinking traces, through both prompting strategies and thinking models, shapes LLM decision-making. We find that enabling thinking in LLMs shifts their behavior toward more human-like behavior, characterized by a mix of random and directed exploration. In a simple stationary setting, thinking-enabled LLMs exhibit similar levels of random and directed exploration compared to humans. However, in more complex, non-stationary environments, LLMs struggle to match human adaptability, particularly in effective directed exploration, despite achieving similar regret in certain scenarios. Our findings highlight both the promise and limits of LLMs as simulators of human behavior and tools for automated decision-making and point to potential areas for improvement.
Authors: Zheng Li, Qingxiu Dong, Jingyuan Ma, Di Zhang, Kai Jia, Zhifang Sui
Abstract: While reasoning models demonstrate exceptional performance on complex tasks, they often exhibit tendencies of overthinking on simple problems. This phenomenon not only leads to excessive computational resource consumption but also significantly degrades user experience. To address this challenge, we propose SelfBudgeter - a novel user-friendly adaptive controllable reasoning framework that incorporates a budget estimation mechanism prior to reasoning. The framework adopts a dual-phase training paradigm: during the cold-start phase, the model learns to predict token budgets before executing reasoning in a standardized format; in the reinforcement learning phase, the model is trained to autonomously plan budgets based on problem difficulty and strictly adhere to them when generating responses. Since the model outputs budget estimates at the initial stage, users can immediately anticipate waiting duration, enabling flexible decisions on whether to interrupt or continue the generation process. Notably, our method supports manual control of reasoning length through pre-filled budget fields. Experimental results demonstrate that SelfBudgeter can dynamically allocate budgets according to problem complexity, yielding an average response length compression of 61% for the 1.5B model on GSM8K, MATH500, and AIME2025, and 48% for the 7B model, while maintaining nearly undiminished accuracy.
Authors: Ruizhe Shi, Minhak Song, Runlong Zhou, Zihan Zhang, Maryam Fazel, Simon S. Du
Abstract: We present a fine-grained theoretical analysis of the performance gap between reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) under a representation gap. Our study decomposes this gap into two sources: an explicit representation gap under exact optimization and an implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is implicitly sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model -- highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.
Authors: Yilin Zhang, Xinran Zhao, Zora Zhiruo Wang, Chenyang Yang, Jiayi Wei, Tongshuang Wu
Abstract: Retrieval-Augmented Generation (RAG) has become essential for large-scale code generation, grounding predictions in external code corpora to improve actuality. However, a critical yet underexplored aspect of RAG pipelines is chunking -- the process of dividing documents into retrievable units. Existing line-based chunking heuristics often break semantic structures, splitting functions or merging unrelated code, which can degrade generation quality. We propose chunking via Abstract Syntax Trees (\ourwork), a structure-aware method that recursively breaks large AST nodes into smaller chunks and merges sibling nodes while respecting size limits. This approach generates self-contained, semantically coherent units across programming languages and tasks, improving performance on diverse code generation tasks, e.g., boosting Recall@5 by 4.3 points on RepoEval retrieval and Pass@1 by 2.67 points on SWE-bench generation. Our work highlights the importance of structure-aware chunking for scaling retrieval-enhanced code intelligence.
Authors: Xiangfeng Wang, Xiao Li, Yadong Wei, Xueyu Song, Yang Song, Xiaoqiang Xia, Fangrui Zeng, Zaiyi Chen, Liu Liu, Gu Xu, Tong Xu
Abstract: The rapid growth of online video content, especially on short video platforms, has created a growing demand for efficient video editing techniques that can condense long-form videos into concise and engaging clips. Existing automatic editing methods predominantly rely on textual cues from ASR transcripts and end-to-end segment selection, often neglecting the rich visual context and leading to incoherent outputs. In this paper, we propose a human-inspired automatic video editing framework (HIVE) that leverages multimodal narrative understanding to address these limitations. Our approach incorporates character extraction, dialogue analysis, and narrative summarization through multimodal large language models, enabling a holistic understanding of the video content. To further enhance coherence, we apply scene-level segmentation and decompose the editing process into three subtasks: highlight detection, opening/ending selection, and pruning of irrelevant content. To facilitate research in this area, we introduce DramaAD, a novel benchmark dataset comprising over 800 short drama episodes and 500 professionally edited advertisement clips. Experimental results demonstrate that our framework consistently outperforms existing baselines across both general and advertisement-oriented editing tasks, significantly narrowing the quality gap between automatic and human-edited videos.
Authors: Anisha Gunjal, Anthony Wang, Elaine Lau, Vaskar Nath, Yunzhong He, Bing Liu, Sean Hendryx
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for complex reasoning tasks with clear correctness signals such as math and coding. However, extending it to real-world reasoning tasks is challenging, as evaluation depends on nuanced, multi-criteria judgments rather than binary correctness. Instance-specific rubrics have recently been used in evaluation benchmarks to capture such judgments, but their potential as reward signals for on-policy post-training remains underexplored. We introduce $\textbf{Rubrics as Rewards}$ (RaR), an on-policy reinforcement learning method that extends RLVR beyond verifiable domains by using rubric-based feedback. Across both medical and science domains, we evaluate multiple strategies for aggregating rubric feedback into rewards. The best RaR variant achieves relative improvements of up to $31\%$ on HealthBench and $7\%$ on GPQA-Diamond over popular LLM-as-judge baselines that rely on direct Likert-based rewards. These results demonstrate that RaR-trained policies adapt well to diverse evaluation formats, performing strongly on both rubric-based and multiple-choice tasks. Moreover, we find that using rubrics as structured reward signals yields better alignment for smaller judges and reduces performance variance across judge scales.
Authors: Chanyeol Choi, Jihoon Kwon, Alejandro Lopez-Lira, Chaewoon Kim, Minjae Kim, Juneha Hwang, Jaeseon Ha, Hojun Choi, Suyeol Yun, Yongjin Kim, Yongjae Lee
Abstract: Accurate information retrieval (IR) is critical in the financial domain, where investors must identify relevant information from large collections of documents. Traditional IR methods -- whether sparse or dense -- often fall short in retrieval accuracy, as it requires not only capturing semantic similarity but also performing fine-grained reasoning over document structure and domain-specific knowledge. Recent advances in large language models (LLMs) have opened up new opportunities for retrieval with multi-step reasoning, where the model ranks passages through iterative reasoning about which information is most relevant to a given query. However, there exists no benchmark to evaluate such capabilities in the financial domain. To address this gap, we introduce FinAgentBench, the first large-scale benchmark for evaluating retrieval with multi-step reasoning in finance -- a setting we term agentic retrieval. The benchmark consists of 26K expert-annotated examples on S&P-500 listed firms and assesses whether LLM agents can (1) identify the most relevant document type among candidates, and (2) pinpoint the key passage within the selected document. Our evaluation framework explicitly separates these two reasoning steps to address context limitations. This design enables to provide a quantitative basis for understanding retrieval-centric LLM behavior in finance. We evaluate a suite of state-of-the-art models and further demonstrated how targeted fine-tuning can significantly improve agentic retrieval performance. Our benchmark provides a foundation for studying retrieval-centric LLM behavior in complex, domain-specific tasks for finance.
Authors: Robert Dilworth
Abstract: When using a public communication channel -- whether formal or informal, such as commenting or posting on social media -- end users have no expectation of privacy: they compose a message and broadcast it for the world to see. Even if an end user takes utmost precautions to anonymize their online presence -- using an alias or pseudonym; masking their IP address; spoofing their geolocation; concealing their operating system and user agent; deploying encryption; registering with a disposable phone number or email; disabling non-essential settings; revoking permissions; and blocking cookies and fingerprinting -- one obvious element still lingers: the message itself. Assuming they avoid lapses in judgment or accidental self-exposure, there should be little evidence to validate their actual identity, right? Wrong. The content of their message -- necessarily open for public consumption -- exposes an attack vector: stylometric analysis, or author profiling. In this paper, we dissect the technique of stylometry, discuss an antithetical counter-strategy in adversarial stylometry, and devise enhancements through Unicode steganography.
Authors: Qikai Chang, Zhenrong Zhang, Pengfei Hu, Jun Du, Jiefeng Ma, Yicheng Pan, Jianshu Zhang, Quan Liu, Jianqing Gao
Abstract: Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. Despite recent advances, existing methods struggle with three key challenges: constructing tool-integrated reasoning data, performing fine-grained optimization, and enhancing inference. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent actor-critic-based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both episode-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer's correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks. Our code will be publicly available at https://github.com/JingMog/THOR.
Authors: Hieu Tran, Zonghai Yao, Nguyen Luong Tran, Zhichao Yang, Feiyun Ouyang, Shuo Han, Razieh Rahimi, Hong Yu
Abstract: Inspired by the dual-process theory of human cognition from \textit{Thinking, Fast and Slow}, we introduce \textbf{PRIME} (Planning and Retrieval-Integrated Memory for Enhanced Reasoning), a multi-agent reasoning framework that dynamically integrates \textbf{System 1} (fast, intuitive thinking) and \textbf{System 2} (slow, deliberate thinking). PRIME first employs a Quick Thinking Agent (System 1) to generate a rapid answer; if uncertainty is detected, it then triggers a structured System 2 reasoning pipeline composed of specialized agents for \textit{planning}, \textit{hypothesis generation}, \textit{retrieval}, \textit{information integration}, and \textit{decision-making}. This multi-agent design faithfully mimics human cognitive processes and enhances both efficiency and accuracy. Experimental results with LLaMA 3 models demonstrate that PRIME enables open-source LLMs to perform competitively with state-of-the-art closed-source models like GPT-4 and GPT-4o on benchmarks requiring multi-hop and knowledge-grounded reasoning. This research establishes PRIME as a scalable solution for improving LLMs in domains requiring complex, knowledge-intensive reasoning.
Authors: He Zhu, Junyou Su, Peng Lai, Ren Ma, Wenjia Zhang, Linyi Yang, Guanhua Chen
Abstract: Post-training of large language models involves a fundamental trade-off between supervised fine-tuning (SFT), which efficiently mimics demonstrations but tends to memorize, and reinforcement learning (RL), which achieves better generalization at higher computational cost. Dynamic Fine-Tuning (DFT) recently emerged as a promising middle ground, reweighting SFT objectives with token probabilities and achieving improvements in certain reasoning domains, though it exhibits instability in other tasks. We provide a analysis of DFT through the reward-weighted regression (RWR) framework, revealing that it corresponds to a specific auxiliary distribution choice that yields provably tighter RL bounds than standard SFT. However, our analysis also uncovers a critical limitation: this construction lacks distributional anchoring, leading to progressive drift that undermines training stability. To address this, we propose Anchored Supervised Fine-Tuning (ASFT), which augments DFT's reweighting with lightweight KL regularization to preserve tightness while ensuring stability. Empirically, ASFT consistently outperforms both SFT and DFT across mathematical reasoning, medical knowledge grounding, and code generation, achieving substantial improvements with minimal computational overhead. Our RWR framework provides a systematic lens for understanding post-training methods and demonstrates that principled theoretical analysis leads to both stronger guarantees and practical gains.
Authors: Claas Beger, Ryan Yi, Shuhao Fu, Arseny Moskvichev, Sarah W. Tsai, Sivasankaran Rajamanickam, Melanie Mitchell
Abstract: OpenAI's o3-preview reasoning model exceeded human accuracy on the ARC-AGI benchmark, but does that mean state-of-the-art models recognize and reason with the abstractions that the task creators intended? We investigate models' abstraction abilities on ConceptARC. We evaluate models under settings that vary the input modality (textual vs. visual), whether the model is permitted to use external Python tools, and, for reasoning models, the amount of reasoning effort. In addition to measuring output accuracy, we perform fine-grained evaluation of the natural-language rules that models generate to explain their solutions. This dual evaluation lets us assess whether models solve tasks using the abstractions ConceptARC was designed to elicit, rather than relying on surface-level patterns. Our results show that, while some models using text-based representations match human output accuracy, the best models' rules are often based on surface-level ``shortcuts'' and capture intended abstractions far less often than humans. Thus their capabilities for general abstract reasoning may be overestimated by evaluations based on accuracy alone. In the visual modality, AI models' output accuracy drops sharply, yet our rule-level analysis reveals that models might be underestimated, as they still exhibit a substantial share of rules that capture intended abstractions, but are often unable to correctly apply these rules. In short, our results show that models still lag humans in abstract reasoning, and that using accuracy alone to evaluate abstract reasoning on ARC-like tasks may overestimate abstract-reasoning capabilities in textual modalities and underestimate it in visual modalities. We believe that our evaluation framework offers a more faithful picture of multimodal models' abstract reasoning abilities and a more principled way to track progress toward human-like, abstraction-centered intelligence.