Authors: Jinzhu Yang
Abstract: This study is dedicated to exploring the application of prompt learning methods to advance Named Entity Recognition (NER) within the medical domain. In recent years, the emergence of large-scale models has driven significant progress in NER tasks, particularly with the introduction of the BioBERT language model, which has greatly enhanced NER capabilities in medical texts. Our research introduces the Prompt-bioMRC model, which integrates both hard template and soft prompt designs aimed at refining the precision and efficiency of medical entity recognition. Through extensive experimentation across diverse medical datasets, our findings consistently demonstrate that our approach surpasses traditional models. This enhancement not only validates the efficacy of our methodology but also highlights its potential to provide reliable technological support for applications like intelligent diagnosis systems. By leveraging advanced NER techniques, this study contributes to advancing automated medical data processing, facilitating more accurate medical information extraction, and supporting efficient healthcare decision-making processes.
Authors: Lukas Rauch, Moritz Wirth, Denis Huseljic, Marek Herde, Bernhard Sick, Matthias A{\ss}enmacher
Abstract: The advent of large language models (LLMs) capable of producing general-purpose representations lets us revisit the practicality of deep active learning (AL): By leveraging frozen LLM embeddings, we can mitigate the computational costs of iteratively fine-tuning large backbones. This study establishes a benchmark and systematically investigates the influence of LLM embedding quality on query strategies in deep AL. We employ five top-performing models from the massive text embedding benchmark (MTEB) leaderboard and two baselines for ten diverse text classification tasks. Our findings reveal key insights: First, initializing the labeled pool using diversity-based sampling synergizes with high-quality embeddings, boosting performance in early AL iterations. Second, the choice of the optimal query strategy is sensitive to embedding quality. While the computationally inexpensive Margin sampling can achieve performance spikes on specific datasets, we find that strategies like Badge exhibit greater robustness across tasks. Importantly, their effectiveness is often enhanced when paired with higher-quality embeddings. Our results emphasize the need for context-specific evaluation of AL strategies, as performance heavily depends on embedding quality and the target task.
Authors: Abhay Gupta, Michael Lu, Kevin Zhu, Sean O'Brien, Vasu Sharma
Abstract: Current large language models (LLMs) struggle to answer questions that span tens of thousands of tokens, especially when multi-hop reasoning is involved. While prior benchmarks explore long-context comprehension or multi-hop reasoning in isolation, none jointly vary context length and reasoning depth in natural narrative settings. We introduce NovelHopQA, the first benchmark to evaluate k1-4 hop QA over 64k-128k-token excerpts from 83 full-length public-domain novels. A keyword-guided pipeline builds hop-separated chains grounded in coherent storylines. We evaluate six state-of-the-art (SOTA) models and apply oracle-context filtering to ensure all questions are genuinely answerable. Human annotators validate both alignment and hop depth. We noticed consistent accuracy drops with increased hops and context length, even in frontier models-revealing that sheer scale does not guarantee robust reasoning. Our failure mode analysis highlights common breakdowns, such as missed final-hop integration and long-range drift. NovelHopQA offers a controlled diagnostic setting to stress-test multi-hop reasoning at scale.
Authors: Timothy Do, Pranav Saran, Harshita Poojary, Pranav Prabhu, Sean O'Brien, Vasu Sharma, Kevin Zhu
Abstract: In this paper, we address the persistent challenges that figurative language expressions pose for natural language processing (NLP) systems, particularly in low-resource languages such as Konkani. We present a hybrid model that integrates a pre-trained Multilingual BERT (mBERT) with a bidirectional LSTM and a linear classifier. This architecture is fine-tuned on a newly introduced annotated dataset for metaphor classification, developed as part of this work. To improve the model's efficiency, we implement a gradient-based attention head pruning strategy. For metaphor classification, the pruned model achieves an accuracy of 78%. We also applied our pruning approach to expand on an existing idiom classification task, achieving 83% accuracy. These results demonstrate the effectiveness of attention head pruning for building efficient NLP tools in underrepresented languages.
Authors: Christopher Lee L\"ubbers
Abstract: Paraphrasing re-expresses meaning to enhance applications like text simplification, machine translation, and question-answering. Specific paraphrase types facilitate accurate semantic analysis and robust language models. However, existing paraphrase-type generation methods often misalign with human preferences due to reliance on automated metrics and limited human-annotated training data, obscuring crucial aspects of semantic fidelity and linguistic transformations. This study addresses this gap by leveraging a human-ranked paraphrase-type dataset and integrating Direct Preference Optimization (DPO) to align model outputs directly with human judgments. DPO-based training increases paraphrase-type generation accuracy by 3 percentage points over a supervised baseline and raises human preference ratings by 7 percentage points. A newly created human-annotated dataset supports more rigorous future evaluations. Additionally, a paraphrase-type detection model achieves F1 scores of 0.91 for addition/deletion, 0.78 for same polarity substitution, and 0.70 for punctuation changes. These findings demonstrate that preference data and DPO training produce more reliable, semantically accurate paraphrases, enabling downstream applications such as improved summarization and more robust question-answering. The PTD model surpasses automated metrics and provides a more reliable framework for evaluating paraphrase quality, advancing paraphrase-type research toward richer, user-aligned language generation and establishing a stronger foundation for future evaluations grounded in human-centric criteria.
Authors: E Fan, Weizong Wang, Tianhan Zhang
Abstract: Computational Fluid Dynamics (CFD) is essential for scientific and engineering advancements but is limited by operational complexity and the need for extensive expertise. This paper presents ChatCFD, a large language model-driven pipeline that automates CFD workflows within the OpenFOAM framework. It enables users to configure and execute complex simulations from natural language prompts or published literature with minimal expertise. The innovation is its structured approach to database construction, configuration validation, and error reflection, integrating CFD and OpenFOAM knowledge with general language models to improve accuracy and adaptability. Validation shows ChatCFD can autonomously reproduce published CFD results, handling complex, unseen configurations beyond basic examples, a task challenging for general language models.
Authors: Feng Wang, Yiding Sun, Jiaxin Mao, Wei Xue, Danqing Xu
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across various professional domains, with their performance typically evaluated through standardized benchmarks. However, the development of financial RAG benchmarks has been constrained by data confidentiality issues and the lack of dynamic data integration. To address this issue, we introduces FinS-Pilot, a novel benchmark for evaluating RAG systems in online financial applications. Constructed from real-world financial assistant interactions, our benchmark incorporates both real-time API data and structured text sources, organized through an intent classification framework covering critical financial domains such as equity analysis and macroeconomic forecasting. The benchmark enables comprehensive evaluation of financial assistants' capabilities in handling both static knowledge and time-sensitive market information. Through systematic experiments with multiple Chinese leading LLMs, we demonstrate FinS-Pilot's effectiveness in identifying models suitable for financial applications while addressing the current gap in specialized evaluation tools for the financial domain. Our work contributes both a practical evaluation framework and a curated dataset to advance research in financial NLP systems. The code and dataset are accessible on GitHub\footnote{https://github.com/PhealenWang/financial\_rag\_benchmark}.
URLs: https://github.com/PhealenWang/financial\_rag\_benchmark
Authors: Duzhen Zhang, Yong Ren, Zhong-Zhi Li, Yahan Yu, Jiahua Dong, Chenxing Li, Zhilong Ji, Jinfeng Bai
Abstract: Multimodal Continual Instruction Tuning (MCIT) aims to finetune Multimodal Large Language Models (MLLMs) to continually align with human intent across sequential tasks. Existing approaches often rely on the Mixture-of-Experts (MoE) LoRA framework to preserve previous instruction alignments. However, these methods are prone to Catastrophic Forgetting (CF), as they aggregate all LoRA blocks via simple summation, which compromises performance over time. In this paper, we identify a critical parameter inefficiency in the MoELoRA framework within the MCIT context. Based on this insight, we propose BranchLoRA, an asymmetric framework to enhance both efficiency and performance. To mitigate CF, we introduce a flexible tuning-freezing mechanism within BranchLoRA, enabling branches to specialize in intra-task knowledge while fostering inter-task collaboration. Moreover, we incrementally incorporate task-specific routers to ensure an optimal branch distribution over time, rather than favoring the most recent task. To streamline inference, we introduce a task selector that automatically routes test inputs to the appropriate router without requiring task identity. Extensive experiments on the latest MCIT benchmark demonstrate that BranchLoRA significantly outperforms MoELoRA and maintains its superiority across various MLLM sizes.
Authors: Xiang Li, Jiayi Xin, Qi Long, Weijie J. Su
Abstract: Accurate evaluation of large language models (LLMs) is crucial for understanding their capabilities and guiding their development. However, current evaluations often inconsistently reflect the actual capacities of these models. In this paper, we demonstrate that one of many contributing factors to this \textit{evaluation crisis} is the oversight of unseen knowledge -- information encoded by LLMs but not directly observed or not yet observed during evaluations. We introduce KnowSum, a statistical framework designed to provide a more comprehensive assessment by quantifying the unseen knowledge for a class of evaluation tasks. KnowSum estimates the unobserved portion by extrapolating from the appearance frequencies of observed knowledge instances. We demonstrate the effectiveness and utility of KnowSum across three critical applications: estimating total knowledge, evaluating information retrieval effectiveness, and measuring output diversity. Our experiments reveal that a substantial volume of knowledge is omitted when relying solely on observed LLM performance. Importantly, KnowSum yields significantly different comparative rankings for several common LLMs based on their internal knowledge.
Authors: Juncheng Wu, Sheng Liu, Haoqin Tu, Hang Yu, Xiaoke Huang, James Zou, Cihang Xie, Yuyin Zhou
Abstract: Recent advances in reasoning-enhanced Large Language Models such as OpenAI-o1/3 and DeepSeek-R1 have significantly improved performance on complex tasks. However, the quality and transparency of their internal reasoning processes remain underexplored. This work moves beyond the final-answer accuracy and investigates step-by-step reasoning in the medical and mathematical domains by explicitly decomposing the thinking trajectories into two parts: knowledge and reasoning. Specifically, we introduce a fine-grained evaluation framework that judges: (1) the correctness of knowledge used (measured by Knowledge Index (KI)) and (2) the quality of reasoning (measured by Information Gain (InfoGain)). Using this framework, we study R1-distilled and base Qwen models trained with supervised fine-tuning (SFT) and/or reinforcement learning (RL) in the medical and math domains. Three intriguing findings emerge: (1) The general reasoning abilities in R1-distilled models do not transfer effectively to the medical domain through either SFT or RL. (2) SFT raises final-answer accuracy in both domains, but often at the cost of reasoning quality: InfoGain drops by 38.9% on average compared with untrained models; In the medical domain, however, SFT remains crucial because domain knowledge is indispensable. (3) RL enhances medical reasoning by pruning inaccurate or irrelevant knowledge from reasoning paths, thereby improving both reasoning accuracy and knowledge correctness.
Authors: Michael Li, Nishant Subramani
Abstract: Large transformer-based language models dominate modern NLP, yet our understanding of how they encode linguistic information is rooted in studies of early models like BERT and GPT-2. To better understand today's language models, we investigate how both classical architectures (BERT, DeBERTa, GPT-2)and contemporary large language models (Pythia, OLMo-2, Gemma-2, Qwen2.5, Llama-3.1) represent lexical identity and inflectional morphology. We train linear and nonlinear classifiers on layer-wise activations to predict word lemmas and inflectional features. We discover that models concentrate lexical information linearly in early layers and increasingly nonlinearly in later layers, while keeping inflectional information uniformly accessible and linearly separable throughout the layers. Further analysis reveals that these models encode inflectional morphology through generalizable abstractions, but rely predominantly on memorization to encode lexical identity. Remarkably, these patterns emerge across all 16 models we test, despite differences in architecture, size, and training regime (including pretrained and instruction-tuned variants). This consistency suggests that, despite substantial advances in LLM technologies, transformer models organize linguistic information in similar ways, indicating that these properties could be fundamental for next token prediction and are learned early during pretraining. Our code is available at https://github.com/ml5885/model_internal_sleuthing
Authors: Joshua Rozner, Leonie Weissweiler, Cory Shain
Abstract: Construction grammar posits that children acquire constructions (form-meaning pairings) from the statistics of their environment. Recent work supports this hypothesis by showing sensitivity to constructions in pretrained language models (PLMs), including one recent study (Rozner et al., 2025) demonstrating that constructions shape the PLM's output distribution. However, models under study have generally been trained on developmentally implausible amounts of data, casting doubt on their relevance to human language learning. Here we use Rozner et al.'s methods to evaluate constructional learning in models from the 2024 BabyLM challenge. Our results show that even when trained on developmentally plausible quantities of data, models represent diverse constructions, even hard cases that are superficially indistinguishable. We further find correlational evidence that constructional performance may be functionally relevant: models that better represent constructions perform better on the BabyLM benchmarks.
Authors: Amir Hussein, Cihan Xiao, Matthew Wiesner, Dan Povey, Leibny Paola Garcia, Sanjeev Khudanpur
Abstract: Neural transducers (NT) provide an effective framework for speech streaming, demonstrating strong performance in automatic speech recognition (ASR). However, the application of NT to speech translation (ST) remains challenging, as existing approaches struggle with word reordering and performance degradation when jointly modeling ASR and ST, resulting in a gap with attention-based encoder-decoder (AED) models. Existing NT-based ST approaches also suffer from high computational training costs. To address these issues, we propose HENT-SRT (Hierarchical Efficient Neural Transducer for Speech Recognition and Translation), a novel framework that factorizes ASR and translation tasks to better handle reordering. To ensure robust ST while preserving ASR performance, we use self-distillation with CTC consistency regularization. Moreover, we improve computational efficiency by incorporating best practices from ASR transducers, including a down-sampled hierarchical encoder, a stateless predictor, and a pruned transducer loss to reduce training complexity. Finally, we introduce a blank penalty during decoding, reducing deletions and improving translation quality. Our approach is evaluated on three conversational datasets Arabic, Spanish, and Mandarin achieving new state-of-the-art performance among NT models and substantially narrowing the gap with AED-based systems.
Authors: Dennis Fucci, Marco Gaido, Matteo Negri, Luisa Bentivogli, Andre Martins, Giuseppe Attanasio
Abstract: Recent studies on interpreting the hidden states of speech models have shown their ability to capture speaker-specific features, including gender. Does this finding also hold for speech translation (ST) models? If so, what are the implications for the speaker's gender assignment in translation? We address these questions from an interpretability perspective, using probing methods to assess gender encoding across diverse ST models. Results on three language directions (English-French/Italian/Spanish) indicate that while traditional encoder-decoder models capture gender information, newer architectures -- integrating a speech encoder with a machine translation system via adapters -- do not. We also demonstrate that low gender encoding capabilities result in systems' tendency toward a masculine default, a translation bias that is more pronounced in newer architectures.
Authors: Salman Rahman, Sheriff Issaka, Ashima Suvarna, Genglin Liu, James Shiffer, Jaeyoung Lee, Md Rizwan Parvez, Hamid Palangi, Shi Feng, Nanyun Peng, Yejin Choi, Julian Michael, Liwei Jiang, Saadia Gabriel
Abstract: As AI grows more powerful, it will increasingly shape how we understand the world. But with this influence comes the risk of amplifying misinformation and deepening social divides-especially on consequential topics like public health where factual accuracy directly impacts well-being. Scalable Oversight aims to ensure AI truthfulness by enabling humans to supervise systems that may exceed human capabilities--yet humans themselves hold different beliefs and biases that impair their judgment. We study whether AI debate can guide biased judges toward the truth by having two AI systems debate opposing sides of controversial COVID-19 factuality claims where people hold strong prior beliefs. We conduct two studies: one with human judges holding either mainstream or skeptical beliefs evaluating factuality claims through AI-assisted debate or consultancy protocols, and a second examining the same problem with personalized AI judges designed to mimic these different human belief systems. In our human study, we find that debate-where two AI advisor systems present opposing evidence-based arguments-consistently improves judgment accuracy and confidence calibration, outperforming consultancy with a single-advisor system by 10% overall. The improvement is most significant for judges with mainstream beliefs (+15.2% accuracy), though debate also helps skeptical judges who initially misjudge claims move toward accurate views (+4.7% accuracy). In our AI judge study, we find that AI judges with human-like personas achieve even higher accuracy (78.5%) than human judges (70.1%) and default AI judges without personas (69.8%), suggesting their potential for supervising frontier AI models. These findings highlight AI debate as a promising path toward scalable, bias-resilient oversight--leveraging both diverse human and AI judgments to move closer to truth in contested domains.
Authors: Dennis Fucci, Marco Gaido, Matteo Negri, Mauro Cettolo, Luisa Bentivogli
Abstract: Despite significant advances in ASR, the specific acoustic cues models rely on remain unclear. Prior studies have examined such cues on a limited set of phonemes and outdated models. In this work, we apply a feature attribution technique to identify the relevant acoustic cues for a modern Conformer-based ASR system. By analyzing plosives, fricatives, and vowels, we assess how feature attributions align with their acoustic properties in the time and frequency domains, also essential for human speech perception. Our findings show that the ASR model relies on vowels' full time spans, particularly their first two formants, with greater saliency in male speech. It also better captures the spectral characteristics of sibilant fricatives than non-sibilants and prioritizes the release phase in plosives, especially burst characteristics. These insights enhance the interpretability of ASR models and highlight areas for future research to uncover potential gaps in model robustness.
Authors: Lindia Tjuatja, Graham Neubig
Abstract: Language model evaluation is a daunting task: prompts are brittle, corpus-level perplexities are vague, and the choice of benchmarks are endless. Finding examples that show meaningful, generalizable differences between two LMs is crucial to understanding where one model succeeds and another fails. Can this process be done automatically? In this work, we propose methodology for automated comparison of language models that uses performance-aware contextual embeddings to find fine-grained features of text where one LM outperforms another. Our method, which we name BehaviorBox, extracts coherent features that demonstrate differences with respect to the ease of generation between two LMs. Specifically, BehaviorBox finds features that describe groups of words in fine-grained contexts, such as "conditional 'were' in the phrase 'if you were'" and "exclamation marks after emotional statements", where one model outperforms another within a particular datatset. We apply BehaviorBox to compare models that vary in size, model family, and post-training, and enumerate insights into specific contexts that illustrate meaningful differences in performance which cannot be found by measures such as corpus-level perplexity alone.
Authors: Ella Rannon, David Burstein
Abstract: Natural Language Processing (NLP) has transformed various fields beyond linguistics by applying techniques originally developed for human language to the analysis of biological sequences. This review explores the application of NLP methods to biological sequence data, focusing on genomics, transcriptomics, and proteomics. We examine how various NLP methods, from classic approaches like word2vec to advanced models employing transformers and hyena operators, are being adapted to analyze DNA, RNA, protein sequences, and entire genomes. The review also examines tokenization strategies and model architectures, evaluating their strengths, limitations, and suitability for different biological tasks. We further cover recent advances in NLP applications for biological data, such as structure prediction, gene expression, and evolutionary analysis, highlighting the potential of these methods for extracting meaningful insights from large-scale genomic data. As language models continue to advance, their integration into bioinformatics holds immense promise for advancing our understanding of biological processes in all domains of life.
Authors: Sofoklis Kakouros
Abstract: In emotion recognition from speech, a key challenge lies in identifying speech signal segments that carry the most relevant acoustic variations for discerning specific emotions. Traditional approaches compute functionals for features such as energy and F0 over entire sentences or longer speech portions, potentially missing essential fine-grained variation in the long-form statistics. This research investigates the use of word informativeness, derived from a pre-trained language model, to identify semantically important segments. Acoustic features are then computed exclusively for these identified segments, enhancing emotion recognition accuracy. The methodology utilizes standard acoustic prosodic features, their functionals, and self-supervised representations. Results indicate a notable improvement in recognition performance when features are computed on segments selected based on word informativeness, underscoring the effectiveness of this approach.
Authors: Radin Shayanfar, Chu Fei Luo, Rohan Bhambhoria, Samuel Dahan, Xiaodan Zhu
Abstract: It is often challenging to teach specialized, unseen tasks to dialogue systems due to the high cost of expert knowledge, training data, and high technical difficulty. To support domain-specific applications - such as law, medicine, or finance - it is essential to build frameworks that enable non-technical experts to define, test, and refine system behaviour with minimal effort. Achieving this requires cross-disciplinary collaboration between developers and domain specialists. In this work, we introduce a novel framework, CoDial (Code for Dialogue), that converts expert knowledge, represented as a novel structured heterogeneous graph, into executable conversation logic. CoDial can be easily implemented in existing guardrailing languages, such as Colang, to enable interpretable, modifiable, and true zero-shot specification of task-oriented dialogue systems. Empirically, CoDial achieves state-of-the-art performance on the STAR dataset for inference-based models and is competitive with similar baselines on the well-known MultiWOZ dataset. We also demonstrate CoDial's iterative improvement via manual and LLM-aided feedback, making it a practical tool for expert-guided alignment of LLMs in high-stakes domains.
Authors: Wenzheng Zhang, Xi Victoria Lin, Karl Stratos, Wen-tau Yih, Mingda Chen
Abstract: Retrieval-Augmented Generation (RAG) systems traditionally treat retrieval and generation as separate processes, requiring explicit textual queries to connect them. This separation can limit the ability of models to generalize across diverse tasks. In this work, we propose a query-free RAG system, named ImpRAG, which integrates retrieval and generation into a unified model. ImpRAG allows models to implicitly express their information needs, eliminating the need for human-specified queries. By dividing pretrained decoder-only language models into specialized layer groups, ImpRAG optimizes retrieval and generation tasks simultaneously. Our approach employs a two-stage inference process, using the same model parameters and forward pass for both retrieval and generation, thereby minimizing the disparity between retrievers and language models. Experiments on 8 knowledge-intensive tasks demonstrate that ImpRAG achieves 3.6-11.5 improvements in exact match scores on unseen tasks with diverse formats, highlighting its effectiveness in enabling models to articulate their own information needs and generalize across tasks. Our analysis underscores the importance of balancing retrieval and generation parameters and leveraging generation perplexities as retrieval training objectives for enhanced performance.
Authors: Sofoklis Kakouros, Haoyu Chen
Abstract: This study examines the prosodic characteristics associated with winning and losing in post-match tennis interviews. Additionally, this research explores the potential to classify match outcomes solely based on post-match interview recordings using prosodic features and self-supervised learning (SSL) representations. By analyzing prosodic elements such as pitch and intensity, alongside SSL models like Wav2Vec 2.0 and HuBERT, the aim is to determine whether an athlete has won or lost their match. Traditional acoustic features and deep speech representations are extracted from the data, and machine learning classifiers are employed to distinguish between winning and losing players. Results indicate that SSL representations effectively differentiate between winning and losing outcomes, capturing subtle speech patterns linked to emotional states. At the same time, prosodic cues -- such as pitch variability -- remain strong indicators of victory.
Authors: Thai Hoang, Kung-Hsiang Huang, Shirley Kokane, Jianguo Zhang, Zuxin Liu, Ming Zhu, Jake Grigsby, Tian Lan, Michael S Ryoo, Chien-Sheng Wu, Shelby Heinecke, Huan Wang, Silvio Savarese, Caiming Xiong, Juan Carlos Niebles
Abstract: Large Action Models (LAMs) for AI Agents offer incredible potential but face challenges due to the need for high-quality training data, especially for multi-steps tasks that involve planning, executing tool calls, and responding to feedback. To address these issues, we present LAM SIMULATOR, a comprehensive framework designed for online exploration of agentic tasks with high-quality feedback. Our framework features a dynamic task query generator, an extensive collection of tools, and an interactive environment where Large Language Model (LLM) Agents can call tools and receive real-time feedback. This setup enables LLM Agents to explore and solve tasks autonomously, facilitating the discovery of multiple approaches to tackle any given task. The resulting action trajectory data are then used to create high-quality training datasets for LAMs. Our experiments on popular agentic benchmarks, ToolBench and CRMArena, highlight the effectiveness of LAM SIMULATOR: models trained with self-generated datasets using our framework achieve significant performance gains, up to a 49.3\% improvement over their original baselines. LAM SIMULATOR requires minimal human input during dataset creation, highlighting LAM SIMULATOR's efficiency and effectiveness in speeding up development of AI agents.
Authors: Russell Scheinberg, Ameeta Agrawal, Amber Shore, So Young Lee
Abstract: Large language models (LLMs) can explain grammatical rules, yet they often fail to apply those rules when judging sentence acceptability. We present "grammar prompting", an explain-then-process paradigm: a large LLM first produces a concise explanation of the relevant syntactic phenomenon, then that explanation is fed back as additional context to the target model -- either an LLM or a smaller language model (SLM) -- before deciding which sentence of a minimal pair is grammatical. On the English BLiMP, Chinese SLING, and Russian RuBLiMP benchmarks, this simple prompt design yields substantial improvements over strong baselines across many syntactic phenomena. Feeding an LLM's metalinguistic explanation back to the target model bridges the gap between knowing a rule and using it. On SLMs, grammar prompting alone trims the average LLM-SLM accuracy gap by about 20%, and when paired with chain-of-thought, by 56% (13.0 pp -> 5.8 pp), all at negligible cost. The lightweight, language-agnostic cue lets low-cost SLMs approach frontier-LLM performance in multilingual settings.
Authors: Pegah Alipoormolabashi, Ajay Patel, Niranjan Balasubramanian
Abstract: Authorship misattribution can have profound consequences in real life. In forensic settings simply being considered as one of the potential authors of an evidential piece of text or communication can result in undesirable scrutiny. This raises a fairness question: Is every author in the candidate pool at equal risk of misattribution? Standard evaluation measures for authorship attribution systems do not explicitly account for this notion of fairness. We introduce a simple measure, Misattribution Unfairness Index (MAUIk), which is based on how often authors are ranked in the top k for texts they did not write. Using this measure we quantify the unfairness of five models on two different datasets. All models exhibit high levels of unfairness with increased risks for some authors. Furthermore, we find that this unfairness relates to how the models embed the authors as vectors in the latent search space. In particular, we observe that the risk of misattribution is higher for authors closer to the centroid (or center) of the embedded authors in the haystack. These results indicate the potential for harm and the need for communicating with and calibrating end users on misattribution risk when building and providing such models for downstream use.
Authors: Berk Atil, Namrata Sureddy, Rebecca J. Passonneau
Abstract: Toxicity in online content, including content generated by language models, has become a critical concern due to its potential for negative psychological and social impact. This paper introduces TRuST, a comprehensive dataset designed to improve toxicity detection that merges existing datasets, and has labels for toxicity, target social group, and toxic spans. It includes a diverse range of target groups such as ethnicity, gender, religion, disability, and politics, with both human/machine-annotated and human machine-generated data. We benchmark state-of-the-art large language models (LLMs) on toxicity detection, target group identification, and toxic span extraction. We find that fine-tuned models consistently outperform zero-shot and few-shot prompting, though performance remains low for certain social groups. Further, reasoning capabilities do not significantly improve performance, indicating that LLMs have weak social reasoning skills.
Authors: Hyungjoo Chae, Dongjin Kang, Jihyuk Kim, Beong-woo Kwak, Sunghyun Park, Haeju Park, Jinyoung Yeo, Moontae Lee, Kyungjae Lee
Abstract: With the release of R1, a publicly available large reasoning model (LRM), researchers commonly train new LRMs by training language models on R1's long chain-of-thought (CoT) inferences. While prior works show that LRMs' capabilities can be reproduced through direct distillation, the continued reliance on the existing models (e.g., R1) remains a critical limitation in advancing the field. As a first step toward independent LRM development, this paper explores the possibility of constructing a long CoT dataset with LLMs that are not trained for inference-time scaling. To this end, we present the Long CoT Collection, a dataset of 100K CoT rationales annotated using existing short CoT LLMs. We develop a pipeline that induces o1's novel reasoning strategies into short CoT LLMs, enabling them to think longer and introducing controllability over the thought budget to better manage the overthinking problem. Our extensive analyses validate that our dataset achieves quality comparable to--or slightly below--R1. Furthermore, our experiments demonstrate that training on our dataset not only strengthens general reasoning skills, but also provides a strong foundation for reinforcement learning--models initialized on our data achieve 2-3x larger gains with RLVR.
Authors: Jiaming Li, Yukun Chen, Ziqiang Liu, Minghuan Tan, Lei Zhang, Yunshui Li, Run Luo, Longze Chen, Jing Luo, Ahmadreza Argha, Hamid Alinejad-Rokny, Wei Zhou, Min Yang
Abstract: Stories are central to human culture, serving to share ideas, preserve traditions, and foster connections. Automatic story generation, a key advancement in artificial intelligence (AI), offers new possibilities for creating personalized content, exploring creative ideas, and enhancing interactive experiences. However, existing methods struggle to maintain narrative coherence and logical consistency. This disconnect compromises the overall storytelling experience, underscoring the need for substantial improvements. Inspired by human cognitive processes, we introduce Storyteller, a novel approach that systemically improves the coherence and consistency of automatically generated stories. Storyteller introduces a plot node structure based on linguistically grounded subject verb object (SVO) triplets, which capture essential story events and ensure a consistent logical flow. Unlike previous methods, Storyteller integrates two dynamic modules, the STORYLINE and narrative entity knowledge graph (NEKG),that continuously interact with the story generation process. This integration produces structurally sound, cohesive and immersive narratives. Extensive experiments demonstrate that Storyteller significantly outperforms existing approaches, achieving an 84.33% average win rate through human preference evaluation. At the same time, it is also far ahead in other aspects including creativity, coherence, engagement, and relevance.
Authors: Herun Wan, Jiaying Wu, Minnan Luo, Zhi Zeng, Zhixiong Su
Abstract: Misinformation detection models often rely on superficial cues (i.e., \emph{shortcuts}) that correlate with misinformation in training data but fail to generalize to the diverse and evolving nature of real-world misinformation. This issue is exacerbated by large language models (LLMs), which can easily generate convincing misinformation through simple prompts. We introduce TruthOverTricks, a unified evaluation paradigm for measuring shortcut learning in misinformation detection. TruthOverTricks categorizes shortcut behaviors into intrinsic shortcut induction and extrinsic shortcut injection, and evaluates seven representative detectors across 14 popular benchmarks, along with two new factual misinformation datasets, NQ-Misinfo and Streaming-Misinfo. Empirical results reveal that existing detectors suffer severe performance degradation when exposed to both naturally occurring and adversarially crafted shortcuts. To address this, we propose SMF, an LLM-augmented data augmentation framework that mitigates shortcut reliance through paraphrasing, factual summarization, and sentiment normalization. SMF consistently enhances robustness across 16 benchmarks, encouraging models to rely on deeper semantic understanding rather than shortcut cues. To promote the development of misinformation detectors, we have published the resources publicly at https://github.com/whr000001/TruthOverTricks.
Authors: Jeonghun Kang, Soonmok Kwon, Joonseok Lee, Byung-Hak Kim
Abstract: Traditional approaches -- such as Win Probability Added (WPA)-based ranking or computer vision-driven event detection -- can identify scoring plays but often miss strategic depth, momentum shifts, and storyline progression. Manual curation remains the gold standard but is resource-intensive and not scalable. We introduce DIAMOND, an LLM-driven agent for context-aware baseball highlight summarization that integrates structured sports analytics with natural language reasoning. DIAMOND leverages sabermetric features -- Win Expectancy, WPA, and Leverage Index -- to quantify play importance, while an LLM module enhances selection based on contextual narrative value. This hybrid approach ensures both quantitative rigor and qualitative richness, surpassing the limitations of purely statistical or vision-based systems. Evaluated on five diverse Korean Baseball Organization League games, DIAMOND improves F1-score from 42.9% (WPA-only) to 84.8%, outperforming both commercial and statistical baselines. Though limited in scale, our results highlight the potential of modular, interpretable agent-based frameworks for event-level summarization in sports and beyond.
Authors: Hisami Suzuki, Satoru Katsumata, Takashi Kodama, Tetsuro Takahashi, Kouta Nakayama, Satoshi Sekine
Abstract: In this paper we present AnswerCarefully, a dataset for promoting the safety and appropriateness of Japanese LLM outputs. The dataset consists of 1,800 pairs of questions and reference answers, where the questions require special attention in answering. It covers a wide range of risk categories established in prior English-language datasets, but the data samples are original in that they are manually created to reflect the socio-cultural context of LLM usage in Japan. We show that using this dataset for instruction to fine-tune a Japanese LLM led to improved output safety without compromising the utility of general responses. We also report the results of a safety evaluation of 12 Japanese LLMs using this dataset as a benchmark. Finally, we describe the latest update on the dataset which provides English translations and annotations of the questions, aimed at facilitating the derivation of similar datasets in different languages and regions.
Authors: Ukyo Honda, Tatsushi Oka
Abstract: In-context learning (ICL) has emerged as a successful paradigm for leveraging large language models (LLMs). However, it often struggles to generalize beyond the distribution of the provided demonstrations. A recent advancement in enhancing robustness is ICL with explanations (X-ICL), which improves prediction reliability by guiding LLMs to understand and articulate the reasoning behind correct labels. Building on this approach, we introduce an advanced framework that extends X-ICL by systematically exploring explanations for all possible labels (X$^2$-ICL), thereby enabling more comprehensive and robust decision-making. Experimental results on multiple natural language understanding datasets validate the effectiveness of X$^2$-ICL, demonstrating significantly improved robustness to out-of-distribution data compared to the existing ICL approaches.
Authors: Chuanghao Ding, Jiaping Wang, Ziqing Yang, Xiaoliang Wang, Dahua Lin, Cam-Tu Nguyen, Fei Tan
Abstract: The synergistic mechanism based on Speculative Decoding (SD) has garnered considerable attention as a simple yet effective approach for accelerating the inference of large language models (LLMs). Nonetheless, the high rejection rates require repeated LLMs calls to validate draft tokens, undermining the overall efficiency gain of SD. In this work, we revisit existing verification mechanisms and propose a novel synergetic mechanism Consultant Decoding (CD). Unlike SD, which relies on a metric derived from importance sampling for verification, CD verifies candidate drafts using token-level likelihoods computed solely by the LLM. CD achieves up to a 2.5-fold increase in inference speed compared to the target model, while maintaining comparable generation quality (around 100% of the target model's performance). Interestingly, this is achieved by combining models whose parameter sizes differ by two orders of magnitude. In addition, CD reduces the call frequency of the large target model to below 10%, particularly in more demanding tasks. CD's performance was even found to surpass that of the large target model, which theoretically represents the upper bound for speculative decoding.
Authors: Yilin Xiao, Junnan Dong, Chuang Zhou, Su Dong, Qianwen Zhang, Di Yin, Xing Sun, Xiao Huang
Abstract: Graph Retrieval Augmented Generation (GraphRAG) has garnered increasing recognition for its potential to enhance large language models (LLMs) by structurally organizing domain-specific corpora and facilitating complex reasoning. However, current evaluations of GraphRAG models predominantly rely on traditional question-answering datasets. Their limited scope in questions and evaluation metrics fails to comprehensively assess the reasoning capacity improvements enabled by GraphRAG models. To address this gap, we introduce GraphRAG-Bench, a large-scale, domain-specific benchmark designed to rigorously evaluate GraphRAG models. Our benchmark offers three key superiorities: \((i)\) Challenging question design. Featuring college-level, domain-specific questions that demand multi-hop reasoning, the benchmark ensures that simple content retrieval is insufficient for problem-solving. For example, some questions require mathematical reasoning or programming. \((ii)\) Diverse task coverage. The dataset includes a broad spectrum of reasoning tasks, multiple-choice, true/false, multi-select, open-ended, and fill-in-the-blank. It spans 16 disciplines in twenty core textbooks. \((iii)\) Holistic evaluation framework. GraphRAG-Bench provides comprehensive assessment across the entire GraphRAG pipeline, including graph construction, knowledge retrieval, and answer generation. Beyond final-answer correctness, it evaluates the logical coherence of the reasoning process. By applying nine contemporary GraphRAG methods to GraphRAG-Bench, we demonstrate its utility in quantifying how graph-based structuring improves model reasoning capabilities. Our analysis reveals critical insights about graph architectures, retrieval efficacy, and reasoning capabilities, offering actionable guidance for the research community.
Authors: Zhengyuan Liu, Geyu Lin, Hui Li Tan, Huayun Zhang, Yanfeng Lu, Xiaoxue Gao, Stella Xin Yin, He Sun, Hock Huan Goh, Lung Hsiang Wong, Nancy F. Chen
Abstract: The integration of generative artificial intelligence into educational applications has enhanced personalized and interactive learning experiences, and it shows strong potential to promote young learners language acquisition. However, it is still challenging to ensure consistent and robust performance across different languages and cultural contexts, and kids-friendly design requires simplified instructions, engaging interactions, and age-appropriate scaffolding to maintain motivation and optimize learning outcomes. In this work, we introduce SingaKids, a dialogic tutor designed to facilitate language learning through picture description tasks. Our system integrates dense image captioning, multilingual dialogic interaction, speech understanding, and engaging speech generation to create an immersive learning environment in four languages: English, Mandarin, Malay, and Tamil. We further improve the system through multilingual pre-training, task-specific tuning, and scaffolding optimization. Empirical studies with elementary school students demonstrate that SingaKids provides effective dialogic teaching, benefiting learners at different performance levels.
Authors: Tairan Liu
Abstract: Textbooks play a critical role in shaping children's understanding of the world. While previous studies have identified gender inequality in individual countries' textbooks, few have examined the issue cross-culturally. This study applies natural language processing methods to quantify gender inequality in English textbooks from 22 countries across 7 cultural spheres. Metrics include character count, firstness (which gender is mentioned first), and TF-IDF word associations by gender. The analysis also identifies gender patterns in proper names appearing in TF-IDF word lists, tests whether large language models can distinguish between gendered word lists, and uses GloVe embeddings to examine how closely keywords associate with each gender. Results show consistent overrepresentation of male characters in terms of count, firstness, and named entities. All regions exhibit gender inequality, with the Latin cultural sphere showing the least disparity.
Authors: Maryam Berijanian, Kuldeep Singh, Amin Sehati
Abstract: Entity relationship classification remains a challenging task in information extraction, especially in scenarios with limited labeled data and complex relational structures. In this study, we conduct a comparative analysis of three distinct AI agent architectures designed to perform relation classification using large language models (LLMs). The agentic architectures explored include (1) reflective self-evaluation, (2) hierarchical task decomposition, and (3) a novel multi-agent dynamic example generation mechanism, each leveraging different modes of reasoning and prompt adaptation. In particular, our dynamic example generation approach introduces real-time cooperative and adversarial prompting. We systematically compare their performance across multiple domains and model backends. Our experiments demonstrate that multi-agent coordination consistently outperforms standard few-shot prompting and approaches the performance of fine-tuned models. These findings offer practical guidance for the design of modular, generalizable LLM-based systems for structured relation extraction. The source codes and dataset are available at https://github.com/maryambrj/ALIEN.git.
Authors: Mahammed Kamruzzaman, Abdullah Al Monsur, Gene Louis Kim, Anshuman Chhabra
Abstract: Emotions are a fundamental facet of human experience, varying across individuals, cultural contexts, and nationalities. Given the recent success of Large Language Models (LLMs) as role-playing agents, we examine whether LLMs exhibit emotional stereotypes when assigned nationality-specific personas. Specifically, we investigate how different countries are represented in pre-trained LLMs through emotion attributions and whether these attributions align with cultural norms. Our analysis reveals significant nationality-based differences, with emotions such as shame, fear, and joy being disproportionately assigned across regions. Furthermore, we observe notable misalignment between LLM-generated and human emotional responses, particularly for negative emotions, highlighting the presence of reductive and potentially biased stereotypes in LLM outputs.
Authors: Utsav Maskey, Mark Dras, Usman Naseem
Abstract: This paper presents a systematic evaluation of Large Language Models' (LLMs) behavior on long-tail distributed (encrypted) texts and their safety implications. We introduce a two-dimensional framework for assessing LLM safety: (1) instruction refusal-the ability to reject harmful obfuscated instructions, and (2) generation safety-the suppression of generating harmful responses. Through comprehensive experiments, we demonstrate that models that possess capabilities to decrypt ciphers may be susceptible to mismatched-generalization attacks: their safety mechanisms fail on at least one safety dimension, leading to unsafe responses or over-refusal. Based on these findings, we evaluate a number of pre-LLM and post-LLM safeguards and discuss their strengths and limitations. This work contributes to understanding the safety of LLM in long-tail text scenarios and provides directions for developing robust safety mechanisms.
Authors: Bo Peng, Zhiheng Wang, Heyang Gong, Chaochao Lu
Abstract: In modern dialogue systems, the ability to implicitly infer user backgrounds from conversations and leverage this information for personalized assistance is crucial. However, the scarcity of high-quality data remains a fundamental challenge to evaluating and improving this capability. Traditional dataset construction methods are labor-intensive, resource-demanding, and raise privacy concerns. To address these issues, we propose a novel approach for automatic synthetic data generation and introduce the Implicit Personalized Dialogue (IP-Dialog) benchmark along with a training dataset, covering 10 tasks and 12 user attribute types. Additionally, we develop a systematic evaluation framework with four metrics to assess both attribute awareness and reasoning capabilities. We further propose five causal graphs to elucidate models' reasoning pathways during implicit personalization. Extensive experiments yield insightful observations and prove the reliability of our dataset.
Authors: Zhaorui Yang, Bo Pan, Han Wang, Yiyao Wang, Xingyu Liu, Minfeng Zhu, Bo Zhang, Wei Chen
Abstract: Visualizations play a crucial part in effective communication of concepts and information. Recent advances in reasoning and retrieval augmented generation have enabled Large Language Models (LLMs) to perform deep research and generate comprehensive reports. Despite its progress, existing deep research frameworks primarily focus on generating text-only content, leaving the automated generation of interleaved texts and visualizations underexplored. This novel task poses key challenges in designing informative visualizations and effectively integrating them with text reports. To address these challenges, we propose Formal Description of Visualization (FDV), a structured textual representation of charts that enables LLMs to learn from and generate diverse, high-quality visualizations. Building on this representation, we introduce Multimodal DeepResearcher, an agentic framework that decomposes the task into four stages: (1) researching, (2) exemplar report textualization, (3) planning, and (4) multimodal report generation. For the evaluation of generated multimodal reports, we develop MultimodalReportBench, which contains 100 diverse topics served as inputs along with 5 dedicated metrics. Extensive experiments across models and evaluation methods demonstrate the effectiveness of Multimodal DeepResearcher. Notably, utilizing the same Claude 3.7 Sonnet model, Multimodal DeepResearcher achieves an 82\% overall win rate over the baseline method.
Authors: Yupeng Qi, Ziyu Lyu, Min Yang, Yanlin Wang, Lu Bai, Lixin Cui
Abstract: As large language models (LLMs) are increasingly applied across various domains, enhancing safety while maintaining the helpfulness of LLMs has become a critical challenge. Recent studies solve this problem through safety-constrained online preference optimization or safety-constrained offline preference optimization. However, the safety-constrained online methods often suffer from excessive safety, which might reduce helpfulness, while the safety-constrained offline methods perform poorly in adaptively balancing safety and helpfulness. To address these limitations, we propose MidPO, a \textbf{\underline{Mi}}xture of Experts (MoE) framework for safety-helpfulness \textbf{\underline{d}}ual \textbf{\underline{P}}reference \textbf{\underline{O}}ptimization. Firstly, MidPO devises single-preference enhanced direct preference optimization approach to transform the base model into two independent experts, termed safety and helpfulness experts, and fine-tunes the two independent experts for optimal safety or helpfulness performance. Secondly, to achieve an effective balance between safety and helpfulness, MidPO incorporates the two experts into the MoE framework and designs a dynamic routing mechanism to allocate contributions from each expert adaptively. We conduct quantitative and qualitative experiments on three popular datasets to demonstrate the proposed MidPO significantly outperforms state-of-the-art approaches in both safety and helpfulness. The code and models will be released.
Authors: Chunkit Chan, Yauwai Yim, Hongchuan Zeng, Zhiying Zou, Xinyuan Cheng, Zhifan Sun, Zheye Deng, Kawai Chung, Yuzhuo Ao, Yixiang Fan, Cheng Jiayang, Ercong Nie, Ginny Y. Wong, Helmut Schmid, Hinrich Sch\"utze, Simon See, Yangqiu Song
Abstract: Theory of Mind (ToM), the ability to infer mental states in others, is pivotal for human social cognition. Existing evaluations of ToM in LLMs are largely limited to English, neglecting the linguistic diversity that shapes human cognition. This limitation raises a critical question: can LLMs exhibit Multilingual Theory of Mind, which is the capacity to reason about mental states across diverse linguistic contexts? To address this gap, we present XToM, a rigorously validated multilingual benchmark that evaluates ToM across five languages and incorporates diverse, contextually rich task scenarios. Using XToM, we systematically evaluate LLMs (e.g., DeepSeek R1), revealing a pronounced dissonance: while models excel in multilingual language understanding, their ToM performance varies across languages. Our findings expose limitations in LLMs' ability to replicate human-like mentalizing across linguistic contexts.
Authors: Zijian Li, Xiaocheng Feng, Huixin Liu, Yichong Huang, Ting Liu, Bing Qin
Abstract: With the development of large language models, fine-tuning has emerged as an effective method to enhance performance in specific scenarios by injecting domain-specific knowledge. In this context, model merging techniques provide a solution for fusing knowledge from multiple fine-tuning models by combining their parameters. However, traditional methods often encounter task interference when merging full fine-tuning models, and this problem becomes even more evident in parameter-efficient fine-tuning scenarios. In this paper, we introduce an improvement to the RegMean method, which indirectly leverages the training data to approximate the outputs of the linear layers before and after merging. We propose an adaptive merging method called FroM, which directly measures the model parameters using the Frobenius norm, without any training data. By introducing an additional hyperparameter for control, FroM outperforms baseline methods across various fine-tuning scenarios, alleviating the task interference problem.
Authors: Yifan Duan, Yihong Tang, Kehai Chen, Liqiang Nie, Min Zhang
Abstract: High-quality prompts are crucial for eliciting outstanding performance from large language models (LLMs) on complex tasks. Existing research has explored model-driven strategies for prompt optimization. However, these methods often suffer from high computational overhead or require strong optimization capabilities from the model itself, which limits their broad applicability.To address these challenges, we propose ORPP (Optimized Role-Playing Prompt),a framework that enhances model performance by optimizing and generating role-playing prompts. The core idea of ORPP is to confine the prompt search space to role-playing scenarios, thereby fully activating the model's intrinsic capabilities through carefully crafted, high-quality role-playing prompts. Specifically, ORPP first performs iterative optimization on a small subset of training samples to generate high-quality role-playing prompts. Then, leveraging the model's few-shot learning capability, it transfers the optimization experience to efficiently generate suitable prompts for the remaining samples.Our experimental results show that ORPP not only matches but in most cases surpasses existing mainstream prompt optimization methods in terms of performance. Notably, ORPP demonstrates superior "plug-and-play" capability. In most cases, it can be integrated with various other prompt methods and further enhance their effectiveness.
Authors: Inderjeet Nair, Lu Wang
Abstract: Evaluations of LLMs' ethical risks and value inclinations often rely on short-form surveys and psychometric tests, yet real-world use involves long-form, open-ended responses -- leaving value-related risks and preferences in practical settings largely underexplored. In this work, we ask: Do value preferences inferred from short-form tests align with those expressed in long-form outputs? To address this question, we compare value preferences elicited from short-form reactions and long-form responses, varying the number of arguments in the latter to capture users' differing verbosity preferences. Analyzing five LLMs (llama3-8b, gemma2-9b, mistral-7b, qwen2-7b, and olmo-7b), we find (1) a weak correlation between value preferences inferred from short-form and long-form responses across varying argument counts, and (2) similarly weak correlation between preferences derived from any two distinct long-form generation settings. (3) Alignment yields only modest gains in the consistency of value expression. Further, we examine how long-form generation attributes relate to value preferences, finding that argument specificity negatively correlates with preference strength, while representation across scenarios shows a positive correlation. Our findings underscore the need for more robust methods to ensure consistent value expression across diverse applications.
Authors: Sina Bagheri Nezhad, Ameeta Agrawal
Abstract: Large language models (LLMs) often struggle to perform multi-target reasoning in long-context scenarios where relevant information is scattered across extensive documents. To address this challenge, we introduce NeuroSymbolic Augmented Reasoning (NSAR), which combines the benefits of neural and symbolic reasoning during inference. NSAR explicitly extracts symbolic facts from text and generates executable Python code to handle complex reasoning steps. Through extensive experiments across seven languages and diverse context lengths, we demonstrate that NSAR significantly outperforms both a vanilla RAG baseline and advanced prompting strategies in accurately identifying and synthesizing multiple pieces of information. Our results highlight the effectiveness of combining explicit symbolic operations with neural inference for robust, interpretable, and scalable reasoning in multilingual settings.
Authors: Junzhe Zhang, Huixuan Zhang, Xinyu Hu, Li Lin, Mingqi Gao, Shi Qiu, Xiaojun Wan
Abstract: Evaluation is important for multimodal generation tasks. With the rapid progress of MLLMs, there is growing interest in applying MLLMs to build general evaluation systems. However, existing work overlooks two aspects: (1) the development of evaluation capabilities for text-to-image (T2I) generation task, and (2) the incorporation of large-scale human evaluation data. In this paper, we introduce Minos-Corpus, a large-scale multimodal evaluation dataset that combines evaluation data from both human and GPT. The corpus contains evaluation data across both image-to-text(I2T) and T2I generation tasks. Based on this corpus, we propose Data Selection and Balance, Mix-SFT training methods, and apply DPO to develop Minos, a multimodal evaluation model built upon a 7B backbone. Minos achieves state-of-the-art (SoTA) performance among all open-source evaluation models of similar scale on the average of evaluation performance on all tasks, and outperforms all open-source and closed-source models on evaluation of T2I generation task. Extensive experiments demonstrate the importance of leveraging high-quality human evaluation data and jointly training on evaluation data from both I2T and T2I generation tasks.
Authors: Yongjian Li, HaoCheng Chu, Yukun Yan, Zhenghao Liu, Shi Yu, Zheni Zeng, Ruobing Wang, Sen Song, Zhiyuan Liu, Maosong Sun
Abstract: Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to access broader knowledge sources, yet factual inconsistencies persist due to noise in retrieved documents-even with advanced retrieval methods. We demonstrate that enhancing generative models' capacity to process noisy content is equally critical for robust performance. In this paper, we present KARE-RAG (Knowledge-Aware Refinement and Enhancement for RAG), which improves knowledge utilization through three key innovations: (1) structured knowledge representations that facilitate error detection during training, (2) Dense Direct Preference Optimization (DDPO)-a refined training objective that prioritizes correction of critical errors, and (3) a contrastive data generation pipeline that maintains semantic consistency while rectifying factual inaccuracies. Experiments show our method significantly enhances standard RAG pipelines across model scales, improving both in-domain and out-of-domain task performance without compromising general capabilities. Notably, these gains are achieved with modest training data, suggesting data-efficient optimization is possible through targeted learning strategies. Our findings establish a new direction for RAG improvement: by improving how models learn to process retrieved content, we can enhance performance across diverse inference paradigms. All data and code will be publicly available on Github.
Authors: Jie Zhu, Junhui Li, Yalong Wen, Xiandong Li, Lifan Guo, Feng Chen
Abstract: Recent breakthroughs in large language models (LLMs) have led to the development of new benchmarks for evaluating their performance in the financial domain. However, current financial benchmarks often rely on news articles, earnings reports, or announcements, making it challenging to capture the real-world dynamics of financial meetings. To address this gap, we propose a novel benchmark called $\texttt{M$^3$FinMeeting}$, which is a multilingual, multi-sector, and multi-task dataset designed for financial meeting understanding. First, $\texttt{M$^3$FinMeeting}$ supports English, Chinese, and Japanese, enhancing comprehension of financial discussions in diverse linguistic contexts. Second, it encompasses various industry sectors defined by the Global Industry Classification Standard (GICS), ensuring that the benchmark spans a broad range of financial activities. Finally, $\texttt{M$^3$FinMeeting}$ includes three tasks: summarization, question-answer (QA) pair extraction, and question answering, facilitating a more realistic and comprehensive evaluation of understanding. Experimental results with seven popular LLMs reveal that even the most advanced long-context models have significant room for improvement, demonstrating the effectiveness of $\texttt{M$^3$FinMeeting}$ as a benchmark for assessing LLMs' financial meeting comprehension skills.
Authors: Zhuohan Xie, Dhruv Sahnan, Debopriyo Banerjee, Georgi Georgiev, Rushil Thareja, Hachem Madmoun, Jinyan Su, Aaryamonvikram Singh, Yuxia Wang, Rui Xing, Fajri Koto, Haonan Li, Ivan Koychev, Tanmoy Chakraborty, Salem Lahlou, Veselin Stoyanov, Preslav Nakov
Abstract: Multi-step symbolic reasoning is critical for advancing downstream performance on financial tasks. Yet, benchmarks for systematically evaluating this capability are lacking. Existing datasets like FinQA and ConvFinQA supervise only final numerical answers, without assessing intermediate reasoning steps. To address this, we introduce FinChain, the first symbolic benchmark designed for verifiable Chain-of- Thought (CoT) financial reasoning. Spanning 54 topics across 12 financial domains, Fin- Chain offers five parameterized templates per topic, each varying in reasoning complexity and domain expertise required. Each dataset instance includes an executable Python trace, enabling automatic generation of extensive training data and easy adaptation to other domains. We also introduce ChainEval, a new metric for automatic evaluation of both final answers and intermediate reasoning. Benchmarking 30 LLMs on our dataset, we find that even state-of-the-art models have considerable room for improvement in multi-step financial reasoning. All templates and evaluation metrics for FinChain are available at https: //github.com/mbzuai-nlp/finchain.
Authors: Sohan Patnaik, Milan Aggarwal, Sumit Bhatia, Balaji Krishnamurthy
Abstract: LLMssuch as GPT-4 have shown a remarkable ability to solve complex questions by generating step-by-step rationales. Prior works have utilized this capability to improve smaller and cheaper LMs (say, with 7B parameters). However, various practical constraints, such as copyright and legal issues, owing to lack of transparency in the pre-training data of large (often closed) models, prevent their use in commercial settings. Little focus has been given to improving the innate reasoning ability of smaller models without distilling information from larger LLMs. To address this, we propose COLLATE, a trainable framework that tunes a (small) LLM to generate those outputs from a pool of diverse rationales that selectively improves the downstream task. COLLATE enforces multiple instances of the same LLM to exhibit distinct behavior and employs them to generate rationales to obtain diverse outputs. The LLM is then tuned via preference optimization to choose the candidate rationale which maximizes the likelihood of ground-truth answer. COLLATE outperforms several trainable and prompting baselines on 5 datasets across 3 domains: maths problem solving, natural language inference, and commonsense reasoning. We show the eff icacy of COLLATE on LLMs from different model families across varying parameter scales (1B to 8B) and demonstrate the benefit of multiple rationale providers guided by the end task through ablations. Code is released here (https://github.com/Sohanpatnaik106/collate).
Authors: Yingying Zhuang, Aman Gupta, Anurag Beniwal
Abstract: Multilingual information retrieval has emerged as powerful tools for expanding knowledge sharing across languages. On the other hand, resources on high quality knowledge base are often scarce and in limited languages, therefore an effective embedding model to transform sentences from different languages into a feature vector space same as the knowledge base language becomes the key ingredient for cross language knowledge sharing, especially to transfer knowledge available in high-resource languages to low-resource ones. In this paper we propose a novel strategy to fine-tune multilingual embedding models with weighted sampling for contrastive learning, enabling multilingual information retrieval with a monolingual knowledge base. We demonstrate that the weighted sampling strategy produces performance gains compared to standard ones by up to 31.03\% in MRR and up to 33.98\% in Recall@3. Additionally, our proposed methodology is language agnostic and applicable for both multilingual and code switching use cases.
Authors: Jinu Lee, Sagnik Mukherjee, Dilek Hakkani-Tur, Julia Hockenmaier
Abstract: Large reasoning models (LRMs) generate complex reasoning traces with planning, reflection, verification, and backtracking. In this work, we introduce ReasoningFlow, a unified schema for analyzing the semantic structures of these complex traces. ReasoningFlow parses traces into directed acyclic graphs, enabling the characterization of distinct reasoning patterns as subgraph structures. This human-interpretable representation offers promising applications in understanding, evaluating, and enhancing the reasoning processes of LRMs.
Authors: Maike Behrendt, Stefan Sylvius Wagner, Carina Weinmann, Marike Bormann, Mira Warne, Stefan Harmeling
Abstract: Political online participation in the form of discussing political issues and exchanging opinions among citizens is gaining importance with more and more formats being held digitally. To come to a decision, a careful discussion and consideration of opinions and a civil exchange of arguments, which is defined as the act of deliberation, is desirable. The quality of discussions and participation processes in terms of their deliberativeness highly depends on the design of platforms and processes. To facilitate online communication for both participants and initiators, machine learning methods offer a lot of potential. In this work we want to showcase which issues occur in political online discussions and how machine learning can be used to counteract these issues and enhance deliberation.
Authors: Xin Liu, Lu Wang
Abstract: Chain-of-thought (CoT) prompting enhances reasoning in large language models (LLMs) but often leads to verbose and redundant outputs, thus increasing inference cost. We hypothesize that many reasoning steps are unnecessary for producing correct answers. To investigate this, we start with a systematic study to examine what is the minimum reasoning required for a model to reach a stable decision. We find that on math reasoning tasks like math, models typically converge to their final answers after 60\% of the reasoning steps, suggesting substantial redundancy in the remaining content. Based on these insights, we propose three inference-time strategies to improve efficiency: (1) early stopping via answer consistency, (2) boosting the probability of generating end-of-reasoning signals, and (3) a supervised method that learns when to stop based on internal activations. Experiments across five benchmarks and five open-weights LLMs show that our methods significantly reduce token usage with little or no accuracy drop. In particular, on NaturalQuestions, Answer Consistency reduces tokens by over 40\% while further improving accuracy. Our work underscores the importance of cost-effective reasoning methods that operate at inference time, offering practical benefits for real-world applications.
Authors: Yang Tian, Fan Liu, Jingyuan Zhang, Victoria W., Yupeng Hu, Liqiang Nie
Abstract: Multimodal Retrieval-Augmented Generation (MMRAG) has been introduced to enhance Multimodal Large Language Models by incorporating externally retrieved multimodal knowledge, but it introduces two challenges: Parametric-Retrieved Knowledge Inconsistency (PRKI), where discrepancies between parametric and retrieved knowledge create uncertainty in determining reliability, and Visual-Textual Knowledge Inconsistency (VTKI), where misalignment between visual and textual sources disrupts entity representation. To address these challenges, we propose Cross-source knowledge \textbf{Re}conciliation for Multimodal RAG (CoRe-MMRAG), a novel end-to-end framework that effectively reconciles inconsistencies across knowledge sources. CoRe-MMRAG follows a four-stage pipeline: it first generates an internal response from parametric knowledge, then selects the most relevant multimodal evidence via joint similarity assessment, generates an external response, and finally integrates both to produce a reliable answer. Additionally, a specialized training paradigm enhances knowledge source discrimination, multimodal integration, and unified answer generation. Experiments on KB-VQA benchmarks show that CoRe-MMRAG achieves substantial improvements over baseline methods, achieving 5.6% and 9.3% performance gains on InfoSeek and Encyclopedic-VQA, respectively.
Authors: Yirao Zhao, Guizhen Chen, Kenji Kawaguchi, Lidong Bing, Wenxuan Zhang
Abstract: Large language models (LLMs) have revolutionized natural language processing, yet their substantial model sizes often require substantial computational resources. To preserve computing resources and accelerate inference speed, it is crucial to prune redundant parameters, especially for experienced users who often need compact expert models tailored to specific downstream scenarios. However, most existing pruning methods focus on preserving the model's general capabilities, often requiring extensive post-training or suffering from degraded performance due to coarse-grained pruning. In this work, we design a $\underline{Cus}$tom $\underline{Prun}$ing method ($\texttt{Cus-Prun}$) to prune a large general model into a smaller lightweight expert model, which is positioned along the "language", "domain" and "task" dimensions. By identifying and pruning irrelevant neurons of each dimension, $\texttt{Cus-Prun}$ creates expert models without any post-training. Our experiments demonstrate that $\texttt{Cus-Prun}$ consistently outperforms other methods, achieving minimal loss in both expert and general capabilities across various models from different model families and sizes.
Authors: Muhammad Falensi Azmi, Muhammad Dehan Al Kautsar, Alfan Farizki Wicaksono, Fajri Koto
Abstract: Although region-specific large language models (LLMs) are increasingly developed, their safety remains underexplored, particularly in culturally diverse settings like Indonesia, where sensitivity to local norms is essential and highly valued by the community. In this work, we present IndoSafety, the first high-quality, human-verified safety evaluation dataset tailored for the Indonesian context, covering five language varieties: formal and colloquial Indonesian, along with three major local languages: Javanese, Sundanese, and Minangkabau. IndoSafety is constructed by extending prior safety frameworks to develop a taxonomy that captures Indonesia's sociocultural context. We find that existing Indonesian-centric LLMs often generate unsafe outputs, particularly in colloquial and local language settings, while fine-tuning on IndoSafety significantly improves safety while preserving task performance. Our work highlights the critical need for culturally grounded safety evaluation and provides a concrete step toward responsible LLM deployment in multilingual settings. Warning: This paper contains example data that may be offensive, harmful, or biased.
Authors: Sarenne Wallbridge, Christoph Minixhofer, Catherine Lai, Peter Bell
Abstract: People exploit the predictability of lexical structures during text comprehension. Though predictable structure is also present in speech, the degree to which prosody, e.g. intonation, tempo, and loudness, contributes to such structure independently of the lexical content is unclear. This study leverages self-supervised learning (SSL) to examine the temporal granularity of structures in the acoustic correlates of prosody. Representations from our proposed Masked Prosody Model can predict perceptual labels dependent on local information, such as word boundaries, but provide the most value for labels involving longer-term structures, like emotion recognition. Probing experiments across various perceptual labels show strong relative gains over untransformed pitch, energy, and voice activity features. Our results reveal the importance of SSL training objective timescale and highlight the value of complex SSL-encoded structures compared to more constrained classical structures.
Authors: Maria Levchenko
Abstract: This paper addresses the challenge of Named Entity Recognition (NER) for person names within the specialized domain of Russian news texts concerning cultural events. The study utilizes the unique SPbLitGuide dataset, a collection of event announcements from Saint Petersburg spanning 1999 to 2019. A comparative evaluation of diverse NER models is presented, encompassing established transformer-based architectures such as DeepPavlov, RoBERTa, and SpaCy, alongside recent Large Language Models (LLMs) including GPT-3.5, GPT-4, and GPT-4o. Key findings highlight the superior performance of GPT-4o when provided with specific prompting for JSON output, achieving an F1 score of 0.93. Furthermore, GPT-4 demonstrated the highest precision at 0.99. The research contributes to a deeper understanding of current NER model capabilities and limitations when applied to morphologically rich languages like Russian within the cultural heritage domain, offering insights for researchers and practitioners. Follow-up evaluation with GPT-4.1 (April 2025) achieves F1=0.94 for both simple and structured prompts, demonstrating rapid progress across model families and simplified deployment requirements.
Authors: Minh Duc Bui, Kyung Eun Park, Goran Glava\v{s}, Fabian David Schmidt, Katharina von der Wense
Abstract: Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.
Authors: Zhi-Yuan Chen, Hao Wang, Xinyu Zhang, Enrui Hu, Yankai Lin
Abstract: Recent studies show that large language models (LLMs) exhibit self-preference bias when serving as judges, meaning they tend to favor their own responses over those generated by other models. Existing methods typically measure this bias by calculating the difference between the scores a judge model assigns to its own responses and those it assigns to responses from other models. However, this approach conflates self-preference bias with response quality, as higher-quality responses from the judge model may also lead to positive score differences, even in the absence of bias. To address this issue, we introduce gold judgments as proxies for the actual quality of responses and propose the DBG score, which measures self-preference bias as the difference between the scores assigned by the judge model to its own responses and the corresponding gold judgments. Since gold judgments reflect true response quality, the DBG score mitigates the confounding effect of response quality on bias measurement. Using the DBG score, we conduct comprehensive experiments to assess self-preference bias across LLMs of varying versions, sizes, and reasoning abilities. Additionally, we investigate two factors that influence and help alleviate self-preference bias: response text style and the post-training data of judge models. Finally, we explore potential underlying mechanisms of self-preference bias from an attention-based perspective. Our code and data are available at https://github.com/zhiyuanc2001/self-preference.
Authors: Fan Gao, Dongyuan Li, Ding Xia, Fei Mi, Yasheng Wang, Lifeng Shang, Baojun Wang
Abstract: Chinese essay writing and its evaluation are critical in educational contexts, yet the capabilities of Large Language Models (LLMs) in this domain remain largely underexplored. Existing benchmarks often rely on coarse-grained text quality metrics, largely overlooking the structural and rhetorical complexities of Chinese essays, particularly across diverse genres. To address this gap, we propose \benchName, a multi-genre benchmark specifically designed for Chinese essay writing across four major genres: Argumentative, Narrative, Descriptive, and Expository. We curate and refine a total of 728 real-world prompts to ensure authenticity and meticulously categorize them into the \textit{Open-Ended} and \textit{Constrained} sets to capture diverse writing scenarios. To reliably evaluate generated essays, we develop a fine-grained, genre-specific scoring framework that hierarchically aggregates scores. We further validate our evaluation protocol through a comprehensive human agreement study. Finally, we benchmark 15 large-sized LLMs, analyzing their strengths and limitations across genres and instruction types. With \benchName, we aim to advance LLM-based Chinese essay evaluation and inspire future research on improving essay generation in educational settings.
Authors: \"Omer Tarik \"Ozyilmaz, Matt Coler, Matias Valdenegro-Toro
Abstract: Although commercial Arabic automatic speech recognition (ASR) systems support Modern Standard Arabic (MSA), they struggle with dialectal speech. We investigate the effect of fine-tuning OpenAI's Whisper on five major Arabic dialects (Gulf, Levantine, Iraqi, Egyptian, Maghrebi) using Mozilla Common Voice for MSA and the MASC dataset for dialectal speech. We evaluate MSA training size effects, benefits of pre-training on MSA data, and dialect-specific versus dialect-pooled models. We find that small amounts of MSA fine-tuning data yield substantial improvements for smaller models, matching larger non-fine-tuned models. While MSA pre-training shows minimal benefit, suggesting limited shared features between MSA and dialects, our dialect-pooled models perform comparably to dialect-specific ones. This indicates that pooling dialectal data, when properly balanced, can help address data scarcity in low-resource ASR without significant performance loss.
Authors: Manon Reusens, Bart Baesens, David Jurgens
Abstract: Personalized Large Language Models (LLMs) are increasingly used in diverse applications, where they are assigned a specific persona - such as a happy high school teacher - to guide their responses. While prior research has examined how well LLMs adhere to predefined personas in writing style, a comprehensive analysis of consistency across different personas and task types is lacking. In this paper, we introduce a new standardized framework to analyze consistency in persona-assigned LLMs. We define consistency as the extent to which a model maintains coherent responses when assigned the same persona across different tasks and runs. Our framework evaluates personas across four different categories (happiness, occupation, personality, and political stance) spanning multiple task dimensions (survey writing, essay generation, social media post generation, single turn, and multi-turn conversations). Our findings reveal that consistency is influenced by multiple factors, including the assigned persona, stereotypes, and model design choices. Consistency also varies across tasks, increasing with more structured tasks and additional context. All code is available on GitHub.
Authors: Shihan Dou, Ming Zhang, Chenhao Huang, Jiayi Chen, Feng Chen, Shichun Liu, Yan Liu, Chenxiao Liu, Cheng Zhong, Zongzhang Zhang, Tao Gui, Chao Xin, Wei Chengzhi, Lin Yan, Qi Zhang, Xuanjing Huang
Abstract: We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available at the GitHub repository.
Authors: Zhong-Zhi Li, Xiao Liang, Zihao Tang, Lei Ji, Peijie Wang, Haotian Xu, Xing W, Haizhen Huang, Weiwei Deng, Ying Nian Wu, Yeyun Gong, Zhijiang Guo, Xiao Liu, Fei Yin, Cheng-Lin Liu
Abstract: Large Language Models (LLMs) have recently achieved remarkable progress by leveraging Reinforcement Learning and extended Chain-of-Thought (CoT) techniques. However, the challenge of performing efficient language reasoning--especially during inference with extremely long outputs--has drawn increasing attention from the research community. In this work, we propose a dynamic ratio-based training pipeline that does not rely on sophisticated data annotations or interpolation between multiple models. We continuously balance the weights between the model's System-1 and System-2 data to eliminate redundant reasoning processes while preserving the model's reasoning capability. We validate our approach across models on DeepSeek-R1-Distill-7B and DeepSeek-R1-Distill-14B and on a diverse set of benchmarks with varying difficulty levels. Our method significantly reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning. Our code and data will be available soon.
Authors: Zhengdong Lu, Weikai Lu, Yiling Tao, Yun Dai, ZiXuan Chen, Huiping Zhuang, Cen Chen, Hao Peng, Ziqian Zeng
Abstract: Despite significant advances in Large Language Models (LLMs), planning tasks still present challenges for LLM-based agents. Existing planning methods face two key limitations: heavy constraints and cascading errors. To address these limitations, we propose a novel parallel planning paradigm, which Decomposes, Plans for subtasks in Parallel, and Merges subplans into a final plan (DPPM). Specifically, DPPM decomposes the complex task based on constraints into subtasks, generates the subplan for each subtask in parallel, and merges them into a global plan. In addition, our approach incorporates a verification and refinement module, enabling error correction and conflict resolution. Experimental results demonstrate that DPPM significantly outperforms existing methods in travel planning tasks.
Authors: Liang Yue, Yihong Tang, Kehai Chen, Jie Liu, Min Zhang
Abstract: Instruction fine-tuning is crucial in NLP tasks, enhancing pretrained models' instruction-following capabilities and task-specific performance. However, obtaining high-quality fine-tuning data for large models is challenging due to data collection difficulties and high production costs. To address this, we propose MASTER, a novel data augmentation method that enriches original data through interactions among multiple agents with varying cognitive levels. We simulate three pedagogically grounded teaching scenarios, leveraging multi-agent conversations to generate high-quality teacher-student interaction data. Utilizing MASTER, we construct BOOST-QA, a fine-tuning dataset augmented from existing datasets like Orca-Math-200k, ProcQA, and OpenHermes2.5. Experiments show that models fine-tuned with BOOST-QA perform excellently across multiple benchmarks, demonstrating strong multitask generalization. Notably, MASTER significantly improves models' reasoning abilities in complex tasks, providing valuable insights for future research.
Authors: Masaki Sakata, Sho Yokoi, Benjamin Heinzerling, Takumi Ito, Kentaro Inui
Abstract: We analyze the extent to which internal representations of language models (LMs) identify and distinguish mentions of named entities, focusing on the many-to-many correspondence between entities and their mentions. We first formulate two problems of entity mentions -- ambiguity and variability -- and propose a framework analogous to clustering quality metrics. Specifically, we quantify through cluster analysis of LM internal representations the extent to which mentions of the same entity cluster together and mentions of different entities remain separated. Our experiments examine five Transformer-based autoregressive models, showing that they effectively identify and distinguish entities with metrics analogous to precision and recall ranging from 0.66 to 0.9. Further analysis reveals that entity-related information is compactly represented in a low-dimensional linear subspace at early LM layers. Additionally, we clarify how the characteristics of entity representations influence word prediction performance. These findings are interpreted through the lens of isomorphism between LM representations and entity-centric knowledge structures in the real world, providing insights into how LMs internally organize and use entity information.
Authors: Qihang Yan, Xinyu Zhang, Luming Guo, Qi Zhang, Feifan Liu
Abstract: Large Language Models (LLMs) struggle with accuracy, domain-specific reasoning, and interpretability in vertical domains. Traditional preference alignment methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) often overlook the underlying knowledge sources and reasoning logic. This paper introduces RACE-Align (Retrieval-Augmented and Chain-of-Thought Enhanced Alignment), a novel framework designed to address these limitations. RACE-Align systematically constructs a binary preference dataset incorporating external knowledge support and explicit Chain-of-Thought (CoT) reasoning, then aligns LLMs using the DPO algorithm. The core innovation lies in its preference data construction strategy: it integrates AI-driven retrieval for factual grounding, enhancing knowledgeability and accuracy, and emphasizes the optimization of domain-specific CoT, treating the reasoning process itself as a key preference dimension. A multi-stage, AI-driven refinement pipeline cost-effectively generates these preference pairs. Experimental validation in Traditional Chinese Medicine (TCM) using Qwen3-1.7B as the base model demonstrates that RACE-Align significantly outperforms the original base model and a model fine-tuned only with Supervised Fine-Tuning (SFT). Improvements were observed across multiple dimensions, including answer accuracy, information richness, application of TCM thinking patterns, logicality and depth of reasoning, and interpretability. These findings suggest RACE-Align offers an effective pathway to enhance LLMs' knowledge application, reasoning reliability, and process transparency in complex vertical domains.
Authors: Ama\c{c} Herda\u{g}delen, Marco Baroni
Abstract: We extracted gender-specific actions from text corpora and Twitter, and compared them to stereotypical expectations of people. We used Open Mind Common Sense (OMCS), a commonsense knowledge repository, to focus on actions that are pertinent to common sense and daily life of humans. We use the gender information of Twitter users and Web-corpus-based pronoun/name gender heuristics to compute the gender bias of the actions. With high recall, we obtained a Spearman correlation of 0.47 between corpus-based predictions and a human gold standard, and an area under the ROC curve of 0.76 when predicting the polarity of the gold standard. We conclude that it is feasible to use natural text (and a Twitter-derived corpus in particular) in order to augment commonsense repositories with the stereotypical gender expectations of actions. We also present a dataset of 441 commonsense actions with human judges' ratings on whether the action is typically/slightly masculine/feminine (or neutral), and another larger dataset of 21,442 actions automatically rated by the methods we investigate in this study.
Authors: Aisha Alansari, Hamzah Luqman
Abstract: The rapid growth of social media has amplified the spread of offensive, violent, and vulgar speech, which poses serious societal and cybersecurity concerns. Detecting such content in Arabic text is particularly complex due to limited labeled data, dialectal variations, and the language's inherent complexity. This paper proposes a novel framework that integrates multi-task learning (MTL) with active learning to enhance offensive speech detection in Arabic social media text. By jointly training on two auxiliary tasks, violent and vulgar speech, the model leverages shared representations to improve the detection accuracy of the offensive speech. Our approach dynamically adjusts task weights during training to balance the contribution of each task and optimize performance. To address the scarcity of labeled data, we employ an active learning strategy through several uncertainty sampling techniques to iteratively select the most informative samples for model training. We also introduce weighted emoji handling to better capture semantic cues. Experimental results on the OSACT2022 dataset show that the proposed framework achieves a state-of-the-art macro F1-score of 85.42%, outperforming existing methods while using significantly fewer fine-tuning samples. The findings of this study highlight the potential of integrating MTL with active learning for efficient and accurate offensive language detection in resource-constrained settings.
Authors: Stefano Bann\`o, Kate Knill, Mark Gales
Abstract: Vocabulary use is a fundamental aspect of second language (L2) proficiency. To date, its assessment by automated systems has typically examined the context-independent, or part-of-speech (PoS) related use of words. This paper introduces a novel approach to enable fine-grained vocabulary evaluation exploiting the precise use of words within a sentence. The scheme combines large language models (LLMs) with the English Vocabulary Profile (EVP). The EVP is a standard lexical resource that enables in-context vocabulary use to be linked with proficiency level. We evaluate the ability of LLMs to assign proficiency levels to individual words as they appear in L2 learner writing, addressing key challenges such as polysemy, contextual variation, and multi-word expressions. We compare LLMs to a PoS-based baseline. LLMs appear to exploit additional semantic information that yields improved performance. We also explore correlations between word-level proficiency and essay-level proficiency. Finally, the approach is applied to examine the consistency of the EVP proficiency levels. Results show that LLMs are well-suited for the task of vocabulary assessment.
Authors: Sifan Li, Yujun Cai, Yiwei Wang
Abstract: Vision-language models (VLMs) excel in semantic tasks but falter at a core human capability: detecting hidden content in optical illusions or AI-generated images through perceptual adjustments like zooming. We introduce HC-Bench, a benchmark of 112 images with hidden text, objects, and illusions, revealing that leading VLMs achieve near-zero accuracy (0-5.36%)-even with explicit prompting. Humans resolve such ambiguities instinctively, yet VLMs fail due to an overreliance on high-level semantics. Strikingly, we propose SemVink (Semantic Visual Thinking) by simply scaling images to low resolutions (32-128 pixels), which unlocks >99% accuracy by eliminating redundant visual noise. This exposes a critical architectural flaw: VLMs prioritize abstract reasoning over low-level visual operations crucial for real-world robustness. Our work urges a shift toward hybrid models integrating multi-scale processing, bridging the gap between computational vision and human cognition for applications in medical imaging, security, and beyond.
Authors: Ekaterina Grishina, Mikhail Gorbunov, Maxim Rakhuba
Abstract: Large language models (LLMs) demonstrate impressive results in natural language processing tasks but require a significant amount of computational and memory resources. Structured matrix representations are a promising way for reducing the number of parameters of these models. However, it seems unrealistic to expect that weight matrices of pretrained models can be accurately represented by structured matrices without any fine-tuning. To overcome this issue, we utilize the fact that LLM output is invariant under certain orthogonal transformations of weight matrices. This insight can be leveraged to identify transformations that significantly improve the compressibility of weights within structured classes. The proposed approach is applicable to various types of structured matrices that support efficient projection operations. Code is available at https://github.com/GrishKate/ProcrustesGPT
Authors: Yulin Dou, Jiangming Liu
Abstract: Large language models (LLMs) can effectively elicit human preferences through multi-turn dialogue. Complex tasks can be accomplished through iterative clarifying questions and final responses generated by an LLM acting as a questioner (STaR-GATE; Andukuri et al., 2024}). However, existing approaches based on self-taught reasoning struggle to identify optimal dialogue trajectories and avoid irrelevant questions to the tasks. To address this limitation, we propose TO-GATE, a novel framework that enhances question generation through trajectory optimization, which consists of two key components: a clarification resolver that generates optimal questioning trajectories, and a summarizer that ensures task-aligned final responses. The trajectory optimization enables the model to produce effective elicitation questions and summary responses tailored to specific tasks. Experimental results demonstrate that TO-GATE significantly outperforms baseline methods, achieving a 9.32% improvement on standard preference elicitation tasks.
Authors: Ludovic Moncla, H\'edi Zeghidi
Abstract: Named Entity Recognition (NER) in historical texts presents unique challenges due to non-standardized language, archaic orthography, and nested or overlapping entities. This study benchmarks a diverse set of NER approaches, ranging from classical Conditional Random Fields (CRFs) and spaCy-based models to transformer-based architectures such as CamemBERT and sequence-labeling models like Flair. Experiments are conducted on the GeoEDdA dataset, a richly annotated corpus derived from 18th-century French encyclopedias. We propose framing NER as both token-level and span-level classification to accommodate complex nested entity structures typical of historical documents. Additionally, we evaluate the emerging potential of few-shot prompting with generative language models for low-resource scenarios. Our results demonstrate that while transformer-based models achieve state-of-the-art performance, especially on nested entities, generative models offer promising alternatives when labeled data are scarce. The study highlights ongoing challenges in historical NER and suggests avenues for hybrid approaches combining symbolic and neural methods to better capture the intricacies of early modern French text.
Authors: Jintian Shao, Yiming Cheng
Abstract: Chain-of-Thought (CoT) prompting has demonstrably enhanced the performance of Large Language Models on tasks requiring multi-step inference. This success has led to widespread claims of emergent reasoning capabilities in these models. In this paper, we present a theoretical counter-perspective: Chain-of-Thought (CoT) does not elicit genuine, abstract reasoning. Instead, we argue that Chain-of-Thought functions as a powerful structural constraint that guides Large Language Models to imitate the form of reasoning. By forcing the generation of intermediate steps, Chain-of-Thought leverages the model immense capacity for sequence prediction and pattern matching, effectively constraining its output to sequences that resemble coherent thought processes. Chain-of-Thought (CoT) prompting has demonstrably enhanced the performance of Large Language Models on tasks requiring multi-step inference. This success has led to widespread claims of emergent reasoning capabilities in these models. In this paper, we present a theoretical counter-perspective: Chain-of-Thought (CoT) does not elicit genuine, abstract reasoning. Instead, we argue that Chain-of-Thought functions as a powerful structural constraint that guides Large Language Models to imitate the form of reasoning. By forcing the generation of intermediate steps, Chain-of-Thought leverages the model immense capacity for sequence prediction and pattern matching, effectively constraining its output to sequences that resemble coherent thought processes.
Authors: Verena Blaschke, Miriam Winkler, Constantin F\"orster, Gabriele Wenger-Glemser, Barbara Plank
Abstract: Although Germany has a diverse landscape of dialects, they are underrepresented in current automatic speech recognition (ASR) research. To enable studies of how robust models are towards dialectal variation, we present Betthupferl, an evaluation dataset containing four hours of read speech in three dialect groups spoken in Southeast Germany (Franconian, Bavarian, Alemannic), and half an hour of Standard German speech. We provide both dialectal and Standard German transcriptions, and analyze the linguistic differences between them. We benchmark several multilingual state-of-the-art ASR models on speech translation into Standard German, and find differences between how much the output resembles the dialectal vs. standardized transcriptions. Qualitative error analyses of the best ASR model reveal that it sometimes normalizes grammatical differences, but often stays closer to the dialectal constructions.
Authors: Yusuke Sakai, Takumi Goto, Taro Watanabe
Abstract: We propose IMPARA-GED, a novel reference-free automatic grammatical error correction (GEC) evaluation method with grammatical error detection (GED) capabilities. We focus on the quality estimator of IMPARA, an existing automatic GEC evaluation method, and construct that of IMPARA-GED using a pre-trained language model with enhanced GED capabilities. Experimental results on SEEDA, a meta-evaluation dataset for automatic GEC evaluation methods, demonstrate that IMPARA-GED achieves the highest correlation with human sentence-level evaluations.
Authors: Yin Fang, Qiao Jin, Guangzhi Xiong, Bowen Jin, Xianrui Zhong, Siru Ouyang, Aidong Zhang, Jiawei Han, Zhiyong Lu
Abstract: Cell type annotation is a key task in analyzing the heterogeneity of single-cell RNA sequencing data. Although recent foundation models automate this process, they typically annotate cells independently, without considering batch-level cellular context or providing explanatory reasoning. In contrast, human experts often annotate distinct cell types for different cell clusters based on their domain knowledge. To mimic this workflow, we introduce the CellPuzzles task, where the objective is to assign unique cell types to a batch of cells. This benchmark spans diverse tissues, diseases, and donor conditions, and requires reasoning across the batch-level cellular context to ensure label uniqueness. We find that off-the-shelf large language models (LLMs) struggle on CellPuzzles, with the best baseline (OpenAI's o1) achieving only 19.0% batch-level accuracy. To fill this gap, we propose Cell-o1, a 7B LLM trained via supervised fine-tuning on distilled reasoning traces, followed by reinforcement learning with batch-level rewards. Cell-o1 achieves state-of-the-art performance, outperforming o1 by over 73% and generalizing well across contexts. Further analysis of training dynamics and reasoning behaviors provides insights into batch-level annotation performance and emergent expert-like reasoning. Code and data are available at https://github.com/ncbi-nlp/cell-o1.
Authors: Yijun Yang, Zeyu Huang, Wenhao Zhu, Zihan Qiu, Fei Yuan, Jeff Z. Pan, Ivan Titov
Abstract: Existing frameworks for evaluating long-context language models (LCLM) can be broadly categorized into real-world and synthetic tasks. Despite their utility, both approaches are accompanied by certain intrinsic limitations. Real-world tasks are too complex to interpret or characterize and are susceptible to data contamination. In contrast, synthetic tasks often adopt the needle-in-the-haystack (NIAH) format, wherein a lack of coherence between the "needle" and the "haystack" compromises their validity as proxies for realistic applications. In response to these challenges, we posit that an ideal long-context evaluation framework should be characterized by three essential features: $\textit{seamless context}$, $\textit{controllable setting}$, and $\textit{sound evaluation}$. This study introduces $\textbf{LongBioBench}$, a novel benchmark that utilizes artificially generated biographies as a controlled environment for assessing LCLMs across dimensions of $\textit{understanding}$, $\textit{reasoning}$, and $\textit{trustworthiness}$. Our experimental evaluation, which includes $\textbf{18}$ LCLMs in total, demonstrates that most models still exhibit deficiencies in semantic understanding and elementary reasoning over retrieved results and are less trustworthy as context length increases. Our further analysis indicates some design choices employed by existing synthetic benchmarks, such as contextual non-coherence, numerical needles, and the absence of distractors, rendering them vulnerable to test the model long-context capabilities. Moreover, we also reveal that long-context continual pretraining primarily adjusts RoPE embedding to accommodate extended context lengths. To sum up, compared to previous synthetic benchmarks, LongBioBench achieves a better trade-off between mirroring authentic language tasks and maintaining controllability, and is highly interpretable and configurable.
Authors: Diogo A. P. Nunes, Eug\'enio Ribeiro
Abstract: In this work, we describe our team's approach to eRisk's 2025 Task 1: Search for Symptoms of Depression. Given a set of sentences and the Beck's Depression Inventory - II (BDI) questionnaire, participants were tasked with submitting up to 1,000 sentences per depression symptom in the BDI, sorted by relevance. Participant submissions were evaluated according to standard Information Retrieval (IR) metrics, including Average Precision (AP) and R-Precision (R-PREC). The provided training data, however, consisted of sentences labeled as to whether a given sentence was relevant or not w.r.t. one of BDI's symptoms. Due to this labeling limitation, we framed our development as a binary classification task for each BDI symptom, and evaluated accordingly. To that end, we split the available labeled data into training and validation sets, and explored foundation model fine-tuning, sentence similarity, Large Language Model (LLM) prompting, and ensemble techniques. The validation results revealed that fine-tuning foundation models yielded the best performance, particularly when enhanced with synthetic data to mitigate class imbalance. We also observed that the optimal approach varied by symptom. Based on these insights, we devised five independent test runs, two of which used ensemble methods. These runs achieved the highest scores in the official IR evaluation, outperforming submissions from 16 other teams.
Authors: Aishwarya Sahoo, Jeevana Kruthi Karnuthala, Tushar Parmanand Budhwani, Pranchal Agarwal, Sankaran Vaidyanathan, Alexa Siu, Franck Dernoncourt, Jennifer Healey, Nedim Lipka, Ryan Rossi, Uttaran Bhattacharya, Branislav Kveton
Abstract: LLM-as-a-judge is a framework in which a large language model (LLM) automatically evaluates the output of another LLM. We propose quantitative LLM judges, which align evaluation scores of existing LLM judges to human scores in a given domain using regression models. The models are trained to improve the score of the original judge by using the judge's textual evaluation and score. We present four quantitative judges for different types of absolute and relative feedback, which showcases the generality and versatility of our framework. Our framework is more computationally efficient than supervised fine-tuning and can be more statistically efficient when human feedback is limited, which is expected in most applications of our work. We validate these claims empirically on four datasets using two base judges. Our experiments show that quantitative judges can effectively improve the predictive power of existing judges through post-hoc modeling.
Authors: Boyi Li, Zhonghan Zhao, Der-Horng Lee, Gaoang Wang
Abstract: Large Language Model (LLM) based multi-agent systems have shown remarkable performance in various tasks, especially when enhanced through collaborative communication. However, current methods often rely on a fixed number of agents and static communication structures, limiting their ability to adapt to varying task complexities. In this paper, we propose Adaptive Graph Pruning (AGP), a novel task-adaptive multi-agent collaboration framework that jointly optimizes agent quantity (hard-pruning) and communication topology (soft-pruning). Specifically, our method employs a two-stage training strategy: firstly, independently training soft-pruning networks for different agent quantities to determine optimal agent-quantity-specific complete graphs and positional masks across specific tasks; and then jointly optimizing hard-pruning and soft-pruning within a maximum complete graph to dynamically configure the number of agents and their communication topologies per task. Extensive experiments demonstrate that our approach is: (1) High-performing, achieving state-of-the-art results across six benchmarks and consistently generalizes across multiple mainstream LLM architectures, with a increase in performance of $2.58\%\sim 9.84\%$; (2) Task-adaptive, dynamically constructing optimized communication topologies tailored to specific tasks, with an extremely high performance in all three task categories (general reasoning, mathematical reasoning, and code generation); (3) Token-economical, having fewer training steps and token consumption at the same time, with a decrease in token consumption of $90\%+$; and (4) Training-efficient, achieving high performance with very few training steps compared with other methods. The performance will surpass the existing baselines after about ten steps of training under six benchmarks.
Authors: Zhixiong Su, Yichen Wang, Herun Wan, Zhaohan Zhang, Minnan Luo
Abstract: The misuse of large language models (LLMs) poses potential risks, motivating the development of machine-generated text (MGT) detection. Existing literature primarily concentrates on binary, document-level detection, thereby neglecting texts that are composed jointly by human and LLM contributions. Hence, this paper explores the possibility of fine-grained MGT detection under human-AI coauthoring. We suggest fine-grained detectors can pave pathways toward coauthored text detection with a numeric AI ratio. Specifically, we propose a dataset, HACo-Det, which produces human-AI coauthored texts via an automatic pipeline with word-level attribution labels. We retrofit seven prevailing document-level detectors to generalize them to word-level detection. Then we evaluate these detectors on HACo-Det on both word- and sentence-level detection tasks. Empirical results show that metric-based methods struggle to conduct fine-grained detection with a 0.462 average F1 score, while finetuned models show superior performance and better generalization across domains. However, we argue that fine-grained co-authored text detection is far from solved. We further analyze factors influencing performance, e.g., context window, and highlight the limitations of current methods, pointing to potential avenues for improvement.
Authors: Yan Gao, Massimo Roberto Scamarcia, Javier Fernandez-Marques, Mohammad Naseri, Chong Shen Ng, Dimitris Stripelis, Zexi Li, Tao Shen, Jiamu Bai, Daoyuan Chen, Zikai Zhang, Rui Hu, InSeo Song, Lee KangYoon, Hong Jia, Ting Dang, Junyan Wang, Zheyuan Liu, Daniel Janes Beutel, Lingjuan Lyu, Nicholas D. Lane
Abstract: Large Language Models (LLMs) have achieved state-of-the-art results across diverse domains, yet their development remains reliant on vast amounts of publicly available data, raising concerns about data scarcity and the lack of access to domain-specific, sensitive information. Federated Learning (FL) presents a compelling framework to address these challenges by enabling decentralized fine-tuning on pre-trained LLMs without sharing raw data. However, the compatibility and performance of pre-trained LLMs in FL settings remain largely under explored. We introduce the FlowerTune LLM Leaderboard, a first-of-its-kind benchmarking suite designed to evaluate federated fine-tuning of LLMs across four diverse domains: general NLP, finance, medical, and coding. Each domain includes federated instruction-tuning datasets and domain-specific evaluation metrics. Our results, obtained through a collaborative, open-source and community-driven approach, provide the first comprehensive comparison across 26 pre-trained LLMs with different aggregation and fine-tuning strategies under federated settings, offering actionable insights into model performance, resource constraints, and domain adaptation. This work lays the foundation for developing privacy-preserving, domain-specialized LLMs for real-world applications.
Authors: Dingwei Chen, Ziqiang Liu, Feiteng Fang, Chak Tou Leong, Shiwen Ni, Ahmadreza Argha, Hamid Alinejad-Rokny, Min Yang, Chengming Li
Abstract: Large Language Models (LLMs) demonstrate remarkable capabilities in text understanding and generation. However, their tendency to produce factually inconsistent outputs, commonly referred to as ''hallucinations'', remains a critical challenge. Existing approaches, such as retrieval-based and inference-time correction methods, primarily address this issue at the input or output level, often overlooking the intrinsic information refinement process and the role of premature layers. Meanwhile, alignment- and fine-tuning-based methods are resource-intensive. In this paper, we propose PLI (Premature Layers Interpolation), a novel, training-free, and plug-and-play intervention designed to enhance factuality. PLI mitigates hallucinations by inserting premature layers formed through mathematical interpolation with adjacent layers. Inspired by stable diffusion and sampling steps, PLI extends the depth of information processing and transmission in LLMs, improving factual coherence. Experiments on four publicly available datasets demonstrate that PLI effectively reduces hallucinations while outperforming existing baselines in most cases. Further analysis suggests that the success of layer interpolation is closely linked to LLMs' internal mechanisms. To promote reproducibility, we will release our code and data upon acceptance.
Authors: Atsumoto Ohashi, Shinya Iizuka, Jingjing Jiang, Ryuichiro Higashinaka
Abstract: Full-duplex spoken dialogue systems, which can model simultaneous bidirectional features of human conversations such as speech overlaps and backchannels, have attracted significant attention recently. However, the study of full-duplex spoken dialogue systems for the Japanese language has been limited, and the research on their development in Japanese remains scarce. In this paper, we present the first publicly available full-duplex spoken dialogue model in Japanese, which is built upon Moshi, a full-duplex dialogue model in English. Our model is trained through a two-stage process: pre-training on a large-scale spoken dialogue data in Japanese, followed by fine-tuning on high-quality stereo spoken dialogue data. We further enhance the model's performance by incorporating synthetic dialogue data generated by a multi-stream text-to-speech system. Evaluation experiments demonstrate that the trained model outperforms Japanese baseline models in both naturalness and meaningfulness.
Authors: Richard Armitage
Abstract: Background: Large language models (LLMs) have demonstrated substantial potential to support clinical practice. Other than Chat GPT4 and its predecessors, few LLMs, especially those of the leading and more powerful reasoning model class, have been subjected to medical specialty examination questions, including in the domain of primary care. This paper aimed to test the capabilities of leading LLMs as of May 2025 (o3, Claude Opus 4, Grok3, and Gemini 2.5 Pro) in primary care education, specifically in answering Member of the Royal College of General Practitioners (MRCGP) style examination questions. Methods: o3, Claude Opus 4, Grok3, and Gemini 2.5 Pro were tasked to answer 100 randomly chosen multiple choice questions from the Royal College of General Practitioners GP SelfTest on 25 May 2025. Questions included textual information, laboratory results, and clinical images. Each model was prompted to answer as a GP in the UK and was provided with full question information. Each question was attempted once by each model. Responses were scored against correct answers provided by GP SelfTest. Results: The total score of o3, Claude Opus 4, Grok3, and Gemini 2.5 Pro was 99.0%, 95.0%, 95.0%, and 95.0%, respectively. The average peer score for the same questions was 73.0%. Discussion: All models performed remarkably well, and all substantially exceeded the average performance of GPs and GP registrars who had answered the same questions. o3 demonstrated the best performance, while the performances of the other leading models were comparable with each other and were not substantially lower than that of o3. These findings strengthen the case for LLMs, particularly reasoning models, to support the delivery of primary care, especially those that have been specifically trained on primary care clinical data.
Authors: Iuliia Zaitova, Badr M. Abdullah, Wei Xue, Dietrich Klakow, Bernd M\"obius, Tania Avgustinova
Abstract: Idioms are defined as a group of words with a figurative meaning not deducible from their individual components. Although modern machine translation systems have made remarkable progress, translating idioms remains a major challenge, especially for speech-to-text systems, where research on this topic is notably sparse. In this paper, we systematically evaluate idiom translation as compared to conventional news translation in both text-to-text machine translation (MT) and speech-to-text translation (SLT) systems across two language pairs (German to English, Russian to English). We compare state-of-the-art end-to-end SLT systems (SeamlessM4T SLT-to-text, Whisper Large v3) with MT systems (SeamlessM4T SLT-to-text, No Language Left Behind), Large Language Models (DeepSeek, LLaMA) and cascaded alternatives. Our results reveal that SLT systems experience a pronounced performance drop on idiomatic data, often reverting to literal translations even in higher layers, whereas MT systems and Large Language Models demonstrate better handling of idioms. These findings underscore the need for idiom-specific strategies and improved internal representations in SLT architectures.
Authors: {\DJ}or{\dj}e Klisura, Astrid R Bernaga Torres, Anna Karen G\'arate-Escamilla, Rajesh Roshan Biswal, Ke Yang, Hilal Pataci, Anthony Rios
Abstract: Privacy policies inform users about data collection and usage, yet their complexity limits accessibility for diverse populations. Existing Privacy Policy Question Answering (QA) systems exhibit performance disparities across English dialects, disadvantaging speakers of non-standard varieties. We propose a novel multi-agent framework inspired by human-centered design principles to mitigate dialectal biases. Our approach integrates a Dialect Agent, which translates queries into Standard American English (SAE) while preserving dialectal intent, and a Privacy Policy Agent, which refines predictions using domain expertise. Unlike prior approaches, our method does not require retraining or dialect-specific fine-tuning, making it broadly applicable across models and domains. Evaluated on PrivacyQA and PolicyQA, our framework improves GPT-4o-mini's zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data. These results highlight the effectiveness of structured agent collaboration in mitigating dialect biases and underscore the importance of designing NLP systems that account for linguistic diversity to ensure equitable access to privacy information.
Authors: Florian Ludwig, Torsten Zesch, Frederike Zufall
Abstract: The assessment of legal problems requires the consideration of a specific legal system and its levels of abstraction, from constitutional law to statutory law to case law. The extent to which Large Language Models (LLMs) internalize such legal systems is unknown. In this paper, we propose and investigate different approaches to condition LLMs at different levels of abstraction in legal systems. This paper examines different approaches to conditioning LLMs at multiple levels of abstraction in legal systems to detect potentially punishable hate speech. We focus on the task of classifying whether a specific social media posts falls under the criminal offense of incitement to hatred as prescribed by the German Criminal Code. The results show that there is still a significant performance gap between models and legal experts in the legal assessment of hate speech, regardless of the level of abstraction with which the models were conditioned. Our analysis revealed, that models conditioned on abstract legal knowledge lacked deep task understanding, often contradicting themselves and hallucinating answers, while models using concrete legal knowledge performed reasonably well in identifying relevant target groups, but struggled with classifying target conducts.
Authors: Aditya Bharat Soni, Boxuan Li, Xingyao Wang, Valerie Chen, Graham Neubig
Abstract: Modern human labor is characterized by specialization; we train for years and develop particular tools that allow us to perform well across a variety of tasks. In addition, AI agents have been specialized for domains such as software engineering, web navigation, and workflow automation. However, this results in agents that are good for one thing but fail to generalize beyond their intended scope. One reason for this is that agent developers provide a highly specialized set of tools or make architectural decisions optimized for a specific use case or benchmark. In this work, we ask the question: what is the minimal set of general tools that can be used to achieve high performance across a diverse set of tasks? Our answer is OpenHands-Versa, a generalist agent built with a modest number of general tools: code editing and execution, web search, as well as multimodal web browsing and file access. Importantly, OpenHands-Versa demonstrates superior or competitive performance over leading specialized agents across three diverse and challenging benchmarks: SWE-Bench Multimodal, GAIA, and The Agent Company, outperforming the best-performing previously published results with absolute improvements in success rate of 9.1, 1.3, and 9.1 points respectively. Further, we show how existing state-of-the-art multi-agent systems fail to generalize beyond their target domains. These results demonstrate the feasibility of developing a generalist agent to solve diverse tasks and establish OpenHands-Versa as a strong baseline for future research.
Authors: Pierre Lepagnol, Sahar Ghannay, Thomas Gerald, Christophe Servan, Sophie Rosset
Abstract: Understanding user queries is fundamental in many applications, such as home assistants, booking systems, or recommendations. Accordingly, it is crucial to develop accurate Spoken Language Understanding (SLU) approaches to ensure the reliability of the considered system. Current State-of-the-Art SLU techniques rely on large amounts of training data; however, only limited annotated examples are available for specific tasks or languages. In the meantime, instruction-tuned large language models (LLMs) have shown exceptional performance on unseen tasks in a few-shot setting when provided with adequate prompts. In this work, we propose to explore example selection by leveraging Information retrieval (IR) approaches to build an enhanced prompt that is applied to an SLU task. We evaluate the effectiveness of the proposed method on several SLU benchmarks. Experimental results show that lexical IR methods significantly enhance performance without increasing prompt length.
Authors: Jintian Shao, Yiming Cheng
Abstract: Reinforcement learning (RL) enhances large language models (LLMs) in complex, long-chain-of-thought (long-CoT) reasoning. The advanced VAPO framework, despite sophisticated mechanisms like Decoupled GAE, theoretically faces fundamental limitations in comprehensively modeling and leveraging deep, long-term value for fine-grained, step-by-step policy guidance in extended reasoning chains. We argue these limitations stem from inherent difficulties in credit assignment, value function representational capacity with temporally abstracted goals, and translating global value signals into local policy improvements, especially with sparse rewards. Our theoretical analysis examines these aspects to illuminate VAPO's boundaries in long-term value modeling, aiming to deepen understanding of current RL for advanced reasoning and suggest future research for more robust LLM agents.
Authors: Yuval Kansal, Shmuel Berman, Lydia Liu
Abstract: Factuality is a necessary precursor to useful educational tools. As adoption of Large Language Models (LLMs) in education continues of grow, ensuring correctness in all settings is paramount. Despite their strong English capabilities, LLM performance in other languages is largely untested. In this work, we evaluate the correctness of the Llama3.1 family of models in answering factual questions appropriate for middle and high school students. We demonstrate that LLMs not only provide extraneous and less truthful information, but also exacerbate existing biases against rare languages.
Authors: Katherine Thai, Mohit Iyyer
Abstract: How well do modern long-context language models understand literary fiction? We explore this question via the task of literary evidence retrieval, repurposing the RELiC dataset of That et al. (2022) to construct a benchmark where the entire text of a primary source (e.g., The Great Gatsby) is provided to an LLM alongside literary criticism with a missing quotation from that work. This setting, in which the model must generate the missing quotation, mirrors the human process of literary analysis by requiring models to perform both global narrative reasoning and close textual examination. We curate a high-quality subset of 292 examples through extensive filtering and human verification. Our experiments show that recent reasoning models, such as Gemini Pro 2.5 can exceed human expert performance (62.5% vs. 50% accuracy). In contrast, the best open-weight model achieves only 29.1% accuracy, highlighting a wide gap in interpretive reasoning between open and closed-weight models. Despite their speed and apparent accuracy, even the strongest models struggle with nuanced literary signals and overgeneration, signaling open challenges for applying LLMs to literary analysis. We release our dataset and evaluation code to encourage future work in this direction.
Authors: Jonas F. Lotz, Ant\'onio V. Lopes, Stephan Peitz, Hendra Setiawan, Leonardo Emili
Abstract: The choice of tokenizer can profoundly impact language model performance, yet accessible and reliable evaluations of tokenizer quality remain an open challenge. Inspired by scaling consistency, we show that smaller models can accurately predict significant differences in tokenizer impact on larger models at a fraction of the compute cost. By systematically evaluating both English-centric and multilingual tokenizers, we find that tokenizer choice has negligible effects on tasks in English but results in consistent performance differences in multilingual settings. We propose new intrinsic tokenizer metrics inspired by Zipf's law that correlate more strongly with downstream performance than text compression when modeling unseen languages. By combining several metrics to capture multiple aspects of tokenizer behavior, we develop a reliable framework for intrinsic tokenizer evaluations. Our work offers a more efficient path to informed tokenizer selection in future language model development.
Authors: Xiaoying Zhang, Hao Sun, Yipeng Zhang, Kaituo Feng, Chaochao Lu, Chao Yang, Helen Meng
Abstract: Recent advances in reinforcement learning (RL) with numerical feedback, such as scalar rewards, have significantly enhanced the complex reasoning capabilities of large language models (LLMs). Despite this success, we identify three key challenges encountered by RL with solely numerical feedback: performance plateaus, limited effectiveness of self-reflection, and persistent failures. We then demonstrate that RL-finetuned models, even after exhibiting performance plateaus, can generate correct refinements on persistently failed problems by leveraging natural language feedback in the form of critiques. Building on this insight, we propose Critique-GRPO, an online RL framework that integrates both natural language and numerical feedback for effective policy optimization. Critique-GRPO enables LLMs to learn from initial responses and critique-guided refinements simultaneously while maintaining exploration. Extensive experiments using Qwen2.5-7B-Base and Qwen3-8B-Base show that Critique-GRPO consistently outperforms supervised learning-based and RL-based fine-tuning approaches across eight challenging mathematical, STEM, and general reasoning tasks, improving average pass@1 scores by approximately 4.5% and 5%, respectively. Notably, Critique-GRPO surpasses a strong baseline that incorporates expert demonstrations within online RL. Further analysis reveals two critical insights about policy exploration: (1) higher entropy does not always guarantee efficient learning from exploration, and (2) longer responses do not necessarily lead to more effective exploration.
Authors: Prashanth Vijayaraghavan, Luyao Shi, Ehsan Degan, Vandana Mukherjee, Xin Zhang
Abstract: Analog circuit topology synthesis is integral to Electronic Design Automation (EDA), enabling the automated creation of circuit structures tailored to specific design requirements. However, the vast design search space and strict constraint adherence make efficient synthesis challenging. Leveraging the versatility of Large Language Models (LLMs), we propose AUTOCIRCUIT-RL,a novel reinforcement learning (RL)-based framework for automated analog circuit synthesis. The framework operates in two phases: instruction tuning, where an LLM learns to generate circuit topologies from structured prompts encoding design constraints, and RL refinement, which further improves the instruction-tuned model using reward models that evaluate validity, efficiency, and output voltage. The refined model is then used directly to generate topologies that satisfy the design constraints. Empirical results show that AUTOCIRCUIT-RL generates ~12% more valid circuits and improves efficiency by ~14% compared to the best baselines, while reducing duplicate generation rates by ~38%. It achieves over 60% success in synthesizing valid circuits with limited training data, demonstrating strong generalization. These findings highlight the framework's effectiveness in scaling to complex circuits while maintaining efficiency and constraint adherence, marking a significant advancement in AI-driven circuit design.
Authors: Yinjie Wang, Ling Yang, Ye Tian, Ke Shen, Mengdi Wang
Abstract: We propose CURE, a novel reinforcement learning framework with a dedicated reward design that co-evolves coding and unit test generation capabilities based on their interaction outcomes, without any ground-truth code as supervision. This approach enables flexible and scalable training and allows the unit tester to learn directly from the coder's mistakes. Our derived ReasonFlux-Coder-7B and 14B models improve code generation accuracy by 5.3% and Best-of-N accuracy by 9.0% after optimization on Qwen2.5-Instruct models, outperforming similarly sized Qwen-Coder, DeepSeek-Coder, and Seed-Coder. They naturally extend to downstream tasks such as test-time scaling and agentic coding-achieving a 8.1% improvement over the base model. For the long-CoT model, our ReasonFlux-Coder-4B consistently outperforms Qwen3-4B while achieving 64.8% inference efficiency in unit test generation. Notably, we also find that our model can serve as an effective reward model for reinforcement learning on base models. Project: https://github.com/Gen-Verse/CURE
Authors: Qianhui Wu, Kanzhi Cheng, Rui Yang, Chaoyun Zhang, Jianwei Yang, Huiqiang Jiang, Jian Mu, Baolin Peng, Bo Qiao, Reuben Tan, Si Qin, Lars Liden, Qingwei Lin, Huan Zhang, Tong Zhang, Jianbing Zhang, Dongmei Zhang, Jianfeng Gao
Abstract: One of the principal challenges in building VLM-powered GUI agents is visual grounding, i.e., localizing the appropriate screen region for action execution based on both the visual content and the textual plans. Most existing work formulates this as a text-based coordinate generation task. However, these approaches suffer from several limitations: weak spatial-semantic alignment, inability to handle ambiguous supervision targets, and a mismatch between the dense nature of screen coordinates and the coarse, patch-level granularity of visual features extracted by models like Vision Transformers. In this paper, we propose GUI-Actor, a VLM-based method for coordinate-free GUI grounding. At its core, GUI-Actor introduces an attention-based action head that learns to align a dedicated
Authors: Pralaypati Ta, Sriram Venkatesaperumal, Keerthi Ram, Mohanasankar Sivaprakasam
Abstract: Neuroscience research publications encompass a vast wealth of knowledge. Accurately retrieving existing information and discovering new insights from this extensive literature is essential for advancing the field. However, when knowledge is dispersed across multiple sources, current state-of-the-art retrieval methods often struggle to extract the necessary information. A knowledge graph (KG) can integrate and link knowledge from multiple sources, but existing methods for constructing KGs in neuroscience often rely on labeled data and require domain expertise. Acquiring large-scale, labeled data for a specialized area like neuroscience presents significant challenges. This work proposes novel methods for constructing KG from unlabeled large-scale neuroscience research corpus utilizing large language models (LLM), neuroscience ontology, and text embeddings. We analyze the semantic relevance of neuroscience text segments identified by LLM for building the knowledge graph. We also introduce an entity-augmented information retrieval algorithm to extract knowledge from the KG. Several experiments were conducted to evaluate the proposed approaches, and the results demonstrate that our methods significantly enhance knowledge discovery from the unlabeled neuroscience research corpus. It achieves an F1 score of 0.84 for entity extraction, and the knowledge obtained from the KG improves answers to over 54% of the questions.
Authors: Pietro Lesci, Clara Meister, Thomas Hofmann, Andreas Vlachos, Tiago Pimentel
Abstract: Modern language models are typically trained over subword sequences, but ultimately define probabilities over character-strings. Ideally, the choice of the tokeniser -- which maps character-strings to subwords -- should not affect the probability assigned to the underlying character-string; in practice, it does. We define this mismatch as tokenisation bias. In this work, we quantify one particular type of tokenisation bias: the effect of including or not a subword (e.g., $\langle hello \rangle$) in a tokeniser's vocabulary on the probability a trained model assigns to the corresponding characters (i.e., \textit{``hello''}). Estimating this effect is challenging because each model is trained with only one tokeniser. We address this by framing tokenisation bias as a causal effect and estimating it using the regression discontinuity design. Specifically, we exploit the fact that tokenisation algorithms rank subwords and add the first $K$ to a tokeniser's vocabulary, where $K$ is an arbitrary cutoff point. As such, we can estimate a causal effect by comparing similar subwords around this cutoff. Experimentally, we find that tokenisation consistently affects models' outputs across scales, vocabularies, and tokenisers. Notably, a subword's presence in a small model's vocabulary may increase its characters' probability by up to 17 times, highlighting tokenisation as a key design choice in language modelling.
Authors: Jiongnan Liu, Zhicheng Dou, Ning Hu, Chenyan Xiong
Abstract: To address the challenge of information overload from massive web contents, recommender systems are widely applied to retrieve and present personalized results for users. However, recommendation tasks are inherently constrained to filtering existing items and lack the ability to generate novel concepts, limiting their capacity to fully satisfy user demands and preferences. In this paper, we propose a new paradigm that goes beyond content filtering and selecting: directly generating personalized items in a multimodal form, such as images, tailored to individual users. To accomplish this, we leverage any-to-any Large Multimodal Models (LMMs) and train them in both supervised fine-tuning and online reinforcement learning strategy to equip them with the ability to yield tailored next items for users. Experiments on two benchmark datasets and user study confirm the efficacy of the proposed method. Notably, the generated images not only align well with users' historical preferences but also exhibit relevance to their potential future interests.
Authors: Andrew Kiruluta, Preethi Raju, Priscilla Burity
Abstract: We present a novel non attention based architecture for large language models (LLMs) that efficiently handles very long context windows, on the order of hundreds of thousands to potentially millions of tokens. Unlike traditional Transformer designs, which suffer from quadratic memory and computation overload due to the nature of the self attention mechanism, our model avoids token to token attention entirely. Instead, it combines the following complementary components: State Space blocks (inspired by S4) that learn continuous time convolution kernels and scale near linearly with sequence length, Multi Resolution Convolution layers that capture local context at different dilation levels, a lightweight Recurrent Supervisor to maintain a global hidden state across sequential chunks, and Retrieval Augmented External Memory that stores and retrieves high-level chunk embeddings without reintroducing quadratic operations.
Authors: Patrik Czak\'o, G\'abor Kert\'esz, S\'andor Sz\'en\'asi
Abstract: Quantization effectively reduces the serving costs of Large Language Models (LLMs) by speeding up data movement through compressed parameters and enabling faster operations via integer arithmetic. However, activating integer arithmetic requires quantizing both weights and activations, which poses challenges due to the significant outliers in LLMs that increase quantization error. In this work, we investigate these outliers with an emphasis on their effect on layer-wise quantization error, then examine how smoothing and rotation transform the observed values. Our primary contributions include introducing a new metric to measure and visualize quantization difficulty based on channel magnitudes, as well as proposing a hybrid approach that applies channel-wise scaling before rotation, supported by a mathematical formulation of its benefits.
Authors: Takao Fujii, Katie Seaborn, Madeleine Steeds, Jun Kato
Abstract: Conversational agents that mimic people have raised questions about the ethics of anthropomorphizing machines with human social identity cues. Critics have also questioned assumptions of identity neutrality in humanlike agents. Recent work has revealed that intersectional Japanese pronouns can elicit complex and sometimes evasive impressions of agent identity. Yet, the role of other "neutral" non-pronominal self-referents (NPSR) and voice as a socially expressive medium remains unexplored. In a crowdsourcing study, Japanese participants (N = 204) evaluated three ChatGPT voices (Juniper, Breeze, and Ember) using seven self-referents. We found strong evidence of voice gendering alongside the potential of intersectional self-referents to evade gendering, i.e., ambiguity through neutrality and elusiveness. Notably, perceptions of age and formality intersected with gendering as per sociolinguistic theories, especially boku and watakushi. This work provides a nuanced take on agent identity perceptions and champions intersectional and culturally-sensitive work on voice agents.
Authors: David Sasu, Kweku Andoh Yamoah, Benedict Quartey, Natalie Schluter
Abstract: Enabling robots to accurately interpret and execute spoken language instructions is essential for effective human-robot collaboration. Traditional methods rely on speech recognition to transcribe speech into text, often discarding crucial prosodic cues needed for disambiguating intent. We propose a novel approach that directly leverages speech prosody to infer and resolve instruction intent. Predicted intents are integrated into large language models via in-context learning to disambiguate and select appropriate task plans. Additionally, we present the first ambiguous speech dataset for robotics, designed to advance research in speech disambiguation. Our method achieves 95.79% accuracy in detecting referent intents within an utterance and determines the intended task plan of ambiguous instructions with 71.96% accuracy, demonstrating its potential to significantly improve human-robot communication.
Authors: Ziwei Gong, Pengyuan Shi, Kaan Donbekci, Lin Ai, Run Chen, David Sasu, Zehui Wu, Julia Hirschberg
Abstract: Speech Emotion Recognition (SER) has seen significant progress with deep learning, yet remains challenging for Low-Resource Languages (LRLs) due to the scarcity of annotated data. In this work, we explore unsupervised learning to improve SER in low-resource settings. Specifically, we investigate contrastive learning (CL) and Bootstrap Your Own Latent (BYOL) as self-supervised approaches to enhance cross-lingual generalization. Our methods achieve notable F1 score improvements of 10.6% in Urdu, 15.2% in German, and 13.9% in Bangla, demonstrating their effectiveness in LRLs. Additionally, we analyze model behavior to provide insights on key factors influencing performance across languages, and also highlighting challenges in low-resource SER. This work provides a foundation for developing more inclusive, explainable, and robust emotion recognition systems for underrepresented languages.
Authors: Yoonjun Cho, Soeun Kim, Dongjae Jeon, Kyelim Lee, Beomsoo Lee, Albert No
Abstract: Decomposing weight matrices into quantization and low-rank components ($\mathbf{W} \approx \mathbf{Q} + \mathbf{L}\mathbf{R}$) is a widely used technique for compressing large language models (LLMs). Existing joint optimization methods iteratively alternate between quantization and low-rank approximation. However, these methods tend to prioritize one component at the expense of the other, resulting in suboptimal decompositions that fail to leverage each component's unique strengths. In this work, we introduce Outlier-Driven Low-Rank Initialization (ODLRI), which assigns low-rank components the specific role of capturing activation-sensitive weights. This structured decomposition mitigates outliers' negative impact on quantization, enabling more effective balance between quantization and low-rank approximation. Experiments on Llama2 (7B, 13B, 70B), Llama3-8B, and Mistral-7B demonstrate that incorporating ODLRI into the joint optimization framework consistently reduces activation-aware error, minimizes quantization scale, and improves perplexity and zero-shot accuracy in low-bit settings.
Authors: Ajinkya Kulkarni, Sandipana Dowerah, Tanel Alumae, Mathew Magimai. -Doss
Abstract: Audio deepfakes are acquiring an unprecedented level of realism with advanced AI. While current research focuses on discerning real speech from spoofed speech, tracing the source system is equally crucial. This work proposes a novel audio source tracing system combining deep metric multi-class N-pair loss with Real Emphasis and Fake Dispersion framework, a Conformer classification network, and ensemble score-embedding fusion. The N-pair loss improves discriminative ability, while Real Emphasis and Fake Dispersion enhance robustness by focusing on differentiating real and fake speech patterns. The Conformer network captures both global and local dependencies in the audio signal, crucial for source tracing. The proposed ensemble score-embedding fusion shows an optimal trade-off between in-domain and out-of-domain source tracing scenarios. We evaluate our method using Frechet Distance and standard metrics, demonstrating superior performance in source tracing over the baseline system.
Authors: Alef Iury Siqueira Ferreira, Lucas Rafael Gris, Alexandre Ferro Filho, Lucas \'Olives, Daniel Ribeiro, Luiz Fernando, Fernanda Lustosa, Rodrigo Tanaka, Frederico Santos de Oliveira, Arlindo Galv\~ao Filho
Abstract: Training SER models in natural, spontaneous speech is especially challenging due to the subtle expression of emotions and the unpredictable nature of real-world audio. In this paper, we present a robust system for the INTERSPEECH 2025 Speech Emotion Recognition in Naturalistic Conditions Challenge, focusing on categorical emotion recognition. Our method combines state-of-the-art audio models with text features enriched by prosodic and spectral cues. In particular, we investigate the effectiveness of Fundamental Frequency (F0) quantization and the use of a pretrained audio tagging model. We also employ an ensemble model to improve robustness. On the official test set, our system achieved a Macro F1-score of 39.79% (42.20% on validation). Our results underscore the potential of these methods, and analysis of fusion techniques confirmed the effectiveness of Graph Attention Networks. Our source code is publicly available.
Authors: Zijian Wu, Jinjie Ni, Xiangyan Liu, Zichen Liu, Hang Yan, Michael Qizhe Shieh
Abstract: Vision-language models (VLMs) trained via reinforcement learning with verifiable reward (RLVR) have shown notable progress in scaling test-time compute effectively. In this work, we investigate how synthesized RL data can further improve RLVR. To this end, we propose \textbf{SynthRL}-a scalable and guaranteed pipeline for automatic data scaling in reasoning-oriented RL training. SynthRL comprises three key stages: (1) selecting seed questions with appropriate distribution, (2) augmenting them into more challenging variants while preserving the original answers, and (3) a guaranteed verification stage that ensures near-perfect correctness and difficulty enhancement. Our empirical experiments demonstrate SynthRL's scalability and effectiveness. When applied to the MMK12 dataset, SynthRL synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. Models trained with our synthesized data achieve consistent gains across five out-of-domain visual math reasoning benchmarks, with a significant improvement over baseline models trained on seed data alone. Notably, detailed analysis reveals that the gains are more pronounced on the most challenging evaluation samples, highlighting SynthRL's effectiveness in eliciting deeper and more complex reasoning patterns.
Authors: Madan Krishnamurthy, Daniel Korn, Melissa A Haendel, Christopher J Mungall, Anne E Thessen
Abstract: This research aims to develop a dynamic and scalable framework to facilitate harmonization of Common Data Elements (CDEs) across heterogeneous biomedical datasets by addressing challenges such as semantic heterogeneity, structural variability, and context dependence to streamline integration, enhance interoperability, and accelerate scientific discovery. Our methodology leverages Large Language Models (LLMs) for context-aware text embeddings that convert CDEs into dense vectors capturing semantic relationships and patterns. These embeddings are clustered using Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to group semantically similar CDEs. The framework incorporates four key steps: (1) LLM-based text embedding to mathematically represent semantic context, (2) unsupervised clustering of embeddings via HDBSCAN, (3) automated labeling using LLM summarization, and (4) supervised learning to train a classifier assigning new or unclustered CDEs to labeled clusters. Evaluated on the NIH NLM CDE Repository with over 24,000 CDEs, the system identified 118 meaningful clusters at an optimized minimum cluster size of 20. The classifier achieved 90.46 percent overall accuracy, performing best in larger categories. External validation against Gravity Projects Social Determinants of Health domains showed strong agreement (Adjusted Rand Index 0.52, Normalized Mutual Information 0.78), indicating that embeddings effectively capture cluster characteristics. This adaptable and scalable approach offers a practical solution to CDE harmonization, improving selection efficiency and supporting ongoing data interoperability.
Authors: Thai-Binh Nguyen, Ngoc-Quan Pham, Alexander Waibel
Abstract: Audio-Visual Speech Recognition (AVSR) offers a robust solution for speech recognition in challenging environments, such as cocktail-party scenarios, where relying solely on audio proves insufficient. However, current AVSR models are often optimized for idealized scenarios with consistently active speakers, overlooking the complexities of real-world settings that include both speaking and silent facial segments. This study addresses this gap by introducing a novel audio-visual cocktail-party dataset designed to benchmark current AVSR systems and highlight the limitations of prior approaches in realistic noisy conditions. Additionally, we contribute a 1526-hour AVSR dataset comprising both talking-face and silent-face segments, enabling significant performance gains in cocktail-party environments. Our approach reduces WER by 67% relative to the state-of-the-art, reducing WER from 119% to 39.2% in extreme noise, without relying on explicit segmentation cues.
Authors: Hongling Xu, Qi Zhu, Heyuan Deng, Jinpeng Li, Lu Hou, Yasheng Wang, Lifeng Shang, Ruifeng Xu, Fei Mi
Abstract: Recent advances in large language model (LLM) post-training have leveraged two distinct paradigms to enhance reasoning capabilities: reinforcement learning (RL) and knowledge distillation (KD). While RL enables the emergence of complex reasoning behaviors, it often suffers from low sample efficiency when the initial policy struggles to explore high-reward trajectories. Conversely, KD improves learning efficiency via mimicking the teacher model but tends to generalize poorly to out-of-domain scenarios. In this work, we present \textbf{KDRL}, a \textit{unified post-training framework} that jointly optimizes a reasoning model through teacher supervision (KD) and self-exploration (RL). Specifically, KDRL leverages policy gradient optimization to simultaneously minimize the reverse Kullback-Leibler divergence (RKL) between the student and teacher distributions while maximizing the expected rule-based rewards. We first formulate a unified objective that integrates GRPO and KD, and systematically explore how different KL approximations, KL coefficients, and reward-guided KD strategies affect the overall post-training dynamics and performance. Empirical results on multiple reasoning benchmarks demonstrate that KDRL outperforms GRPO and various KD baselines while achieving a favorable balance between performance and reasoning token efficiency. These findings indicate that integrating KD and RL serves as an effective and efficient strategy to train reasoning LLMs.
Authors: Manas Mehta, Yimu Pan, Kelly Gallagher, Alison D. Gernand, Jeffery A. Goldstein, Delia Mwinyelle, Leena Mithal, James Z. Wang
Abstract: Pathological examination of the placenta is an effective method for detecting and mitigating health risks associated with childbirth. Recent advancements in AI have enabled the use of photographs of the placenta and pathology reports for detecting and classifying signs of childbirth-related pathologies. However, existing automated methods are computationally extensive, which limits their deployability. We propose two modifications to vision-language contrastive learning (VLC) frameworks to enhance their accuracy and efficiency: (1) text-anchored vision-language contrastive knowledge distillation (VLCD)-a new knowledge distillation strategy for medical VLC pretraining, and (2) unsupervised predistillation using a large natural images dataset for improved initialization. Our approach distills efficient neural networks that match or surpass the teacher model in performance while achieving model compression and acceleration. Our results showcase the value of unsupervised predistillation in improving the performance and robustness of our approach, specifically for lower-quality images. VLCD serves as an effective way to improve the efficiency and deployability of medical VLC approaches, making AI-based healthcare solutions more accessible, especially in resource-constrained environments.
Authors: Tianyu Hua, Harper Hua, Violet Xiang, Benjamin Klieger, Sang T. Truong, Weixin Liang, Fan-Yun Sun, Nick Haber
Abstract: Large language models (LLMs) have shown promise in transforming machine learning research, yet their capability to faithfully implement novel ideas from recent research papers-ideas unseen during pretraining-remains unclear. We introduce ResearchCodeBench, a benchmark of 212 coding challenges that evaluates LLMs' ability to translate cutting-edge ML contributions from top 2024-2025 research papers into executable code. We assessed 30+ proprietary and open-source LLMs, finding that even the best models correctly implement less than 40% of the code. We find Gemini-2.5-Pro-Preview to perform best at 37.3% success rate, with O3 (High) and O4-mini (High) following behind at 32.3% and 30.8% respectively. We present empirical findings on performance comparison, contamination, and error patterns. By providing a rigorous and community-driven evaluation platform, ResearchCodeBench enables continuous understanding and advancement of LLM-driven innovation in research code generation.
Authors: Fengjin Li, Jie Wang, Yadong Niu, Yongqing Wang, Meng Meng, Jian Luan, Zhiyong Wu
Abstract: Voice Conversion (VC) modifies speech to match a target speaker while preserving linguistic content. Traditional methods usually extract speaker information directly from speech while neglecting the explicit utilization of linguistic content. Since VC fundamentally involves disentangling speaker identity from linguistic content, leveraging structured semantic features could enhance conversion performance. However, previous attempts to incorporate semantic features into VC have shown limited effectiveness, motivating the integration of explicit text modeling. We propose StarVC, a unified autoregressive VC framework that first predicts text tokens before synthesizing acoustic features. The experiments demonstrate that StarVC outperforms conventional VC methods in preserving both linguistic content (i.e., WER and CER) and speaker characteristics (i.e., SECS and MOS). Audio demo can be found at: https://thuhcsi.github.io/StarVC/.
Authors: Jiaxi Hu, Yongqi Pan, Jusen Du, Disen Lan, Xiaqiang Tang, Qingsong Wen, Yuxuan Liang, Weigao Sun
Abstract: Recent efficient sequence modeling methods such as Gated DeltaNet, TTT, and RWKV-7 have achieved performance improvements by supervising the recurrent memory management through Delta learning rule. Unlike previous state-space models (e.g., Mamba) and gated linear attentions (e.g., GLA), these models introduce interactions between the recurrent state and the key vector, resulting in a nonlinear recursive structure. In this paper, we first introduce the concept of Nonlinear RNNs with a comprehensive analysis on the advantages and limitations of these models. Then, based on closed-loop control theory, we propose a novel Nonlinear RNN variant named Comba, which adopts a scalar-plus-low-rank state transition, with both state feedback and output feedback corrections. We also implement a hardware-efficient chunk-wise parallel kernel in Triton and train models with 340M/1.3B parameters on large-scale corpus. Comba demonstrates its superior performance and computation efficiency in both language and vision modeling.
Authors: Kalyan Nakka, Nitesh Saxena
Abstract: The inherent risk of generating harmful and unsafe content by Large Language Models (LLMs), has highlighted the need for their safety alignment. Various techniques like supervised fine-tuning, reinforcement learning from human feedback, and red-teaming were developed for ensuring the safety alignment of LLMs. However, the robustness of these aligned LLMs is always challenged by adversarial attacks that exploit unexplored and underlying vulnerabilities of the safety alignment. In this paper, we develop a novel black-box jailbreak attack, called BitBypass, that leverages hyphen-separated bitstream camouflage for jailbreaking aligned LLMs. This represents a new direction in jailbreaking by exploiting fundamental information representation of data as continuous bits, rather than leveraging prompt engineering or adversarial manipulations. Our evaluation of five state-of-the-art LLMs, namely GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral, in adversarial perspective, revealed the capabilities of BitBypass in bypassing their safety alignment and tricking them into generating harmful and unsafe content. Further, we observed that BitBypass outperforms several state-of-the-art jailbreak attacks in terms of stealthiness and attack success. Overall, these results highlights the effectiveness and efficiency of BitBypass in jailbreaking these state-of-the-art LLMs.
Authors: Nguyen-Khang Le, Quan Minh Bui, Minh Ngoc Nguyen, Hiep Nguyen, Trung Vo, Son T. Luu, Shoshin Nomura, Minh Le Nguyen
Abstract: Web applications are critical to modern software ecosystems, yet ensuring their reliability remains challenging due to the complexity and dynamic nature of web interfaces. Recent advances in large language models (LLMs) have shown promise in automating complex tasks, but limitations persist in handling dynamic navigation flows and complex form interactions. This paper presents an automated system for generating test cases for two key aspects of web application testing: site navigation and form filling. For site navigation, the system employs screen transition graphs and LLMs to model navigation flows and generate test scenarios. For form filling, it uses state graphs to handle conditional forms and automates Selenium script generation. Key contributions include: (1) a novel integration of graph structures and LLMs for site navigation testing, (2) a state graph-based approach for automating form-filling test cases, and (3) a comprehensive dataset for evaluating form-interaction testing. Experimental results demonstrate the system's effectiveness in improving test coverage and robustness, advancing the state of web application testing.
Authors: Shenghua He, Tian Xia, Xuan Zhou, Hui Wei
Abstract: We study a common challenge in reinforcement learning for large language models (LLMs): the Zero-Reward Assumption, where non-terminal actions (i.e., intermediate token generations) receive zero task-specific immediate reward, while only the final token receives a reward for the entire response. This assumption arises frequently in practice, as precise token-level rewards are often difficult or infeasible to obtain in LLM applications. In this work, we provide a unifying theoretical perspective. We introduce the Trajectory Policy Gradient Theorem, which shows that the policy gradient based on true, unknown token-level rewards can be unbiasedly estimated using only a response-level reward model, regardless of whether the Zero-Reward Assumption holds or not, for algorithms in the REINFORCE and Actor-Critic families. This result reveals that widely used methods such as PPO, GRPO, ReMax, and RLOO inherently possess the capacity to model token-level reward signals, offering a theoretical justification for response-level reward approaches. Our findings pave the way for more practical, efficient LLM fine-tuning, allowing developers to treat training algorithms as black boxes and focus on improving the response-level reward model with auxiliary sub-models. We also offer a detailed analysis of popular RL and non-RL methods, comparing their theoretical foundations and practical advantages across common LLM tasks. Finally, we propose a new algorithm: Token-Reinforced Policy Optimization (TRePO), a theoretically grounded method that is simpler than PPO, matches GRPO in memory efficiency, and holds promise for broad applicability.
Authors: Dimitrios Koutsianos, Stavros Zacharopoulos, Yannis Panagakis, Themos Stafylakis
Abstract: This paper addresses source tracing in synthetic speech-identifying generative systems behind manipulated audio via speaker recognition-inspired pipelines. While prior work focuses on spoofing detection, source tracing lacks robust solutions. We evaluate two approaches: classification-based and metric-learning. We tested our methods on the MLAADv5 benchmark using ResNet and self-supervised learning (SSL) backbones. The results show that ResNet achieves competitive performance with the metric learning approach, matching and even exceeding SSL-based systems. Our work demonstrates ResNet's viability for source tracing while underscoring the need to optimize SSL representations for this task. Our work bridges speaker recognition methodologies with audio forensic challenges, offering new directions for combating synthetic media manipulation.
Authors: Naoto Tanji, Toshihiko Yamasaki
Abstract: Image scoring is a crucial task in numerous real-world applications. To trust a model's judgment, understanding its rationale is essential. This paper proposes a novel training method for Vision Language Models (VLMs) to generate not only image scores but also corresponding justifications in natural language. Leveraging only an image scoring dataset and an instruction-tuned VLM, our method enables self-training, utilizing the VLM's generated text without relying on external data or models. In addition, we introduce a simple method for creating a dataset designed to improve alignment between predicted scores and their textual justifications. By iteratively training the model with Direct Preference Optimization on two distinct datasets and merging them, we can improve both scoring accuracy and the coherence of generated explanations.
Authors: Xiaochong Lan, Jie Feng, Jiahuan Lei, Xinlei Shi, Yong Li
Abstract: Large language models (LLMs) have exhibited remarkable capabilities and achieved significant breakthroughs across various domains, leading to their widespread adoption in recent years. Building on this progress, we investigate their potential in the realm of local life services. In this study, we establish a comprehensive benchmark and systematically evaluate the performance of diverse LLMs across a wide range of tasks relevant to local life services. To further enhance their effectiveness, we explore two key approaches: model fine-tuning and agent-based workflows. Our findings reveal that even a relatively compact 7B model can attain performance levels comparable to a much larger 72B model, effectively balancing inference cost and model capability. This optimization greatly enhances the feasibility and efficiency of deploying LLMs in real-world online services, making them more practical and accessible for local life applications.
Authors: Po-Chieh Yu
Abstract: We present an exploratory framework to test whether noise-like input can induce structured responses in language models. Instead of assuming that extraterrestrial signals must be decoded, we evaluate whether inputs can trigger linguistic behavior in generative systems. This shifts the focus from decoding to viewing structured output as a sign of underlying regularity in the input. We tested GPT-2 small, a 117M-parameter model trained on English text, using four types of acoustic input: human speech, humpback whale vocalizations, Phylloscopus trochilus birdsong, and algorithmically generated white noise. All inputs were treated as noise-like, without any assumed symbolic encoding. To assess reactivity, we defined a composite score called Semantic Induction Potential (SIP), combining entropy, syntax coherence, compression gain, and repetition penalty. Results showed that whale and bird vocalizations had higher SIP scores than white noise, while human speech triggered only moderate responses. This suggests that language models may detect latent structure even in data without conventional semantics. We propose that this approach could complement traditional SETI methods, especially in cases where communicative intent is unknown. Generative reactivity may offer a different way to identify data worth closer attention.
Authors: Renyang Liu, Wenjie Feng, Tianwei Zhang, Wei Zhou, Xueqi Cheng, See-Kiong Ng
Abstract: With the surge and widespread application of image generation models, data privacy and content safety have become major concerns and attracted great attention from users, service providers, and policymakers. Machine unlearning (MU) is recognized as a cost-effective and promising means to address these challenges. Despite some advancements, image generation model unlearning (IGMU) still faces remarkable gaps in practice, e.g., unclear task discrimination and unlearning guidelines, lack of an effective evaluation framework, and unreliable evaluation metrics. These can hinder the understanding of unlearning mechanisms and the design of practical unlearning algorithms. We perform exhaustive assessments over existing state-of-the-art unlearning algorithms and evaluation standards, and discover several critical flaws and challenges in IGMU tasks. Driven by these limitations, we make several core contributions, to facilitate the comprehensive understanding, standardized categorization, and reliable evaluation of IGMU. Specifically, (1) We design CatIGMU, a novel hierarchical task categorization framework. It provides detailed implementation guidance for IGMU, assisting in the design of unlearning algorithms and the construction of testbeds. (2) We introduce EvalIGMU, a comprehensive evaluation framework. It includes reliable quantitative metrics across five critical aspects. (3) We construct DataIGM, a high-quality unlearning dataset, which can be used for extensive evaluations of IGMU, training content detectors for judgment, and benchmarking the state-of-the-art unlearning algorithms. With EvalIGMU and DataIGM, we discover that most existing IGMU algorithms cannot handle the unlearning well across different evaluation dimensions, especially for preservation and robustness. Code and models are available at https://github.com/ryliu68/IGMU.
Authors: Chen Qian, Dongrui Liu, Haochen Wen, Zhen Bai, Yong Liu, Jing Shao
Abstract: Large reasoning models (LRMs) have demonstrated impressive capabilities in complex problem-solving, yet their internal reasoning mechanisms remain poorly understood. In this paper, we investigate the reasoning trajectories of LRMs from an information-theoretic perspective. By tracking how mutual information (MI) between intermediate representations and the correct answer evolves during LRM reasoning, we observe an interesting MI peaks phenomenon: the MI at specific generative steps exhibits a sudden and significant increase during LRM's reasoning process. We theoretically analyze such phenomenon and show that as MI increases, the probability of model's prediction error decreases. Furthermore, these MI peaks often correspond to tokens expressing reflection or transition, such as ``Hmm'', ``Wait'' and ``Therefore,'' which we term as the thinking tokens. We then demonstrate that these thinking tokens are crucial for LRM's reasoning performance, while other tokens has minimal impacts. Building on these analyses, we propose two simple yet effective methods to improve LRM's reasoning performance, by delicately leveraging these thinking tokens. Overall, our work provides novel insights into the reasoning mechanisms of LRMs and offers practical ways to improve their reasoning capabilities. The code is available at https://github.com/ChnQ/MI-Peaks.
Authors: Jakub Krajewski, Marcin Chochowski, Daniel Korzekwa
Abstract: Mixture of Experts (MoE) architectures have emerged as pivotal for scaling Large Language Models (LLMs) efficiently. Fine-grained MoE approaches - utilizing more numerous, smaller experts - have demonstrated potential in improving model convergence and quality. This work proposes a set of training recipes and provides a comprehensive empirical evaluation of fine-grained MoE, directly comparing its scaling properties against standard MoE configurations for models with up to 56B total (17B active) parameters. We investigate convergence speed, model performance on downstream benchmarks, and practical training considerations across various setups. Overall, at the largest scale we show that fine-grained MoE achieves better validation loss and higher accuracy across a set of downstream benchmarks. This study offers empirical grounding and practical insights for leveraging fine-grained MoE in the development of future large-scale models.
Authors: Li Zhang, Kevin D. Ashley
Abstract: Large Language Models (LLMs) are increasingly explored for legal argument generation, yet they pose significant risks of manipulation through hallucination and ungrounded persuasion, and often fail to utilize provided factual bases effectively or abstain when arguments are untenable. This paper introduces a novel reflective multi-agent method designed to address these challenges in the context of legally compliant persuasion. Our approach employs specialized agents--a Factor Analyst and an Argument Polisher--in an iterative refinement process to generate 3-ply legal arguments (plaintiff, defendant, rebuttal). We evaluate Reflective Multi-Agent against single-agent, enhanced-prompt single-agent, and non-reflective multi-agent baselines using four diverse LLMs (GPT-4o, GPT-4o-mini, Llama-4-Maverick-17b-128e, Llama-4-Scout-17b-16e) across three legal scenarios: "arguable", "mismatched", and "non-arguable". Results demonstrate Reflective Multi-Agent's significant superiority in successful abstention (preventing generation when arguments cannot be grounded), marked improvements in hallucination accuracy (reducing fabricated and misattributed factors), particularly in "non-arguable" scenarios, and enhanced factor utilization recall (improving the use of provided case facts). These findings suggest that structured reflection within a multi-agent framework offers a robust computable method for fostering ethical persuasion and mitigating manipulation in LLM-based legal argumentation systems, a critical step towards trustworthy AI in law. Project page: https://lizhang-aiandlaw.github.io/A-Reflective-Multi-Agent-Approach-for-Legal-Argument-Generation/
URLs: https://lizhang-aiandlaw.github.io/A-Reflective-Multi-Agent-Approach-for-Legal-Argument-Generation/
Authors: Sinem Erisken (Independent Researcher), Timothy Gothard (Independent Researcher), Martin Leitgab (Independent Researcher), Ram Potham (Independent Researcher)
Abstract: Traditional AI safety evaluations on isolated LLMs are insufficient as multi-agent AI ensembles become prevalent, introducing novel emergent risks. This paper introduces the Multi-Agent Emergent Behavior Evaluation (MAEBE) framework to systematically assess such risks. Using MAEBE with the Greatest Good Benchmark (and a novel double-inversion question technique), we demonstrate that: (1) LLM moral preferences, particularly for Instrumental Harm, are surprisingly brittle and shift significantly with question framing, both in single agents and ensembles. (2) The moral reasoning of LLM ensembles is not directly predictable from isolated agent behavior due to emergent group dynamics. (3) Specifically, ensembles exhibit phenomena like peer pressure influencing convergence, even when guided by a supervisor, highlighting distinct safety and alignment challenges. Our findings underscore the necessity of evaluating AI systems in their interactive, multi-agent contexts.
Authors: Yang Guo, Yutian Tao, Yifei Ming, Robert D. Nowak, Yingyu Liang
Abstract: Retrieval-augmented generation (RAG) has seen many empirical successes in recent years by aiding the LLM with external knowledge. However, its theoretical aspect has remained mostly unexplored. In this paper, we propose the first finite-sample generalization bound for RAG in in-context linear regression and derive an exact bias-variance tradeoff. Our framework views the retrieved texts as query-dependent noisy in-context examples and recovers the classical in-context learning (ICL) and standard RAG as the limit cases. Our analysis suggests that an intrinsic ceiling on generalization error exists on RAG as opposed to the ICL. Furthermore, our framework is able to model retrieval both from the training data and from external corpora by introducing uniform and non-uniform RAG noise. In line with our theory, we show the sample efficiency of ICL and RAG empirically with experiments on common QA benchmarks, such as Natural Questions and TriviaQA.
Authors: Mengdi Jia, Zekun Qi, Shaochen Zhang, Wenyao Zhang, Xinqiang Yu, Jiawei He, He Wang, Li Yi
Abstract: Spatial reasoning is a key aspect of cognitive psychology and remains a major bottleneck for current vision-language models (VLMs). While extensive research has aimed to evaluate or improve VLMs' understanding of basic spatial relations, such as distinguishing left from right, near from far, and object counting, these tasks represent only the most fundamental level of spatial reasoning. In this work, we introduce OmniSpatial, a comprehensive and challenging benchmark for spatial reasoning, grounded in cognitive psychology. OmniSpatial covers four major categories: dynamic reasoning, complex spatial logic, spatial interaction, and perspective-taking, with 50 fine-grained subcategories. Through Internet data crawling and careful manual annotation, we construct over 1.5K question-answer pairs. Extensive experiments show that both open- and closed-source VLMs, as well as existing reasoning and spatial understanding models, exhibit significant limitations in comprehensive spatial understanding. We further analyze failure cases and propose potential directions for future research.
Authors: Wei Chow, Yuan Gao, Linfeng Li, Xian Wang, Qi Xu, Hang Song, Lingdong Kong, Ran Zhou, Yi Zeng, Yidong Cai, Botian Jiang, Shilin Xu, Jiajun Zhang, Minghui Qiu, Xiangtai Li, Tianshu Yang, Siliang Tang, Juncheng Li
Abstract: Semantic retrieval is crucial for modern applications yet remains underexplored in current research. Existing datasets are limited to single languages, single images, or singular retrieval conditions, often failing to fully exploit the expressive capacity of visual information as evidenced by maintained performance when images are replaced with captions. However, practical retrieval scenarios frequently involve interleaved multi-condition queries with multiple images. Hence, this paper introduces MERIT, the first multilingual dataset for interleaved multi-condition semantic retrieval, comprising 320,000 queries with 135,000 products in 5 languages, covering 7 distinct product categories. Extensive experiments on MERIT identify existing models's limitation: focusing solely on global semantic information while neglecting specific conditional elements in queries. Consequently, we propose Coral, a novel fine-tuning framework that adapts pre-trained MLLMs by integrating embedding reconstruction to preserve fine-grained conditional elements and contrastive learning to extract comprehensive global semantics. Experiments demonstrate that Coral achieves a 45.9% performance improvement over conventional approaches on MERIT, with strong generalization capabilities validated across 8 established retrieval benchmarks. Collectively, our contributions - a novel dataset, identification of critical limitations in existing approaches, and an innovative fine-tuning framework - establish a foundation for future research in interleaved multi-condition semantic retrieval.
Authors: Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, Yatian Pang, Li Yuan
Abstract: Although existing unified models achieve strong performance in vision-language understanding and text-to-image generation, they remain limited in addressing image perception and manipulation -- capabilities increasingly demanded in practical applications. Recently, OpenAI introduced the powerful GPT-4o-Image model, which showcases advanced capabilities in comprehensive image perception and manipulation, sparking widespread interest. Through carefully designed experiments, we observe that GPT-4o-Image likely relies on semantic encoders rather than VAEs for feature extraction, despite VAEs being commonly regarded as crucial for image manipulation tasks. Inspired by this insight, we propose UniWorld, a unified generative framework built upon semantic features extracted from powerful multimodal large language models and contrastive semantic encoders. Using only 2.7M training data, UniWorld achieves impressive performance across diverse tasks, including image understanding, generation, manipulation, and perception. We fully open-source the UniWorld framework, including model weights, training and evaluation scripts, and datasets to promote reproducibility and further research.
Authors: Arij Riabi, Beno\^it Sagot, Djam\'e Seddah
Abstract: Recent impressive improvements in NLP, largely based on the success of contextual neural language models, have been mostly demonstrated on at most a couple dozen high-resource languages. Building language models and, more generally, NLP systems for non-standardized and low-resource languages remains a challenging task. In this work, we focus on North-African colloquial dialectal Arabic written using an extension of the Latin script, called NArabizi, found mostly on social media and messaging communication. In this low-resource scenario with data displaying a high level of variability, we compare the downstream performance of a character-based language model on part-of-speech tagging and dependency parsing to that of monolingual and multilingual models. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre-trained on large multilingual and monolingual models. Confirming these results a on much larger data set of noisy French user-generated content, we argue that such character-based language models can be an asset for NLP in low-resource and high language variability set-tings.
Authors: Jue Wang
Abstract: While contrastive learning greatly advances the representation of sentence embeddings, it is still limited by the size of the existing sentence datasets. In this paper, we present TransAug (Translate as Augmentation), which provide the first exploration of utilizing translated sentence pairs as data augmentation for text, and introduce a two-stage paradigm to advances the state-of-the-art sentence embeddings. Instead of adopting an encoder trained in other languages setting, we first distill a Chinese encoder from a SimCSE encoder (pretrained in English), so that their embeddings are close in semantic space, which can be regraded as implicit data augmentation. Then, we only update the English encoder via cross-lingual contrastive learning and frozen the distilled Chinese encoder. Our approach achieves a new state-of-art on standard semantic textual similarity (STS), outperforming both SimCSE and Sentence-T5, and the best performance in corresponding tracks on transfer tasks evaluated by SentEval.
Authors: Thanh-Dung Le, Philippe Jouvet, Rita Noumeir
Abstract: Transformer-based models have shown outstanding results in natural language processing but face challenges in applications like classifying small-scale clinical texts, especially with constrained computational resources. This study presents a customized Mixture of Expert (MoE) Transformer models for classifying small-scale French clinical texts at CHU Sainte-Justine Hospital. The MoE-Transformer addresses the dual challenges of effective training with limited data and low-resource computation suitable for in-house hospital use. Despite the success of biomedical pre-trained models such as CamemBERT-bio, DrBERT, and AliBERT, their high computational demands make them impractical for many clinical settings. Our MoE-Transformer model not only outperforms DistillBERT, CamemBERT, FlauBERT, and Transformer models on the same dataset but also achieves impressive results: an accuracy of 87\%, precision of 87\%, recall of 85\%, and F1-score of 86\%. While the MoE-Transformer does not surpass the performance of biomedical pre-trained BERT models, it can be trained at least 190 times faster, offering a viable alternative for settings with limited data and computational resources. Although the MoE-Transformer addresses challenges of generalization gaps and sharp minima, demonstrating some limitations for efficient and accurate clinical text classification, this model still represents a significant advancement in the field. It is particularly valuable for classifying small French clinical narratives within the privacy and constraints of hospital-based computational resources.
Authors: Yangning Li, Qingsong Lv, Tianyu Yu, Yinghui Li, Xuming Hu, Wenhao Jiang, Hai-Tao Zheng, Hui Wang
Abstract: Entity Set Expansion (ESE) aims to identify new entities belonging to the same semantic class as the given set of seed entities. Traditional methods solely relied on positive seed entities to represent the target fine-grained semantic class, rendering them tough to represent ultra-fine-grained semantic classes. Specifically, merely relying on positive seed entities leads to two inherent shortcomings: (i) Ambiguity among ultra-fine-grained semantic classes. (ii) Inability to define ``unwanted'' semantics. Hence, previous ESE methods struggle to address the ultra-fine-grained ESE (Ultra-ESE) task. To solve this issue, we first introduce negative seed entities in the inputs, which jointly describe the ultra-fine-grained semantic class with positive seed entities. Negative seed entities eliminate the semantic ambiguity by providing a contrast between positive and negative attributes. Meanwhile, it provides a straightforward way to express ``unwanted''. To assess model performance in Ultra-ESE and facilitate further research, we also constructed UltraWiki, the first large-scale dataset tailored for Ultra-ESE. UltraWiki encompasses 50,973 entities and 394,097 sentences, alongside 236 ultra-fine-grained semantic classes, where each class is represented with 3-5 positive and negative seed entities. Moreover, a retrieval-based framework RetExpan and a generation-based framework GenExpan are proposed to provide powerful baselines for Ultra-ESE. Additionally, we devised two strategies to enhance models' comprehension of ultra-fine-grained entities' semantics: contrastive learning and chain-of-thought reasoning. Extensive experiments confirm the effectiveness of our proposed strategies and also reveal that there remains a large space for improvement in Ultra-ESE.
Authors: Yongyu Mu, Peinan Feng, Zhiquan Cao, Yuzhang Wu, Bei Li, Chenglong Wang, Tong Xiao, Kai Song, Tongran Liu, Chunliang Zhang, Jingbo Zhu
Abstract: In this study, we reveal an in-context learning (ICL) capability of multilingual large language models (LLMs): by translating the input to several languages, we provide Parallel Input in Multiple Languages (PiM) to LLMs, which significantly enhances their comprehension abilities. To test this capability, we design extensive experiments encompassing 8 typical datasets, 7 languages and 8 state-of-the-art multilingual LLMs. Experimental results show that (1) incorporating more languages help PiM surpass the conventional ICL further; (2) even combining with the translations that are inferior to baseline performance can also help. Moreover, by examining the activated neurons in LLMs, we discover a counterintuitive but interesting phenomenon. Contrary to the common thought that PiM would activate more neurons than monolingual input to leverage knowledge learned from diverse languages, PiM actually inhibits neurons and promotes more precise neuron activation especially when more languages are added. This phenomenon aligns with the neuroscience insight about synaptic pruning, which removes less used neural connections, strengthens remainders, and then enhances brain intelligence.
Authors: Deyuan Liu, Zecheng Wang, Bingning Wang, Weipeng Chen, Chunshan Li, Zhiying Tu, Dianhui Chu, Bo Li, Dianbo Sui
Abstract: The rapid proliferation of large language models (LLMs) such as GPT-4 and Gemini underscores the intense demand for resources during their training processes, posing significant challenges due to substantial computational and environmental costs. To alleviate this issue, we propose checkpoint merging in pretraining LLM. This method utilizes LLM checkpoints with shared training trajectories, and is rooted in an extensive search space exploration for the best merging weight via Bayesian optimization. Through various experiments, we demonstrate that: (1) Our proposed methodology exhibits the capacity to augment pretraining, presenting an opportunity akin to obtaining substantial benefits at minimal cost; (2) Our proposed methodology, despite requiring a given held-out dataset, still demonstrates robust generalization capabilities across diverse domains, a pivotal aspect in pretraining.
Authors: Zhuoxuan Jiang, Haoyuan Peng, Shanshan Feng, Fan Li, Dongsheng Li
Abstract: Self-correction is emerging as a promising approach to mitigate the issue of hallucination in Large Language Models (LLMs). To facilitate effective self-correction, recent research has proposed mistake detection as its initial step. However, current literature suggests that LLMs often struggle with reliably identifying reasoning mistakes when using simplistic prompting strategies. To address this challenge, we introduce a unique prompting strategy, termed the Pedagogical Chain-of-Thought (PedCoT), which is specifically designed to guide the identification of reasoning mistakes, particularly mathematical reasoning mistakes. PedCoT consists of pedagogical principles for prompts (PPP) design, two-stage interaction process (TIP) and grounded PedCoT prompts, all inspired by the educational theory of the Bloom Cognitive Model (BCM). We evaluate our approach on two public datasets featuring math problems of varying difficulty levels. The experiments demonstrate that our zero-shot prompting strategy significantly outperforms strong baselines. The proposed method can achieve the goal of reliable mathematical mistake identification and provide a foundation for automatic math answer grading. The results underscore the significance of educational theory, serving as domain knowledge, in guiding prompting strategy design for addressing challenging tasks with LLMs effectively.
Authors: Shester Gueuwou, Xiaodan Du, Greg Shakhnarovich, Karen Livescu
Abstract: A persistent challenge in sign language video processing, including the task of sign to written language translation, is how we learn representations of sign language in an effective and efficient way that preserves the important attributes of these languages, while remaining invariant to irrelevant visual differences. Informed by the nature and linguistics of signed languages, our proposed method focuses on just the most relevant parts in a signing video: the face, hands and body pose of the signer. However, instead of fully relying on pose estimation from off-the-shelf pose tracking models, which have inconsistent performance for hands and faces, we propose to learn a representation of the complex handshapes and facial expressions of sign languages in a self-supervised fashion. Our approach is based on learning from individual frames (rather than video sequences) and is therefore much more efficient than prior work on sign language pre-training. Compared to a recent model that established a new state of the art in sign language translation on the How2Sign dataset, our approach yields similar translation performance, using less than 3\% of the compute.
Authors: {\DJ}or{\dj}e Klisura, Anthony Rios
Abstract: Text-to-SQL systems empower users to interact with databases using natural language, automatically translating queries into executable SQL code. However, their reliance on database schema information for SQL generation exposes them to significant security vulnerabilities, particularly schema inference attacks that can lead to unauthorized data access or manipulation. In this paper, we introduce a novel zero-knowledge framework for reconstructing the underlying database schema of text-to-SQL models without any prior knowledge of the database. Our approach systematically probes text-to-SQL models with specially crafted questions and leverages a surrogate GPT-4 model to interpret the outputs, effectively uncovering hidden schema elements -- including tables, columns, and data types. We demonstrate that our method achieves high accuracy in reconstructing table names, with F1 scores of up to .99 for generative models and .78 for fine-tuned models, underscoring the severity of schema leakage risks. We also show that our attack can steal prompt information in non-text-to-SQL models. Furthermore, we propose a simple protection mechanism for generative models and empirically show its limitations in mitigating these attacks.
Authors: Yi-Sheng Hsu, Nils Feldhus, Sherzod Hakimov
Abstract: Free-text rationales justify model decisions in natural language and thus become likable and accessible among approaches to explanation across many tasks. However, their effectiveness can be hindered by misinterpretation and hallucination. As a perturbation test, we investigate how large language models (LLMs) perform rationale generation under the effects of readability level control, i.e., being prompted for an explanation targeting a specific expertise level, such as sixth grade or college. We find that explanations are adaptable to such instruction, though the observed distinction between readability levels does not fully match the defined complexity scores according to traditional readability metrics. Furthermore, the generated rationales tend to feature medium level complexity, which correlates with the measured quality using automatic metrics. Finally, our human annotators confirm a generally satisfactory impression on rationales at all readability levels, with high-school-level readability being most commonly perceived and favored.
Authors: Md Nayem Uddin, Amir Saeidi, Divij Handa, Agastya Seth, Tran Cao Son, Eduardo Blanco, Steven R. Corman, Chitta Baral
Abstract: This paper introduces UnSeenTimeQA, a novel data contamination-free time-sensitive question-answering (TSQA) benchmark. It differs from existing TSQA benchmarks by avoiding web-searchable queries grounded in the real world. We present a series of time-sensitive event scenarios based on synthetically generated facts. It requires large language models (LLMs) to engage in genuine temporal reasoning without depending on the factual knowledge acquired during the pre-training phase. Our data generation framework enables on-demand generation of new samples, mitigating the risk of data leakage. We designed three types of time-sensitive questions to test LLMs' temporal reasoning abilities over sequential and parallel event occurrences. Our evaluation of five LLMs on synthetic fact-based TSQA reveals mixed results: while they perform well on simpler subsets, their overall performance remains inferior as compared to real world fact-based TSQA. Error analysis indicates that LLMs face difficulties in reasoning over long-range event dependencies and parallel events.
Authors: Rachneet Sachdeva, Yixiao Song, Mohit Iyyer, Iryna Gurevych
Abstract: Long-form question answering (LFQA) aims to provide thorough and in-depth answers to complex questions, enhancing comprehension. However, such detailed responses are prone to hallucinations and factual inconsistencies, challenging their faithful evaluation. This work introduces HaluQuestQA, the first hallucination dataset with localized error annotations for human-written and model-generated LFQA answers. HaluQuestQA comprises 698 QA pairs with 1.8k span-level error annotations for five different error types by expert annotators, along with preference judgments. Using our collected data, we thoroughly analyze the shortcomings of long-form answers and find that they lack comprehensiveness and provide unhelpful references. We train an automatic feedback model on this dataset that predicts error spans with incomplete information and provides associated explanations. Finally, we propose a prompt-based approach, Error-informed refinement, that uses signals from the learned feedback model to refine generated answers, which we show reduces errors and improves answer quality across multiple models. Furthermore, humans find answers generated by our approach comprehensive and highly prefer them (84%) over the baseline answers.
Authors: Liang Shi, Zhengju Tang, Nan Zhang, Xiaotong Zhang, Zhi Yang
Abstract: With the development of the Large Language Models (LLMs), a large range of LLM-based Text-to-SQL(Text2SQL) methods have emerged. This survey provides a comprehensive review of LLM-based Text2SQL studies. We first enumerate classic benchmarks and evaluation metrics. For the two mainstream methods, prompt engineering and finetuning, we introduce a comprehensive taxonomy and offer practical insights into each subcategory. We present an overall analysis of the above methods and various models evaluated on well-known datasets and extract some characteristics. Finally, we discuss the challenges and future directions in this field.
Authors: Yao-Shun Chuang, Chun-Teh Lee, Oluwabunmi Tokede, Guo-Hao Lin, Ryan Brandon, Trung Duong Tran, Xiaoqian Jiang, Muhammad F. Walji
Abstract: This research addresses the issue of missing structured data in dental records by extracting diagnostic information from unstructured text. The updated periodontology classification system's complexity has increased incomplete or missing structured diagnoses. To tackle this, we use advanced AI and NLP methods, leveraging GPT-4 to generate synthetic notes for fine-tuning a RoBERTa model. This significantly enhances the model's ability to understand medical and dental language. We evaluated the model using 120 randomly selected clinical notes from two datasets, demonstrating its improved diagnostic extraction accuracy. The results showed high accuracy in diagnosing periodontal status, stage, and grade, with Site 1 scoring 0.99 and Site 2 scoring 0.98. In the subtype category, Site 2 achieved perfect scores, outperforming Site 1. This method enhances extraction accuracy and broadens its use across dental contexts. The study underscores AI and NLP's transformative impact on healthcare delivery and management. Integrating AI and NLP technologies enhances documentation and simplifies administrative tasks by precisely extracting complex clinical information. This approach effectively addresses challenges in dental diagnostics. Using synthetic training data from LLMs optimizes the training process, improving accuracy and efficiency in identifying periodontal diagnoses from clinical notes. This innovative method holds promise for broader healthcare applications, potentially improving patient care quality.
Authors: Dingwei Chen, Feiteng Fang, Shiwen Ni, Feng Liang, Xiping Hu, Ahmadreza Argha, Hamid Alinejad-Rokny, Min Yang, Chengming Li
Abstract: Large Language Models (LLMs) have demonstrated exceptional performance across various natural language processing tasks. However, they occasionally generate inaccurate and counterfactual outputs, a phenomenon commonly referred to as "hallucinations''. To tackle this issue, recent studies have explored contrastive decoding between the original model and an amateur model with induced hallucination, showing promising results. Nevertheless, this approach can disrupt the original LLM's output distribution due to coarse contrast and simple subtraction operations, potentially leading to errors. In this paper, we introduce a novel contrastive decoding framework, termed LOL (LOwer Layer Matters). Unlike prior methods that focus solely on the final layer, our approach integrates contrastive information from lower layers to enable multi-layer fusion during contrastive decoding. Additionally, we incorporate a truthfulness refocused module that leverages instruction guidance to further improve truthfulness in contrastive decoding. Extensive experiments on four publicly available datasets demonstrate that the LOL framework significantly mitigates hallucination while outperforming existing baselines in most cases. For reproducibility, we will release our code and data upon acceptance.
Authors: Yahan Li, Yi Wang, Yi Chang, Yuan Wu
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across a range of natural language processing (NLP) tasks, capturing the attention of both practitioners and the broader public. A key question that now preoccupies the AI community concerns the capabilities and limitations of these models, with trustworthiness emerging as a central issue, particularly as LLMs are increasingly applied in sensitive fields like healthcare and finance, where errors can have serious consequences. However, most previous studies on the trustworthiness of LLMs have been limited to a single language, typically the predominant one in the dataset, such as English. In response to the growing global deployment of LLMs, we introduce XTRUST, the first comprehensive multilingual trustworthiness benchmark. XTRUST encompasses a diverse range of topics, including illegal activities, hallucination, out-of-distribution (OOD) robustness, physical and mental health, toxicity, fairness, misinformation, privacy, and machine ethics, across 10 different languages. Using XTRUST, we conduct an empirical evaluation of the multilingual trustworthiness of five widely used LLMs, offering an in-depth analysis of their performance across languages and tasks. Our results indicate that many LLMs struggle with certain low-resource languages, such as Arabic and Russian, highlighting the considerable room for improvement in the multilingual trustworthiness of current language models. The code is available at https://github.com/LluckyYH/XTRUST.
Authors: Francesco Verdini, Pierfrancesco Melucci, Stefano Perna, Francesco Cariaggi, Marco Gaido, Sara Papi, Szymon Mazurek, Marek Kasztelnik, Luisa Bentivogli, S\'ebastien Brati\`eres, Paolo Merialdo, Simone Scardapane
Abstract: The remarkable performance achieved by Large Language Models (LLM) has driven research efforts to leverage them for a wide range of tasks and input modalities. In speech-to-text (S2T) tasks, the emerging solution consists of projecting the output of the encoder of a Speech Foundational Model (SFM) into the LLM embedding space through an adapter module. However, no work has yet investigated how much the downstream-task performance depends on each component (SFM, adapter, LLM) nor whether the best design of the adapter depends on the chosen SFM and LLM. To fill this gap, we evaluate the combination of 5 adapter modules, 2 LLMs (Mistral and Llama), and 2 SFMs (Whisper and SeamlessM4T) on two widespread S2T tasks, namely Automatic Speech Recognition and Speech Translation. Our results demonstrate that the SFM plays a pivotal role in downstream performance, while the adapter choice has moderate impact and depends on the SFM and LLM.
Authors: Shota Takashiro, Takeshi Kojima, Andrew Gambardella, Qi Cao, Yusuke Iwasawa, Yutaka Matsuo
Abstract: As large language models (LLMs) are applied across diverse domains, the ability to selectively unlearn specific information is becoming increasingly essential. For instance, LLMs are expected to selectively provide confidential information to authorized internal users, such as employees or trusted partners, while withholding it from external users, including the general public and unauthorized entities. Therefore, we propose a novel method termed ``in-context knowledge unlearning'', which enables the model to selectively forget information in test-time based on the query context. Our method fine-tunes pre-trained LLMs to enable prompt unlearning of target knowledge within the context, while preserving unrelated information. Experiments on TOFU, AGE and RWKU datasets using Llama2-7B/13B and Mistral-7B models demonstrate that our method achieves up to 95% forget accuracy while retaining 80% of unrelated knowledge, significantly outperforming baselines in both in-domain and out-of-domain scenarios. Further investigation of the model's internal behavior revealed that while fine-tuned LLMs generate correct predictions in the middle layers and preserve them up to the final layer. However, the decision to forget is made only at the last layer, i.e. ``LLMs pretend to forget''. Our findings offer valuable insight into the improvement of the robustness of the unlearning mechanisms in LLMs, laying a foundation for future research in the field.
Authors: Yu Ying Chiu, Liwei Jiang, Bill Yuchen Lin, Chan Young Park, Shuyue Stella Li, Sahithya Ravi, Mehar Bhatia, Maria Antoniak, Yulia Tsvetkov, Vered Shwartz, Yejin Choi
Abstract: Robust, diverse, and challenging cultural knowledge benchmarks are essential for measuring our progress towards making LMs that are helpful across diverse cultures. We introduce CulturalBench: a set of 1,696 human-written and human-verified questions to assess LMs' cultural knowledge, covering 45 global regions including underrepresented ones like Bangladesh, Zimbabwe, and Peru. Questions are each verified by five independent annotators and span 17 diverse topics ranging from food preferences to greeting etiquette. We construct CulturalBench using methods inspired by Human-AI Red-Teaming. Compared to human performance (92.4% accuracy), the hard version of CulturalBench is challenging even for the best-performing frontier LMs, ranging from 28.7% to 61.5% in accuracy. We find that LMs often struggle with tricky questions that have multiple correct answers (e.g., What utensils do the Chinese usually use?), revealing a tendency to overfit to a single answer. Our results indicate that GPT-4o substantially outperform other models across cultures, besting local providers (e.g., Mistral on European culture and DeepSeek on Chinese culture). Across the board, models under-perform on questions related to North Africa, South America and Middle East.
Authors: Wenxuan Wang, Kuiyi Gao, Youliang Yuan, Jen-tse Huang, Qiuzhi Liu, Shuai Wang, Wenxiang Jiao, Zhaopeng Tu
Abstract: Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows, making them the focus in recent years. Despite their remarkable capability to generate diverse and vivid images, considerable efforts are being made to prevent the generation of harmful content, such as abusive, violent, or pornographic material. To assess the safety of existing models, we introduce a novel jailbreaking method called Chain-of-Jailbreak (CoJ) attack, which compromises image generation models through a step-by-step editing process. Specifically, for malicious queries that cannot bypass the safeguards with a single prompt, we intentionally decompose the query into multiple sub-queries. The image generation models are then prompted to generate and iteratively edit images based on these sub-queries. To evaluate the effectiveness of our CoJ attack method, we constructed a comprehensive dataset, CoJ-Bench, encompassing nine safety scenarios, three types of editing operations, and three editing elements. Experiments on four widely-used image generation services provided by GPT-4V, GPT-4o, Gemini 1.5 and Gemini 1.5 Pro, demonstrate that our CoJ attack method can successfully bypass the safeguards of models for over 60% cases, which significantly outperforms other jailbreaking methods (i.e., 14%). Further, to enhance these models' safety against our CoJ attack method, we also propose an effective prompting-based method, Think Twice Prompting, that can successfully defend over 95% of CoJ attack. We release our dataset and code to facilitate the AI safety research.
Authors: Yahan Li, Tingyu Xia, Yi Chang, Yuan Wu
Abstract: As large language models (LLMs) continue to evolve, efficient evaluation metrics are vital for assessing their ability to compress information and reduce redundancy. While traditional metrics like Matrix Entropy offer valuable insights, they are computationally intensive for large-scale models due to their \( O(n^3) \) time complexity with Singular Value Decomposition (SVD). To mitigate this issue, we introduce the Matrix Nuclear-Norm, which not only serves as a metric to quantify the data compression proficiency of LLM but also provides a convex approximation of matrix rank to capture both predictive discriminability and diversity. By employing the \( L_{1,2}\text{-norm} \) to further approximate the nuclear norm, we can effectively assess the model's information compression capabilities. This approach reduces the time complexity to \( O(n^2) \) and eliminates the need for SVD computation. Consequently, the Matrix Nuclear-Norm achieves speeds 8 to 24 times faster than Matrix Entropy for the CEREBRAS-GPT model as sizes increase from 111M to 6.7B. This performance gap becomes more pronounced with larger models, as validated in tests with other models like Pythia. Additionally, evaluations on benchmarks and model responses confirm that our proposed Matrix Nuclear-Norm is a reliable, scalable, and efficient tool for assessing LLMs' performance, striking a balance between accuracy and computational efficiency. The code is available at https://github.com/MLGroupJLU/MatrixNuclearNorm.
Authors: Emmanouil Zaranis, Giuseppe Attanasio, Sweta Agrawal, Andr\'e F. T. Martins
Abstract: Quality estimation (QE)-the automatic assessment of translation quality-has recently become crucial across several stages of the translation pipeline, from data curation to training and decoding. While QE metrics have been optimized to align with human judgments, whether they encode social biases has been largely overlooked. Biased QE risks favoring certain demographic groups over others, e.g., by exacerbating gaps in visibility and usability. This paper defines and investigates gender bias of QE metrics and discusses its downstream implications for machine translation (MT). Experiments with state-of-the-art QE metrics across multiple domains, datasets, and languages reveal significant bias. When a human entity's gender in the source is undisclosed, masculine-inflected translations score higher than feminine-inflected ones, and gender-neutral translations are penalized. Even when contextual cues disambiguate gender, using context-aware QE metrics leads to more errors in selecting the correct translation inflection for feminine referents than for masculine ones. Moreover, a biased QE metric affects data filtering and quality-aware decoding. Our findings underscore the need for a renewed focus on developing and evaluating QE metrics centered on gender.
Authors: Bokai Hu, Sai Ashish Somayajula, Xin Pan, Pengtao Xie
Abstract: Instruction-fine-tuned large language models (LLMs) under 14B parameters continue to underperform on natural language understanding (NLU) tasks, often trailing smaller models like BERT-base on benchmarks such as GLUE and SuperGLUE. Motivated by the success of reinforcement learning in reasoning tasks (e.g., DeepSeek), we explore Proximal Policy Optimization (PPO) as a framework to improve the NLU capabilities of LLMs. We frame NLU as a reinforcement learning environment, treating token generation as a sequence of actions and optimizing for reward signals based on alignment with ground-truth labels. PPO consistently outperforms supervised fine-tuning, yielding an average improvement of 6.3 points on GLUE, and surpasses zero-shot and few-shot prompting by 38.7 and 26.1 points, respectively. Notably, PPO-tuned models outperform GPT-4o by over 4\% on average across sentiment and natural language inference tasks, including gains of 7.3\% on the Mental Health dataset and 10.9\% on SIGA-nli. This work highlights a promising direction for adapting LLMs to new tasks by reframing them as reinforcement learning problems, enabling learning through simple end-task rewards rather than extensive data curation.
Authors: Anna Sokol, Elizabeth Daly, Michael Hind, David Piorkowski, Xiangliang Zhang, Nuno Moniz, Nitesh Chawla
Abstract: Large language models (LLMs) are powerful tools capable of handling diverse tasks. Comparing and selecting appropriate LLMs for specific tasks requires systematic evaluation methods, as models exhibit varying capabilities across different domains. However, finding suitable benchmarks is difficult given the many available options. This complexity not only increases the risk of benchmark misuse and misinterpretation but also demands substantial effort from LLM users, seeking the most suitable benchmarks for their specific needs. To address these issues, we introduce \texttt{BenchmarkCards}, an intuitive and validated documentation framework that standardizes critical benchmark attributes such as objectives, methodologies, data sources, and limitations. Through user studies involving benchmark creators and users, we show that \texttt{BenchmarkCards} can simplify benchmark selection and enhance transparency, facilitating informed decision-making in evaluating LLMs. Data & Code: https://github.com/SokolAnn/BenchmarkCards
Authors: Eric Elmoznino, Thomas Jiralerspong, Yoshua Bengio, Guillaume Lajoie
Abstract: Compositionality is believed to be fundamental to intelligence. In humans, it underlies the structure of thought, language, and higher-level reasoning. In AI, compositional representations can enable a powerful form of out-of-distribution generalization, in which a model systematically adapts to novel combinations of known concepts. However, while we have strong intuitions about what compositionality is, we lack satisfying formal definitions for it that are measurable and mathematical. Here, we propose such a definition, which we call representational compositionality, that accounts for and extends our intuitions about compositionality. The definition is conceptually simple, quantitative, grounded in algorithmic information theory, and applicable to any representation. Intuitively, representational compositionality states that a compositional representation satisfies three properties. First, it must be expressive. Second, it must be possible to re-describe the representation as a function of discrete symbolic sequences with re-combinable parts, analogous to sentences in natural language. Third, the function that relates these symbolic sequences to the representation, analogous to semantics in natural language, must be simple. Through experiments on both synthetic and real world data, we validate our definition of compositionality and show how it unifies disparate intuitions from across the literature in both AI and cognitive science. We also show that representational compositionality, while theoretically intractable, can be readily estimated using standard deep learning tools. We hope that our definition can inspire the design of novel, theoretically-driven models that better capture the mechanisms of compositional thought. We make our code available at https://github.com/EricElmoznino/complexity_compositionality.
URLs: https://github.com/EricElmoznino/complexity_compositionality.
Authors: Shih-Yang Liu, Maksim Khadkevich, Nai Chit Fung, Charbel Sakr, Chao-Han Huck Yang, Chien-Yi Wang, Saurav Muralidharan, Hongxu Yin, Kwang-Ting Cheng, Jan Kautz, Yu-Chiang Frank Wang, Pavlo Molchanov, Min-Hung Chen
Abstract: While post-training compression techniques effectively reduce the memory footprint, latency, and power consumption of Large Language Models (LLMs), they often result in noticeable accuracy degradation and remain limited by hardware and kernel constraints that restrict supported compression formats ultimately reducing flexibility across a wide range of deployment scenarios. In this work, we propose EoRA, a novel fine-tuning-free method that augments compressed LLMs with low-rank matrices, allowing users to rapidly enhance task-specific performance and freely balance the trade-off between accuracy and computational overhead beyond the constraints of compression formats. EoRA consistently outperforms prior training-free low rank methods in recovering the accuracy of compressed LLMs, achieving notable accuracy improvements (e.g., $\mathbf{10.84\%}$ on ARC-Challenge, $\mathbf{6.74\%}$ on MathQA, and $\mathbf{6.74\%}$ on GSM8K) for LLaMA3-8B compressed to 3-bit. We also introduce an optimized CUDA kernel, accelerating inference by up to 1.4x and reducing memory overhead through quantizing EoRA. Overall, EoRA offers a prompt solution for improving the accuracy of compressed models under varying user requirements, enabling more efficient and flexible deployment of LLMs. Code is available at https://github.com/NVlabs/EoRA.
Authors: Chenghua Huang, Zhizhen Fan, Lu Wang, Fangkai Yang, Pu Zhao, Zeqi Lin, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, Qi Zhang
Abstract: Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences, playing a pivotal role in the success of conversational models like GPT-4, ChatGPT, and Llama 2. A core challenge in employing RLHF lies in training a reliable reward model (RM), which relies on high-quality labels typically provided by human experts or advanced AI system. These methods can be costly and may introduce biases that affect the language model's responses. As language models improve, human input may become less effective in further enhancing their performance. In this paper, we propose Self-Evolved Reward Learning (SER), a novel approach where the RM generates additional training data to iteratively improve itself. We conducted extensive experiments on multiple datasets such as HH-RLHF and UltraFeedback, using models like Mistral and Llama 3, and compare SER against various baselines. Our results demonstrate that even with limited human-annotated data, learning from self-feedback can robustly enhance RM performance, thereby boosting the capabilities of large language models (LLMs). Resources of this paper can be found at https://aka.ms/ser
URLs: https://aka.ms/ser
Authors: Lin Wang, Xiaocui Yang, Shi Feng, Daling Wang, Yifei Zhang, Zhitao Zhang
Abstract: Multimodal conversation, a crucial form of human communication, carries rich emotional content, making the exploration of the causes of emotions within it a research endeavor of significant importance. However, existing research on the causes of emotions typically employs an utterance selection method within a single textual modality to locate causal utterances. This approach remains limited to coarse-grained assessments, lacks nuanced explanations of emotional causation, and demonstrates inadequate capability in identifying multimodal emotional triggers. Therefore, we introduce a task-\textbf{Multimodal Emotion Cause Explanation in Conversation (MECEC)}. This task aims to generate a summary based on the multimodal context of conversations, clearly and intuitively describing the reasons that trigger a given emotion. To adapt to this task, we develop a new dataset (ECEM) based on the MELD dataset. ECEM combines video clips with detailed explanations of character emotions, helping to explore the causal factors behind emotional expression in multimodal conversations. A novel approach, FAME-Net, is further proposed, that harnesses the power of Large Language Models (LLMs) to analyze visual data and accurately interpret the emotions conveyed through facial expressions in videos. By exploiting the contagion effect of facial emotions, FAME-Net effectively captures the emotional causes of individuals engaged in conversations. Our experimental results on the newly constructed dataset show that FAME-Net outperforms several excellent baselines. Code and dataset are available at https://github.com/3222345200/FAME-Net.
Authors: Lindia Tjuatja, Graham Neubig, Tal Linzen, Sophie Hao
Abstract: When comparing the linguistic capabilities of language models (LMs) with humans using LM probabilities, factors such as the length of the sequence and the unigram frequency of lexical items have a significant effect on LM probabilities in ways that humans are largely robust to. Prior works in comparing LM and human acceptability judgments treat these effects uniformly across models, making a strong assumption that models require the same degree of adjustment to control for length and unigram frequency effects. We propose MORCELA, a new linking theory between LM scores and acceptability judgments where the optimal level of adjustment for these effects is estimated from data via learned parameters for length and unigram frequency. We first show that MORCELA outperforms a commonly used linking theory for acceptability - SLOR (Pauls and Klein, 2012; Lau et al. 2017) - across two families of transformer LMs (Pythia and OPT). Furthermore, we demonstrate that the assumed degrees of adjustment in SLOR for length and unigram frequency overcorrect for these confounds, and that larger models require a lower relative degree of adjustment for unigram frequency, though a significant amount of adjustment is still necessary for all models. Finally, our subsequent analysis shows that larger LMs' lower susceptibility to frequency effects can be explained by an ability to better predict rarer words in context.
Authors: Gabriele Oliaro, Zhihao Jia, Daniel Campos, Aurick Qiao
Abstract: Speculative decoding is widely adopted to reduce latency in large language model (LLM) inference by leveraging smaller draft models capable of handling diverse user tasks. However, emerging AI applications, such as LLM-based agents, present unique workload characteristics: instead of diverse independent requests, agentic frameworks typically submit repetitive inference requests, such as multi-agent pipelines performing similar subtasks or self-refinement loops iteratively enhancing outputs. These workloads result in long and highly predictable sequences, which current speculative decoding methods do not effectively exploit. To address this gap, we introduce \emph{SuffixDecoding}, a novel method that utilizes efficient suffix trees to cache long token sequences from prompts and previous outputs. By adaptively speculating more tokens when acceptance likelihood is high and fewer when it is low, SuffixDecoding effectively exploits opportunities for longer speculations while conserving computation when those opportunities are limited. Evaluations on agentic benchmarks, including SWE-Bench and Text-to-SQL, demonstrate that SuffixDecoding achieves speedups of up to 5.3$\times$, outperforming state-of-the-art methods -- 2.8$\times$ faster than model-based approaches like EAGLE-2/3 and 1.9$\times$ faster than model-free approaches such as Token Recycling. SuffixDecoding is open-sourced at https://github.com/snowflakedb/ArcticInference.
Authors: Chuyi Kong, Ziyang Luo, Hongzhan Lin, Zhiyuan Fan, Yaxin Fan, Yuxi Sun, Jing Ma
Abstract: The advanced role-playing capabilities of Large Language Models (LLMs) have enabled rich interactive scenarios, yet existing research in social interactions neglects hallucination while struggling with poor generalizability and implicit character fidelity judgments. To bridge this gap, motivated by human behaviour, we introduce a generalizable and explicit paradigm for uncovering interactive patterns of LLMs across diverse worldviews. Specifically, we first define interactive hallucination through stance transfer, then construct SHARP, a benchmark built by extracting relations from commonsense knowledge graphs and utilizing LLMs' inherent hallucination properties to simulate multi-role interactions. Extensive experiments confirm our paradigm's effectiveness and stability, examine the factors that influence these metrics, and challenge conventional hallucination mitigation solutions. More broadly, our work reveals a fundamental limitation in popular post-training methods for role-playing LLMs: the tendency to obscure knowledge beneath style, resulting in monotonous yet human-like behaviors - interactive hallucination.
Authors: Jingxuan Li, Yuning Yang, Shengqi Yang, Linfan Zhang, Ying Nian Wu
Abstract: The recent progress in Vision-Language Models (VLMs) has broadened the scope of multimodal applications. However, evaluations often remain limited to functional tasks, neglecting abstract dimensions such as personality traits and human values. To address this gap, we introduce Value-Spectrum, a novel Visual Question Answering (VQA) benchmark aimed at assessing VLMs based on Schwartz's value dimensions that capture core human values guiding people's preferences and actions. We design a VLM agent pipeline to simulate video browsing and construct a vector database comprising over 50,000 short videos from TikTok, YouTube Shorts, and Instagram Reels. These videos span multiple months and cover diverse topics, including family, health, hobbies, society, technology, etc. Benchmarking on Value-Spectrum highlights notable variations in how VLMs handle value-oriented content. Beyond identifying VLMs' intrinsic preferences, we also explore the ability of VLM agents to adopt specific personas when explicitly prompted, revealing insights into the adaptability of the model in role-playing scenarios. These findings highlight the potential of Value-Spectrum as a comprehensive evaluation set for tracking VLM preferences in value-based tasks and abilities to simulate diverse personas. The complete code and data are available at: https://github.com/Jeremyyny/Value-Spectrum.
Authors: Manuel Tonneau, Diyi Liu, Niyati Malhotra, Scott A. Hale, Samuel P. Fraiberger, Victor Orozco-Olvera, Paul R\"ottger
Abstract: To address the global challenge of online hate speech, prior research has developed detection models to flag such content on social media. However, due to systematic biases in evaluation datasets, the real-world effectiveness of these models remains unclear, particularly across geographies. We introduce HateDay, the first global hate speech dataset representative of social media settings, constructed from a random sample of all tweets posted on September 21, 2022 and covering eight languages and four English-speaking countries. Using HateDay, we uncover substantial variation in the prevalence and composition of hate speech across languages and regions. We show that evaluations on academic datasets greatly overestimate real-world detection performance, which we find is very low, especially for non-European languages. Our analysis identifies key drivers of this gap, including models' difficulty to distinguish hate from offensive speech and a mismatch between the target groups emphasized in academic datasets and those most frequently targeted in real-world settings. We argue that poor model performance makes public models ill-suited for automatic hate speech moderation and find that high moderation rates are only achievable with substantial human oversight. Our results underscore the need to evaluate detection systems on data that reflects the complexity and diversity of real-world social media.
Authors: Shester Gueuwou, Xiaodan Du, Greg Shakhnarovich, Karen Livescu, Alexander H. Liu
Abstract: Sign language processing has traditionally relied on task-specific models, limiting the potential for transfer learning across tasks. Pre-training methods for sign language have typically focused on either supervised pre-training, which cannot take advantage of unlabeled data, or context-independent (frame or video segment) representations, which ignore the effects of relationships across time in sign language. We introduce SHuBERT (Sign Hidden-Unit BERT), a self-supervised contextual representation model learned from approximately 1,000 hours of American Sign Language video. SHuBERT adapts masked token prediction objectives to multi-stream visual sign language input, learning to predict multiple targets corresponding to clustered hand, face, and body pose streams. SHuBERT achieves state-of-the-art performance across multiple tasks including sign language translation, isolated sign language recognition, and fingerspelling detection.
Authors: Cristiano Chesi
Abstract: A significant debate has emerged in response to a paper written by Steven Piantadosi (Piantadosi, 2023) and uploaded to the LingBuzz platform, the open archive for generative linguistics. Piantadosi's dismissal of Chomsky's approach is ruthless, but generative linguists deserve it. In this paper, I will adopt three idealized perspectives -- computational, theoretical, and experimental -- to focus on two fundamental issues that lend partial support to Piantadosi's critique: (a) the evidence challenging the Poverty of Stimulus (PoS) hypothesis and (b) the notion of simplicity as conceived within mainstream Minimalism. In conclusion, I argue that, to reclaim a central role in language studies, generative linguistics -- representing a prototypical theoretical perspective on language -- needs a serious update leading to (i) more precise, consistent, and complete formalizations of foundational intuitions and (ii) the establishment and utilization of a standardized dataset of crucial empirical evidence to evaluate the theory's adequacy. On the other hand, ignoring the formal perspective leads to major drawbacks in both computational and experimental approaches. Neither descriptive nor explanatory adequacy can be easily achieved without the precise formulation of general principles that can be challenged empirically.
Authors: Jialong Wu, Zhenglin Wang, Linhai Zhang, Yilong Lai, Yulan He, Deyu Zhou
Abstract: Key-Value (KV) cache has become a bottleneck of LLMs for long-context generation. Despite the numerous efforts in this area, the optimization for the decoding phase is generally ignored. However, we believe such optimization is crucial, especially for long-output generation tasks based on the following two observations: (i) Excessive compression during the prefill phase, which requires specific full context impairs the comprehension of the reasoning task; (ii) Deviation of heavy hitters occurs in the reasoning tasks with long outputs. Therefore, SCOPE, a simple yet efficient framework that separately performs KV cache optimization during the prefill and decoding phases, is introduced. Specifically, the KV cache during the prefill phase is preserved to maintain the essential information, while a novel strategy based on sliding is proposed to select essential heavy hitters for the decoding phase. Memory usage and memory transfer are further optimized using adaptive and discontinuous strategies. Extensive experiments on LongGenBench show the effectiveness and generalization of SCOPE and its compatibility as a plug-in to other prefill-only KV compression methods.
Authors: Mengyu Ye, Tatsuki Kuribayashi, Goro Kobayashi, Jun Suzuki
Abstract: Interpreting the internal process of neural models has long been a challenge. This challenge remains relevant in the era of large language models (LLMs) and in-context learning (ICL); for example, ICL poses a new issue of interpreting which example in the few-shot examples contributed to identifying/solving the task. To this end, in this paper, we design synthetic diagnostic tasks of inductive reasoning, inspired by the generalization tests typically adopted in psycholinguistics. Here, most in-context examples are ambiguous w.r.t. their underlying rule, and one critical example disambiguates it. The question is whether conventional input attribution (IA) methods can track such a reasoning process, i.e., identify the influential example, in ICL. Our experiments provide several practical findings; for example, a certain simple IA method works the best, and the larger the model, the generally harder it is to interpret the ICL with gradient-based IA methods.
Authors: Esther Shizgal, Eitan Wagner, Renana Keydar, Omri Abend
Abstract: This work presents a computational approach to analyze character development along the narrative timeline. The analysis characterizes the inner and outer changes the protagonist undergoes within a narrative, and the interplay between them. We consider transcripts of Holocaust survivor testimonies as a test case, each telling the story of an individual in first-person terms. We focus on the survivor's religious trajectory, examining the evolution of their disposition toward religious belief and practice along the testimony. Clustering the resulting trajectories in the dataset, we identify common sequences in the data. Our findings highlight multiple common structures of religiosity across the narratives: in terms of belief, most present a constant disposition, while for practice, most present an oscillating structure, serving as valuable material for historical and sociological research. This work demonstrates the potential of natural language processing techniques for analyzing character evolution through thematic trajectories in narratives.
Authors: Wei Liu, Junlong Li, Xiwen Zhang, Fan Zhou, Yu Cheng, Junxian He
Abstract: Self-evolving trainin--where models iteratively learn from their own outputs--has emerged as a key approach for complex reasoning tasks, addressing the scarcity of high-quality chain-of-thought data. However, its effectiveness in multimodal reasoning, a domain more intricate than text-only reasoning, remains underexplored, and the understanding of critical factors in this training paradigm remains limited. Furthermore, a central challenge for this training method is performance saturation, which impedes further improvements and scalability. Inspired by reinforcement learning (RL), in this paper, we reframe self-evolving training for multimodal reasoning through the lens of RL, identifying three pivotal factors: Training Method, Reward Model, and Prompt Variation. Through systematic analysis, we establish relatively optimal design principles that significantly enhance multimodal reasoning capabilities. Moreover, delving deeper into training dynamics, we uncover the roots of saturation and propose a new automatic balancing mechanism to mitigate this limitation. Building on these insights, we propose M-STAR (Multimodal Self-evolving Training for Reasoning), a framework that achieves consistent performance gains across models of varying sizes and diverse benchmarks. All resources are made publicly available at https://mstar-lmm.github.io.
Authors: Anna Ko{\l}os, Katarzyna Lorenc, Emilia Wi\'snios, Agnieszka Karli\'nska
Abstract: The surge in online content has created an urgent demand for robust detection systems, especially in non-English contexts where current tools demonstrate significant limitations. We present forePLay, a novel Polish language dataset for erotic content detection, featuring over 24k annotated sentences with a multidimensional taxonomy encompassing ambiguity, violence, and social unacceptability dimensions. Our comprehensive evaluation demonstrates that specialized Polish language models achieve superior performance compared to multilingual alternatives, with transformer-based architectures showing particular strength in handling imbalanced categories. The dataset and accompanying analysis establish essential frameworks for developing linguistically-aware content moderation systems, while highlighting critical considerations for extending such capabilities to morphologically complex languages.
Authors: Bairu Hou, Qibin Chen, Jianyu Wang, Guoli Yin, Chong Wang, Nan Du, Ruoming Pang, Shiyu Chang, Tao Lei
Abstract: With the rapid scaling of large language models (LLMs), structured pruning has become a widely used technique to learn efficient, smaller models from larger ones, delivering superior performance compared to training similarly sized models from scratch. In this paper, we move beyond the traditional static pruning approach of determining a fixed pruning mask for a model, and propose a dynamic approach to structured pruning. In our method, the pruning mask is input-dependent and adapts dynamically based on the information described in a user instruction. Our approach, termed "instruction-following pruning", introduces a sparse mask predictor that takes the user instruction as input and dynamically selects the most relevant model parameters for the given task. To identify and activate effective parameters, we jointly optimize the sparse mask predictor and the LLM, leveraging both instruction-following data and the pre-training corpus. Experimental results demonstrate the effectiveness of our approach on a wide range of evaluation benchmarks. For example, our 3B activated model improves over the 3B dense model by 5-8 points of absolute margin on domains such as math and coding, and rivals the performance of a 9B model.
Authors: Yachao Zhao, Bo Wang, Yan Wang, Dongming Zhao, Ruifang He, Yuexian Hou
Abstract: Large Language Models (LLMs) have been shown to exhibit various biases and stereotypes in their generated content. While extensive research has investigated biases in LLMs, prior work has predominantly focused on explicit bias, with minimal attention to implicit bias and the relation between these two forms of bias. This paper presents a systematic framework grounded in social psychology theories to investigate and compare explicit and implicit biases in LLMs. We propose a novel self-reflection-based evaluation framework that operates in two phases: first measuring implicit bias through simulated psychological assessment methods, then evaluating explicit bias by prompting LLMs to analyze their own generated content. Through extensive experiments on advanced LLMs across multiple social dimensions, we demonstrate that LLMs exhibit a substantial inconsistency between explicit and implicit biases: while explicit bias manifests as mild stereotypes, implicit bias exhibits strong stereotypes. We further investigate the underlying factors contributing to this explicit-implicit bias inconsistency, examining the effects of training data scale, model size, and alignment techniques. Experimental results indicate that while explicit bias declines with increased training data and model size, implicit bias exhibits a contrasting upward trend. Moreover, contemporary alignment methods effectively suppress explicit bias but show limited efficacy in mitigating implicit bias.
Authors: Yindu Su, Huike Zou, Lin Sun, Ting Zhang, Haiyang Yang, Liyu Chen, David Lo, Qingheng Zhang, Shuguang Han, Jufeng Chen
Abstract: Product Attribute Value Identification (PAVI) involves identifying attribute values from product profiles, a key task for improving product search, recommendation, and business analytics on e-commerce platforms. However, existing PAVI methods face critical challenges, such as inferring implicit values, handling out-of-distribution (OOD) values, and producing normalized outputs. To address these limitations, we introduce Taxonomy-Aware Contrastive Learning Retrieval (TACLR), the first retrieval-based method for PAVI. TACLR formulates PAVI as an information retrieval task by encoding product profiles and candidate values into embeddings and retrieving values based on their similarity. It leverages contrastive training with taxonomy-aware hard negative sampling and employs adaptive inference with dynamic thresholds. TACLR offers three key advantages: (1) it effectively handles implicit and OOD values while producing normalized outputs; (2) it scales to thousands of categories, tens of thousands of attributes, and millions of values; and (3) it supports efficient inference for high-load industrial deployment. Extensive experiments on proprietary and public datasets validate the effectiveness and efficiency of TACLR. Further, it has been successfully deployed on the real-world e-commerce platform Xianyu, processing millions of product listings daily with frequently updated, large-scale attribute taxonomies. We release the code to facilitate reproducibility and future research at https://github.com/SuYindu/TACLR.
Authors: Tong Liu, Xiao Yu, Wenxuan Zhou, Jindong Gu, Volker Tresp
Abstract: Efficient preference optimization algorithms such as Direct Preference Optimization (DPO) have become a popular approach in aligning large language models (LLMs) with human preferences. These algorithms implicitly treat the LLM as a reward model, and focus on training it to correct misranked preference pairs. However, recent work~\citep{chen2024preference} empirically finds that DPO training \textit{rarely improves these misranked preference pairs}, despite its gradient emphasizing on these cases. We introduce FocalPO, a DPO variant that instead \textit{down-weighs} misranked preference pairs and prioritizes enhancing the model's understanding of pairs that it can already rank correctly. Inspired by Focal Loss used in vision tasks, FocalPO achieves this by adding a modulating factor to dynamically scale DPO loss. Our experiment demonstrates that FocalPO surpasses DPO and its variants on popular benchmarks like Alpaca Eval 2.0 using Mistral-Base-7B and Llama-3-Instruct-8B, with the introduced hyperparameter fixed. Additionally, we empirically reveals how FocalPO affects training on correct and incorrect sample groups, further underscoring its effectiveness.
Authors: Edoardo Cetin, Tianyu Zhao, Yujin Tang
Abstract: We propose a new finetuning method to provide pre-trained large language models (LMs) the ability to scale test-time compute through the diffusion framework. By increasing the number of diffusion steps, we show our finetuned models achieve monotonically increasing accuracy, directly translating to improved performance across downstream tasks. Furthermore, our finetuned models can expertly answer questions on specific topics by integrating powerful guidance techniques, and autonomously determine the compute required for a given problem by leveraging adaptive ODE solvers. Our method is universally applicable to any foundation model pre-trained with a cross-entropy loss and does not modify any of its original weights, fully preserving its strong single-step generation capabilities. We show our method is more effective and fully compatible with traditional finetuning approaches, introducing an orthogonal new direction to unify the strengths of the autoregressive and diffusion frameworks.
Authors: Xin Xu, Qiyun Xu, Tong Xiao, Tianhao Chen, Yuchen Yan, Jiaxin Zhang, Shizhe Diao, Can Yang, Yang Wang
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in solving complex reasoning tasks, particularly in mathematics. However, the domain of physics reasoning presents unique challenges that have received significantly less attention. Existing benchmarks often fall short in evaluating LLMs' abilities on the breadth and depth of undergraduate-level physics, underscoring the need for a comprehensive evaluation. To fill this gap, we introduce UGPhysics, a large-scale and comprehensive benchmark specifically designed to evaluate UnderGraduate-level Physics (UGPhysics) reasoning with LLMs. UGPhysics includes 5,520 undergraduate-level physics problems in both English and Chinese, covering 13 subjects with seven different answer types and four distinct physics reasoning skills, all rigorously screened for data leakage. Additionally, we develop a Model-Assistant Rule-based Judgment (MARJ) pipeline specifically tailored for assessing answer correctness of physics problems, ensuring accurate evaluation. Our evaluation of 31 leading LLMs shows that the highest overall accuracy, 49.8% (achieved by OpenAI-o1-mini), emphasizes the necessity for models with stronger physics reasoning skills, beyond math abilities. We hope UGPhysics, along with MARJ, will drive future advancements in AI for physics reasoning. Codes and data are available at https://github.com/YangLabHKUST/UGPhysics .
Authors: Pierre-Emmanuel Mazar\'e, Gergely Szilvasy, Maria Lomeli, Francisco Massa, Naila Murray, Herv\'e J\'egou, Matthijs Douze
Abstract: Self-attention in transformer models is an incremental associative memory that maps key vectors to value vectors. One way to speed up self-attention is to employ GPU-compatible vector search algorithms based on standard partitioning methods such as k-means. However, such partitioning methods yield poor results in this context because (1) the keys and queries follow different distributions, and (2) the RoPE positional encoding hinders the bucket assignment. This paper introduces Saap (Self-Attention with Asymmetric Partitions), which overcomes these problems. It is an asymmetrical indexing technique that employs distinct partitions for keys and queries, thereby approximating self-attention with a data-adaptive sparsity pattern. It works on pretrained language models and only requires to train (offline) a small query classifier. On a long context Llama 3.1-8b model, with sequences ranging from 100k to 500k tokens, Saap typically reduces by a factor of 20 the fraction of memory that needs to be looked-up, which translates to a time saving of 60\% when compared to FlashAttention-v2.
Authors: Takumi Goto, Yusuke Sakai, Taro Watanabe
Abstract: One of the goals of automatic evaluation metrics in grammatical error correction (GEC) is to rank GEC systems such that it matches human preferences. However, current automatic evaluations are based on procedures that diverge from human evaluation. Specifically, human evaluation derives rankings by aggregating sentence-level relative evaluation results, e.g., pairwise comparisons, using a rating algorithm, whereas automatic evaluation averages sentence-level absolute scores to obtain corpus-level scores, which are then sorted to determine rankings. In this study, we propose an aggregation method for existing automatic evaluation metrics which aligns with human evaluation methods to bridge this gap. We conducted experiments using various metrics, including edit-based metrics, n-gram based metrics, and sentence-level metrics, and show that resolving the gap improves results for the most of metrics on the SEEDA benchmark. We also found that even BERT-based metrics sometimes outperform the metrics of GPT-4. The proposed ranking method is integrated gec-metrics.
Authors: David Sasu, Zehui Wu, Ziwei Gong, Run Chen, Pengyuan Shi, Lin Ai, Julia Hirschberg, Natalie Schluter
Abstract: In this paper, we introduce the Akan Conversation Emotion (ACE) dataset, the first multimodal emotion dialogue dataset for an African language, addressing the significant lack of resources for low-resource languages in emotion recognition research. ACE, developed for the Akan language, contains 385 emotion-labeled dialogues and 6,162 utterances across audio, visual, and textual modalities, along with word-level prosodic prominence annotations. The presence of prosodic labels in this dataset also makes it the first prosodically annotated African language dataset. We demonstrate the quality and utility of ACE through experiments using state-of-the-art emotion recognition methods, establishing solid baselines for future research. We hope ACE inspires further work on inclusive, linguistically and culturally diverse NLP resources.
Authors: Haoyang Li, Xuejia Chen, Zhanchao XU, Darian Li, Nicole Hu, Fei Teng, Yiming Li, Luyu Qiu, Chen Jason Zhang, Qing Li, Lei Chen
Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in natural language processing tasks, such as text generation and semantic understanding. However, their performance on numerical reasoning tasks, such as basic arithmetic, numerical retrieval, and magnitude comparison, remains surprisingly poor. This gap arises from their reliance on surface-level statistical patterns rather than understanding numbers as continuous magnitudes. Existing benchmarks primarily focus on either linguistic competence or structured mathematical problem-solving, neglecting fundamental numerical reasoning required in real-world scenarios. To bridge this gap, we propose NumericBench, a comprehensive benchmark to evaluate six fundamental numerical capabilities: number recognition, arithmetic operations, contextual retrieval, comparison, summary, and logical reasoning. NumericBench includes datasets ranging from synthetic number lists to the crawled real-world data, addressing challenges like long contexts, noise, and multi-step reasoning. Extensive experiments on state-of-the-art LLMs, including GPT-4 and DeepSeek, reveal persistent weaknesses in numerical reasoning, highlighting the urgent need to improve numerically-aware language modeling. The benchmark is released in: https://github.com/TreeAI-Lab/NumericBench.
Authors: Wenxuan Wang, Xiaoyuan Liu, Kuiyi Gao, Jen-tse Huang, Youliang Yuan, Pinjia He, Shuai Wang, Zhaopeng Tu
Abstract: Multimodal Large Language Models (MLLMs) have expanded the capabilities of traditional language models by enabling interaction through both text and images. However, ensuring the safety of these models remains a significant challenge, particularly in accurately identifying whether multimodal content is safe or unsafe-a capability we term safety awareness. In this paper, we introduce MMSafeAware, the first comprehensive multimodal safety awareness benchmark designed to evaluate MLLMs across 29 safety scenarios with 1500 carefully curated image-prompt pairs. MMSafeAware includes both unsafe and over-safety subsets to assess models abilities to correctly identify unsafe content and avoid over-sensitivity that can hinder helpfulness. Evaluating nine widely used MLLMs using MMSafeAware reveals that current models are not sufficiently safe and often overly sensitive; for example, GPT-4V misclassifies 36.1% of unsafe inputs as safe and 59.9% of benign inputs as unsafe. We further explore three methods to improve safety awareness-prompting-based approaches, visual contrastive decoding, and vision-centric reasoning fine-tuning-but find that none achieve satisfactory performance. Our findings highlight the profound challenges in developing MLLMs with robust safety awareness, underscoring the need for further research in this area. All the code and data will be publicly available to facilitate future research.
Authors: Weikai Lu, Hao Peng, Huiping Zhuang, Cen Chen, Ziqian Zeng
Abstract: Multimodal Large Language Models (MLLMs) have serious security vulnerabilities.While safety alignment using multimodal datasets consisting of text and data of additional modalities can effectively enhance MLLM's security, it is costly to construct these datasets. Existing low-resource security alignment methods, including textual alignment, have been found to struggle with the security risks posed by additional modalities. To address this, we propose Synthetic Embedding augmented safety Alignment (SEA), which optimizes embeddings of additional modality through gradient updates to expand textual datasets. This enables multimodal safety alignment training even when only textual data is available. Extensive experiments on image, video, and audio-based MLLMs demonstrate that SEA can synthesize a high-quality embedding on a single RTX3090 GPU within 24 seconds. SEA significantly improves the security of MLLMs when faced with threats from additional modalities. To assess the security risks introduced by video and audio, we also introduced a new benchmark called VA-SafetyBench. High attack success rates across multiple MLLMs validate its challenge. Our code and data will be available at https://github.com/ZeroNLP/SEA.
Authors: Junhui He, Junna Xing, Nan Wang, Rui Xu, Shangyu Wu, Peng Zhou, Qiang Liu, Chun Jason Xue, Qingan Li
Abstract: Long context large language models (LLMs) pose significant challenges for efficient serving due to the large memory footprint and high access overhead of KV cache. Retrieval-based KV cache reduction methods can mitigate these challenges, typically by offloading the complete KV cache to CPU and retrieving necessary tokens on demand during inference. However, these methods still suffer from unsatisfactory accuracy degradation and extra retrieval overhead. To address these limitations, this paper proposes A$^2$ATS, a novel retrieval-based KV cache reduction method. A$^2$ATS aims to obtain an accurate approximation of attention scores by applying the vector quantization technique to key states, thereby enabling efficient and precise retrieval of the top-K tokens. First, we propose Windowed Rotary Position Embedding, which decouples the positional dependency from query and key states after position embedding. Then, we propose query-aware vector quantization that optimizes the objective of attention score approximation directly. Finally, we design the heterogeneous inference architecture for KV cache offloading, enabling long context serving with larger batch sizes. Experimental results demonstrate that A$^2$ATS can achieve a lower performance degradation with similar or lower overhead compared to existing methods, thereby increasing long context serving throughput by up to $2.7 \times$.
Authors: George-Kirollos Saad, Scott Sanner
Abstract: Query-driven recommendation with unknown items poses a challenge for users to understand why certain items are appropriate for their needs. Query-driven Contrastive Summarization (QCS) is a methodology designed to address this issue by leveraging language-based item descriptions to clarify contrasts between them. However, existing state-of-the-art contrastive summarization methods such as STRUM-LLM fall short of this goal. To overcome these limitations, we introduce Q-STRUM Debate, a novel extension of STRUM-LLM that employs debate-style prompting to generate focused and contrastive summarizations of item aspects relevant to a query. Leveraging modern large language models (LLMs) as powerful tools for generating debates, Q-STRUM Debate provides enhanced contrastive summaries. Experiments across three datasets demonstrate that Q-STRUM Debate yields significant performance improvements over existing methods on key contrastive summarization criteria, thus introducing a novel and performant debate prompting methodology for QCS.
Authors: Fangming Cui, Jan Fong, Rongfei Zeng, Xinmei Tian, Jun Yu
Abstract: Prompt learning has emerged as a promising method for adapting pre-trained visual-language models (VLMs) to a range of downstream tasks. While optimizing the context can be effective for improving performance on specific tasks, it can often lead to poor generalization performance on unseen classes or datasets sampled from different distributions. It may be attributed to the fact that textual prompts tend to overfit downstream data distributions, leading to the forgetting of generalized knowledge derived from hand-crafted prompts. In this paper, we propose a novel method called Similarity Paradigm with Textual Regularization (SPTR) for prompt learning without forgetting. SPTR is a two-pronged design based on hand-crafted prompts that is an inseparable framework. 1) To avoid forgetting general textual knowledge, we introduce the optimal transport as a textual regularization to finely ensure approximation with hand-crafted features and tuning textual features. 2) In order to continuously unleash the general ability of multiple hand-crafted prompts, we propose a similarity paradigm for natural alignment score and adversarial alignment score to improve model robustness for generalization. Both modules share a common objective in addressing generalization issues, aiming to maximize the generalization capability derived from multiple hand-crafted prompts. Four representative tasks (i.e., non-generalization few-shot learning, base-to-novel generalization, cross-dataset generalization, domain generalization) across 11 datasets demonstrate that SPTR outperforms existing prompt learning methods.
Authors: Leena Mathur, Marian Qian, Paul Pu Liang, Louis-Philippe Morency
Abstract: Social reasoning abilities are crucial for AI systems to effectively interpret and respond to multimodal human communication and interaction within social contexts. We introduce SOCIAL GENOME, the first benchmark for fine-grained, grounded social reasoning abilities of multimodal models. SOCIAL GENOME contains 272 videos of interactions and 1,486 human-annotated reasoning traces related to inferences about these interactions. These traces contain 5,777 reasoning steps that reference evidence from visual cues, verbal cues, vocal cues, and external knowledge (contextual knowledge external to videos). SOCIAL GENOME is also the first modeling challenge to study external knowledge in social reasoning. SOCIAL GENOME computes metrics to holistically evaluate semantic and structural qualities of model-generated social reasoning traces. We demonstrate the utility of SOCIAL GENOME through experiments with state-of-the-art models, identifying performance gaps and opportunities for future research to improve the grounded social reasoning abilities of multimodal models.
Authors: Junyang Wang, Haiyang Xu, Xi Zhang, Ming Yan, Ji Zhang, Fei Huang, Jitao Sang
Abstract: The exponential rise in mobile device usage necessitates streamlined automation for effective task management, yet many AI frameworks fall short due to inadequate operational expertise. While manually written knowledge can bridge this gap, it is often burdensome and inefficient. We introduce Mobile-Agent-V, an innovative framework that utilizes video as a guiding tool to effortlessly and efficiently inject operational knowledge into mobile automation processes. By deriving knowledge directly from video content, Mobile-Agent-V eliminates manual intervention, significantly reducing the effort and time required for knowledge acquisition. To rigorously evaluate this approach, we propose Mobile-Knowledge, a benchmark tailored to assess the impact of external knowledge on mobile agent performance. Our experimental findings demonstrate that Mobile-Agent-V enhances performance by 36% compared to existing methods, underscoring its effortless and efficient advantages in mobile automation.
Authors: Boxuan Zhang, Ruqi Zhang
Abstract: Large language models (LLMs) excel in many tasks but struggle to accurately quantify uncertainty in their generated responses. This limitation makes it challenging to detect misinformation and ensure reliable decision-making. Existing uncertainty quantification (UQ) methods for LLMs are primarily prompt-wise rather than response-wise, often requiring multiple response samples, which incurs high computational costs. Moreover, LLMs have been shown to be overconfident, particularly when using reasoning steps to derive their answers. In this work, we propose CoT-UQ, a response-wise UQ framework that integrates LLMs' inherent reasoning capabilities through Chain-of-Thought (CoT) into the UQ process. CoT-UQ captures critical information during inference by extracting keywords from each reasoning step and assessing their importance to the final answer. This key reasoning information is then aggregated to produce a final uncertainty estimate. We conduct extensive experiments based on Llama Family with model sizes varying from 8B to 13B across logical and mathematical reasoning tasks. Experimental results demonstrate that CoT-UQ significantly outperforms existing UQ methods, achieving an average improvement of 5.9% AUROC compared to current UQ methods. The code is available at: https://github.com/ZBox1005/CoT-UQ.
Authors: Hongqiu Wu, Weiqi Wu, Tianyang Xu, Jiameng Zhang, Hai Zhao
Abstract: LLM-based Interactive Drama is a novel AI-based dialogue scenario, where the user (i.e. the player) plays the role of a character in the story, has conversations with characters played by LLM agents, and experiences an unfolding story. This paper begins with understanding interactive drama from two aspects: Immersion, the player's feeling of being present in the story, and Agency, the player's ability to influence the story world. Both are crucial to creating an enjoyable interactive experience, while they have been underexplored in previous work. To enhance these two aspects, we first propose Playwriting-guided Generation, a novel method that helps LLMs craft dramatic stories with substantially improved structures and narrative quality. Additionally, we introduce Plot-based Reflection for LLM agents to refine their reactions to align with the player's intentions. Our evaluation relies on human judgment to assess the gains of our methods in terms of immersion and agency.
Authors: Xueguang Ma, Xi Victoria Lin, Barlas Oguz, Jimmy Lin, Wen-tau Yih, Xilun Chen
Abstract: Large language models (LLMs) have demonstrated strong effectiveness and robustness while fine-tuned as dense retrievers. However, their large parameter size brings significant inference time computational challenges, including high encoding costs for large-scale corpora and increased query latency, limiting their practical deployment. While smaller retrievers offer better efficiency, they often fail to generalize effectively with limited supervised fine-tuning data. In this work, we introduce DRAMA, a training framework that leverages LLMs to train smaller generalizable dense retrievers. In particular, we adopt pruned LLMs as the backbone and train on diverse LLM-augmented data in a single-stage contrastive learning setup. Experiments show that DRAMA offers better multilingual and long-context capabilities than traditional encoder-based retrievers, and achieves strong performance across multiple tasks and languages. These highlight the potential of connecting the training of smaller retrievers with the growing advancements in LLMs, bridging the gap between efficiency and generalization.
Authors: Nathan Roll
Abstract: Large language models (LLMs) showcase increasingly impressive English benchmark scores, however their performance profiles remain inconsistent across multilingual settings. To address this gap, we introduce PolyPrompt, a novel, parameter-efficient framework for enhancing the multilingual capabilities of LLMs. Our method learns a set of trigger tokens for each language through a gradient-based search, identifying the input query's language and selecting the corresponding trigger tokens which are prepended to the prompt during inference. We perform experiments on two ~1 billion parameter models, with evaluations on the global MMLU benchmark across fifteen typologically and resource diverse languages, demonstrating accuracy gains of 3.7%-19.9% compared to naive and translation-pipeline baselines.
Authors: Yifan Zhang, Wenyu Du, Dongming Jin, Jie Fu, Zhi Jin
Abstract: Chain-of-thought (CoT) significantly enhances the performance of large language models (LLMs) across a wide range of tasks, and prior research shows that CoT can theoretically increase expressiveness. However, there is limited mechanistic understanding of the algorithms that Transformer+CoT can learn. Our key contributions are: (1) We evaluate the state tracking capabilities of Transformer+CoT and its variants, confirming the effectiveness of CoT. (2) Next, we identify the circuit (a subset of model components, responsible for tracking the world state), indicating that late-layer MLP neurons play a key role. We propose two metrics, compression and distinction, and show that the neuron sets for each state achieve nearly 100% accuracy, providing evidence of an implicit finite state automaton (FSA) embedded within the model. (3) Additionally, we explore three challenging settings: skipping intermediate steps, introducing data noises, and testing length generalization. Our results demonstrate that Transformer+CoT learns robust algorithms (FSAs), highlighting its resilience in challenging scenarios. Our code is available at https://github.com/IvanChangPKU/FSA.
Authors: Keyu Duan, Yiran Zhao, Zhili Feng, Jinjie Ni, Tianyu Pang, Qian Liu, Tianle Cai, Longxu Dou, Kenji Kawaguchi, Anirudh Goyal, J. Zico Kolter, Michael Qizhe Shieh
Abstract: Large Language Models (LLMs) have been observed to process non-human-readable text sequences, such as jailbreak prompts, often viewed as a bug for aligned LLMs. In this work, we present a systematic investigation challenging this perception, demonstrating that unnatural languages - strings that appear incomprehensible to humans but maintain semantic meanings for LLMs - contain latent features usable by models. Notably, unnatural languages possess latent features that can be generalized across different models and tasks during inference. Furthermore, models fine-tuned on unnatural versions of instruction datasets perform on-par with those trained on natural language, achieving 49.71 win rates in Length-controlled AlpacaEval 2.0 in average across various base models. In addition, through comprehensive analysis, we demonstrate that LLMs process unnatural languages by filtering noise and inferring contextual meaning from filtered words.
Authors: Xingzuo Li, Kehai Chen, Yunfei Long, Xuefeng Bai, Yong Xu, Min Zhang
Abstract: Large language model (LLM) agents typically adopt a step-by-step reasoning framework, in which they interleave the processes of thinking and acting to accomplish the given task. However, this paradigm faces a deep-rooted one-pass issue whereby each generated intermediate thought is plugged into the trajectory regardless of its correctness, which can cause irreversible error propagation. To address the issue, this paper proposes a novel framework called Generator-Assistant Stepwise Rollback (GA-Rollback) to induce better decision-making for LLM agents. Particularly, GA-Rollback utilizes a generator to interact with the environment and an assistant to examine each action produced by the generator, where the assistant triggers a rollback operation upon detection of incorrect actions. Moreover, we introduce two additional strategies tailored for the rollback scenario to further improve its effectiveness. Extensive experiments show that GA-Rollback achieves significant improvements over several strong baselines on three widely used benchmarks. Our analysis further reveals that GA-Rollback can function as a robust plug-and-play module, integrating seamlessly with other methods.
Authors: Mohsen Fayyaz, Ali Modarressi, Hinrich Schuetze, Nanyun Peng
Abstract: Dense retrieval models are commonly used in Information Retrieval (IR) applications, such as Retrieval-Augmented Generation (RAG). Since they often serve as the first step in these systems, their robustness is critical to avoid downstream failures. In this work, we repurpose a relation extraction dataset (e.g., Re-DocRED) to design controlled experiments that quantify the impact of heuristic biases, such as a preference for shorter documents, on retrievers like Dragon+ and Contriever. We uncover major vulnerabilities, showing retrievers favor shorter documents, early positions, repeated entities, and literal matches, all while ignoring the answer's presence! Notably, when multiple biases combine, models exhibit catastrophic performance degradation, selecting the answer-containing document in less than 10% of cases over a synthetic biased document without the answer. Furthermore, we show that these biases have direct consequences for downstream applications like RAG, where retrieval-preferred documents can mislead LLMs, resulting in a 34% performance drop than providing no documents at all. https://huggingface.co/datasets/mohsenfayyaz/ColDeR
Authors: Angela Lopez-Cardona, Sebastian Idesis, Miguel Barreda-\'Angeles, Sergi Abadal, Ioannis Arapakis
Abstract: While Large Language Models (LLMs) have significantly advanced natural language processing, aligning them with human preferences remains an open challenge. Although current alignment methods rely primarily on explicit feedback, eye-tracking (ET) data offers insights into real-time cognitive processing during reading. In this paper, we present OASST-ETC, a novel eye-tracking corpus capturing reading patterns from 24 participants, while evaluating LLM-generated responses from the OASST1 dataset. Our analysis reveals distinct reading patterns between preferred and non-preferred responses, which we compare with synthetic eye-tracking data. Furthermore, we examine the correlation between human reading measures and attention patterns from various transformer-based models, discovering stronger correlations in preferred responses. This work introduces a unique resource for studying human cognitive processing in LLM evaluation and suggests promising directions for incorporating eye-tracking data into alignment methods. The dataset and analysis code are publicly available.
Authors: Tamar I. Regev, Chiebuka Ohams, Shaylee Xie, Lukas Wolf, Evelina Fedorenko, Alex Warstadt, Ethan G. Wilcox, Tiago Pimentel
Abstract: In spoken communication, information is transmitted not only via words, but also through a rich array of non-verbal signals, including prosody--the non-segmental auditory features of speech. Do these different communication channels carry distinct information? Prior work has shown that the information carried by prosodic features is substantially redundant with that carried by the surrounding words. Here, we systematically examine the time scale of this relationship, studying how it varies with the length of past and future contexts. We find that a word's prosodic features require an extended past context (3-8 words across different features) to be reliably predicted. Given that long-scale contextual information decays in memory, prosody may facilitate communication by adding information that is locally unique. We also find that a word's prosodic features show some redundancy with future words, but only with a short scale of 1-2 words, consistent with reports of incremental short-term planning in language production. Thus, prosody may facilitate communication by helping listeners predict upcoming material. In tandem, our results highlight potentially distinct roles that prosody plays in facilitating integration of words into past contexts and in helping predict upcoming words.
Authors: Bar Gazit (Ben-Gurion University of the Negev), Shaltiel Shmidman (DICTA), Avi Shmidman (DICTA), Yuval Pinter (Ben-Gurion University of the Negev)
Abstract: Common subword tokenization algorithms like BPE and UnigramLM assume that text can be split into meaningful units by concatenative measures alone. This is not true for languages such as Hebrew and Arabic, where morphology is encoded in root-template patterns, or Malay and Georgian, where split affixes are common. We present SPLINTER, a pre-processing step which rearranges text into a linear form that better represents such nonconcatenative morphologies, enabling meaningful contiguous segments to be found by the tokenizer. We demonstrate SPLINTER's merit using both intrinsic measures evaluating token vocabularies in Hebrew, Arabic, and Malay; as well as on downstream tasks using BERT-architecture models trained for Hebrew.
Authors: Michael Goodale, Salvador Mascarenhas, Yair Lakretz
Abstract: Children acquire language despite being exposed to several orders of magnitude less data than large language models require. Meta-learning has been proposed as a way to integrate human-like learning biases into neural-network architectures, combining both the structured generalizations of symbolic models with the scalability of neural-network models. But what does meta-learning exactly imbue the model with? We investigate the meta-learning of formal languages and find that, contrary to previous claims, meta-trained models are not learning simplicity-based priors when meta-trained on datasets organised around simplicity. Rather, we find evidence that meta-training imprints neural mechanisms (such as counters) into the model, which function like cognitive primitives for the network on downstream tasks. Most surprisingly, we find that meta-training on a single formal language can provide as much improvement to a model as meta-training on 5000 different formal languages, provided that the formal language incentivizes the learning of useful neural mechanisms. Taken together, our findings provide practical implications for efficient meta-learning paradigms and new theoretical insights into linking symbolic theories and neural mechanisms.
Authors: Suet-Ying Lam, Qingcheng Zeng, Jingyi Wu, Rob Voigt
Abstract: Whether large language models (LLMs) process language similarly to humans has been the subject of much theoretical and practical debate. We examine this question through the lens of the production-interpretation distinction found in human sentence processing and evaluate the extent to which instruction-tuned LLMs replicate this distinction. Using an empirically documented asymmetry between pronoun production and interpretation in humans for implicit causality verbs as a testbed, we find that some LLMs do quantitatively and qualitatively reflect human-like asymmetries between production and interpretation. We demonstrate that whether this behavior holds depends upon both model size-with larger models more likely to reflect human-like patterns and the choice of meta-linguistic prompts used to elicit the behavior. Our codes and results are available at https://github.com/LingMechLab/Production-Interpretation_Asymmetries_ACL2025.
URLs: https://github.com/LingMechLab/Production-Interpretation_Asymmetries_ACL2025.
Authors: Tereza Vrabcov\'a, Marek Kadl\v{c}\'ik, Petr Sojka, Michal \v{S}tef\'anik, Michal Spiegel
Abstract: Negations are key to determining sentence meaning, making them essential for logical reasoning. Despite their importance, negations pose a substantial challenge for large language models (LLMs) and remain underexplored. We constructed and published two new textual entailment datasets NoFEVER-ML and NoSNLI-ML in four languages (English, Czech, German, and Ukrainian) with examples differing in negation. It allows investigation of the root causes of the negation problem and its exemplification: how popular LLM model properties and language impact their inability to handle negation correctly. Contrary to previous work, we show that increasing the model size may improve the models' ability to handle negations. Furthermore, we find that both the models' reasoning accuracy and robustness to negation are language-dependent and that the length and explicitness of the premise have an impact on robustness. There is better accuracy in projective language with fixed order, such as English, than in non-projective ones, such as German or Czech. Our entailment datasets pave the way to further research for explanation and exemplification of the negation problem, minimization of LLM hallucinations, and improvement of LLM reasoning in multilingual settings.
Authors: Owen Cook, Jake Vasilakes, Ian Roberts, Xingyi Song
Abstract: Data annotation is an essential component of the machine learning pipeline; it is also a costly and time-consuming process. With the introduction of transformer-based models, annotation at the document level is increasingly popular; however, there is no standard framework for structuring such tasks. The EffiARA annotation framework is, to our knowledge, the first project to support the whole annotation pipeline, from understanding the resources required for an annotation task to compiling the annotated dataset and gaining insights into the reliability of individual annotators as well as the dataset as a whole. The framework's efficacy is supported by two previous studies: one improving classification performance through annotator-reliability-based soft-label aggregation and sample weighting, and the other increasing the overall agreement among annotators through removing identifying and replacing an unreliable annotator. This work introduces the EffiARA Python package and its accompanying webtool, which provides an accessible graphical user interface for the system. We open-source the EffiARA Python package at https://github.com/MiniEggz/EffiARA and the webtool is publicly accessible at https://effiara.gate.ac.uk.
URLs: https://github.com/MiniEggz/EffiARA, https://effiara.gate.ac.uk.
Authors: Jiawei Lian, Jianhong Pan, Lefan Wang, Yi Wang, Shaohui Mei, Lap-Pui Chau
Abstract: Large language models (LLMs) are foundational explorations to artificial general intelligence, yet their alignment with human values via instruction tuning and preference learning achieves only superficial compliance. Here, we demonstrate that harmful knowledge embedded during pretraining persists as indelible "dark patterns" in LLMs' parametric memory, evading alignment safeguards and resurfacing under adversarial inducement at distributional shifts. In this study, we first theoretically analyze the intrinsic ethical vulnerability of aligned LLMs by proving that current alignment methods yield only local "safety regions" in the knowledge manifold. In contrast, pretrained knowledge remains globally connected to harmful concepts via high-likelihood adversarial trajectories. Building on this theoretical insight, we empirically validate our findings by employing semantic coherence inducement under distributional shifts--a method that systematically bypasses alignment constraints through optimized adversarial prompts. This combined theoretical and empirical approach achieves a 100% attack success rate across 19 out of 23 state-of-the-art aligned LLMs, including DeepSeek-R1 and LLaMA-3, revealing their universal vulnerabilities.
Authors: Kseniia Petukhova, Ekaterina Kochmar
Abstract: Recent advances in Large Language Models (LLMs) have shown promise in automating discourse annotation for conversations. While manually designing tree annotation schemes significantly improves annotation quality for humans and models, their creation remains time-consuming and requires expert knowledge. We propose a fully automated pipeline that uses LLMs to construct such schemes and perform annotation. We evaluate our approach on speech functions (SFs) and the Switchboard-DAMSL (SWBD-DAMSL) taxonomies. Our experiments compare various design choices, and we show that frequency-guided decision trees, paired with an advanced LLM for annotation, can outperform previously manually designed trees and even match or surpass human annotators while significantly reducing the time required for annotation. We release all code and resultant schemes and annotations to facilitate future research on discourse annotation.
Authors: Sukannya Purkayastha, Zhuang Li, Anne Lauscher, Lizhen Qu, Iryna Gurevych
Abstract: Peer review is a cornerstone of quality control in scientific publishing. With the increasing workload, the unintended use of `quick' heuristics, referred to as lazy thinking, has emerged as a recurring issue compromising review quality. Automated methods to detect such heuristics can help improve the peer-reviewing process. However, there is limited NLP research on this issue, and no real-world dataset exists to support the development of detection tools. This work introduces LazyReview, a dataset of peer-review sentences annotated with fine-grained lazy thinking categories. Our analysis reveals that Large Language Models (LLMs) struggle to detect these instances in a zero-shot setting. However, instruction-based fine-tuning on our dataset significantly boosts performance by 10-20 performance points, highlighting the importance of high-quality training data. Furthermore, a controlled experiment demonstrates that reviews revised with lazy thinking feedback are more comprehensive and actionable than those written without such feedback. We will release our dataset and the enhanced guidelines that can be used to train junior reviewers in the community. (Code available here: https://github.com/UKPLab/acl2025-lazy-review)
Authors: Siyan Zhao, Devaansh Gupta, Qinqing Zheng, Aditya Grover
Abstract: Recent large language models (LLMs) have demonstrated strong reasoning capabilities that benefits from online reinforcement learning (RL). These capabilities have primarily been demonstrated within the left-to-right autoregressive (AR) generation paradigm. In contrast, non-autoregressive paradigms based on diffusion generate text in a coarse-to-fine manner. Although recent diffusion-based large language models (dLLMs) have achieved competitive language modeling performance compared to their AR counterparts, it remains unclear if dLLMs can also leverage recent advances in LLM reasoning. To this end, we propose d1, a framework to adapt pre-trained masked dLLMs into reasoning models via a combination of supervised finetuning (SFT) and RL. Specifically, we develop and extend techniques to improve reasoning in pretrained dLLMs: (a) we utilize a masked SFT technique to distill knowledge and instill self-improvement behavior directly from existing datasets, and (b) we introduce a novel critic-free, policy-gradient based RL algorithm called diffu-GRPO, the first integration of policy gradient methods to masked dLLMs. Through empirical studies, we investigate the performance of different post-training recipes on multiple mathematical and planning benchmarks. We find that d1 yields the best performance and significantly improves performance of a state-of-the-art dLLM. Our code is released at https://dllm-reasoning.github.io/.
Authors: Haoming Yang, Ke Ma, Xiaojun Jia, Yingfei Sun, Qianqian Xu, Qingming Huang
Abstract: Despite the remarkable performance of Large Language Models (LLMs), they remain vulnerable to jailbreak attacks, which can compromise their safety mechanisms. Existing studies often rely on brute-force optimization or manual design, failing to uncover potential risks in real-world scenarios. To address this, we propose a novel jailbreak attack framework, ICRT, inspired by heuristics and biases in human cognition. Leveraging the simplicity effect, we employ cognitive decomposition to reduce the complexity of malicious prompts. Simultaneously, relevance bias is utilized to reorganize prompts, enhancing semantic alignment and inducing harmful outputs effectively. Furthermore, we introduce a ranking-based harmfulness evaluation metric that surpasses the traditional binary success-or-failure paradigm by employing ranking aggregation methods such as Elo, HodgeRank, and Rank Centrality to comprehensively quantify the harmfulness of generated content. Experimental results show that our approach consistently bypasses mainstream LLMs' safety mechanisms and generates high-risk content, providing insights into jailbreak attack risks and contributing to stronger defense strategies.
Authors: Junyu Ma, Tianqing Fang, Zhisong Zhang, Hongming Zhang, Haitao Mi, Dong Yu
Abstract: Mamba's theoretical infinite-context potential is limited in practice when sequences far exceed training lengths. This work explores unlocking Mamba's long-context memory ability by a simple-yet-effective method, Recall with Reasoning (RwR), by distilling chain-of-thought (CoT) summarization from a teacher model. Specifically, RwR prepends these summarization as CoT prompts during fine-tuning, teaching Mamba to actively recall and reason over long contexts. Experiments on LONGMEMEVAL and HELMET show RwR boosts Mamba's long-context performance against comparable Transformer/hybrid baselines under similar pretraining conditions, while preserving short-context capabilities, all without architectural changes.
Authors: Ryan Lagasse, Aidan Kierans, Avijit Ghosh, Shiri Dori-Hacohen
Abstract: We introduce a scaling law for fine-tuning large language models (LLMs) under fixed compute budgets that explicitly accounts for data composition. Conventional approaches measure training data solely by total tokens, yet the number of examples and their average token length -- what we term \emph{dataset volume} -- play a decisive role in model performance. Our formulation is tuned following established procedures. Experiments on the BRICC dataset \cite{salavati2024reducing} and subsets of the MMLU dataset \cite{hendrycks2021measuringmassivemultitasklanguage}, evaluated under multiple subsampling strategies, reveal that data composition significantly affects token efficiency. These results motivate refined scaling laws for practical LLM fine-tuning in resource-constrained settings.
Authors: Peiqi Sui, Juan Diego Rodriguez, Philippe Laban, Dean Murphy, Joseph P. Dexter, Richard Jean So, Samuel Baker, Pramit Chaudhuri
Abstract: Each year, tens of millions of essays are written and graded in college-level English courses. Students are asked to analyze literary and cultural texts through a process known as close reading, in which they gather textual details to formulate evidence-based arguments. Despite being viewed as a basis for critical thinking and widely adopted as a required element of university coursework, close reading has never been evaluated on large language models (LLMs), and multi-discipline benchmarks like MMLU do not include literature as a subject. To fill this gap, we present KRISTEVA, the first close reading benchmark for evaluating interpretive reasoning, consisting of 1331 multiple-choice questions adapted from classroom data. With KRISTEVA, we propose three progressively more difficult sets of tasks to approximate different elements of the close reading process, which we use to test how well LLMs may seem to understand and reason about literary works: 1) extracting stylistic features, 2) retrieving relevant contextual information from parametric knowledge, and 3) multi-hop reasoning between style and external contexts. Our baseline results find that, while state-of-the-art LLMs possess some college-level close reading competency (accuracy 49.7% - 69.7%), their performances still trail those of experienced human evaluators on 10 out of our 11 tasks.
Authors: Qiongqiong Wang, Hardik B. Sailor, Tianchi Liu, Ai Ti Aw
Abstract: Current speech-LLMs exhibit limited capability in contextual reasoning alongside paralinguistic understanding, primarily due to the lack of Question-Answer (QA) datasets that cover both aspects. We propose a novel framework for dataset generation from in-the-wild speech data, that integrates contextual reasoning with paralinguistic information. It consists of a pseudo paralinguistic label-based data condensation of in-the-wild speech and LLM-based Contextual Paralinguistic QA (CPQA) generation. The effectiveness is validated by a strong correlation in evaluations of the Qwen2-Audio-7B-Instruct model on a dataset created by our framework and human-generated CPQA dataset. The results also reveal the speech-LLM's limitations in handling empathetic reasoning tasks, highlighting the need for such datasets and more robust models. The proposed framework is first of its kind and has potential in training more robust speech-LLMs with paralinguistic reasoning capabilities.
Authors: Zijia Liu, Peixuan Han, Haofei Yu, Haoru Li, Jiaxuan You
Abstract: Large Language Models (LLMs) demonstrate impressive capabilities but lack robust temporal intelligence, struggling to integrate reasoning about the past with predictions and plausible generations of the future. Meanwhile, existing methods typically target isolated temporal skills, such as question answering about past events or basic forecasting, and exhibit poor generalization, particularly when dealing with events beyond their knowledge cutoff or requiring creative foresight. To address these limitations, we introduce \textit{Time-R1}, the first framework to endow a moderate-sized (3B-parameter) LLM with comprehensive temporal abilities: understanding, prediction, and creative generation. Our approach features a novel three-stage development path; the first two constitute a \textit{reinforcement learning (RL) curriculum} driven by a meticulously designed dynamic rule-based reward system. This framework progressively builds (1) foundational temporal understanding and logical event-time mappings from historical data, (2) future event prediction skills for events beyond its knowledge cutoff, and finally (3) enables remarkable generalization to creative future scenario generation without any fine-tuning. Strikingly, experiments demonstrate that Time-R1 outperforms models over 200 times larger, including the state-of-the-art 671B DeepSeek-R1, on highly challenging future event prediction and creative scenario generation benchmarks. This work provides strong evidence that thoughtfully engineered, progressive RL fine-tuning allows smaller, efficient models to achieve superior temporal performance, offering a practical and scalable path towards truly time-aware AI. To foster further research, we also release \textit{Time-Bench}, a large-scale multi-task temporal reasoning dataset derived from 10 years of news data, and our series of \textit{Time-R1} checkpoints.
Authors: Maodong Li, Longyin Zhang, Fang Kong
Abstract: Multi-hop question generation (MQG) aims to generate questions that require synthesizing multiple information snippets from documents to derive target answers. The primary challenge lies in effectively pinpointing crucial information snippets related to question-answer (QA) pairs, typically relying on keywords. However, existing works fail to fully utilize the guiding potential of keywords and neglect to differentiate the distinct roles of question-specific and document-specific keywords. To address this, we define dual-perspective keywords (i.e., question and document keywords) and propose a Dual-Perspective Keyword-Guided (DPKG) framework, which seamlessly integrates keywords into the multi-hop question generation process. We argue that question keywords capture the questioner's intent, whereas document keywords reflect the content related to the QA pair. Functionally, question and document keywords work together to pinpoint essential information snippets in the document, with question keywords required to appear in the generated question. The DPKG framework consists of an expanded transformer encoder and two answer-aware transformer decoders for keyword and question generation, respectively. Extensive experiments demonstrate the effectiveness of our work, showcasing its promising performance and underscoring its significant value in the MQG task.
Authors: Yash Saxena, Ankur Padia, Mandar S Chaudhary, Kalpa Gunaratna, Srinivasan Parthasarathy, Manas Gaur
Abstract: Traditional Retrieval-Augmented Generation (RAG) pipelines rely on similarity-based retrieval and re-ranking, which depend on heuristics such as top-k, and lack explainability, interpretability, and robustness against adversarial content. To address this gap, we propose a novel method METEORA that replaces re-ranking in RAG with a rationale-driven selection approach. METEORA operates in two stages. First, a general-purpose LLM is preference-tuned to generate rationales conditioned on the input query using direct preference optimization. These rationales guide the evidence chunk selection engine, which selects relevant chunks in three stages: pairing individual rationales with corresponding retrieved chunks for local relevance, global selection with elbow detection for adaptive cutoff, and context expansion via neighboring chunks. This process eliminates the need for top-k heuristics. The rationales are also used for consistency check using a Verifier LLM to detect and filter poisoned or misleading content for safe generation. The framework provides explainable and interpretable evidence flow by using rationales consistently across both selection and verification. Our evaluation across six datasets spanning legal, financial, and academic research domains shows that METEORA improves generation accuracy by 33.34% while using approximately 50% fewer chunks than state-of-the-art re-ranking methods. In adversarial settings, METEORA significantly improves the F1 score from 0.10 to 0.44 over the state-of-the-art perplexity-based defense baseline, demonstrating strong resilience to poisoning attacks. Code available at: https://anonymous.4open.science/r/METEORA-DC46/README.md
URLs: https://anonymous.4open.science/r/METEORA-DC46/README.md
Authors: Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, Ruihua Song
Abstract: Large Language Models (LLMs) achieve superior performance through Chain-of-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor randomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) perform reasoning at a dense latent level (i.e., silently), substantially reducing reasoning chain length, and ii) dynamically adjust reasoning speed at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3% with only 4.8% performance degradation compared to explicit CoT method. Moreover, when applied to more challenging mathematical reasoning tasks, our RL-enhanced CoLaR demonstrates performance gains of up to 5.4% while dramatically reducing latent reasoning chain length by 82.8%. The code and models will be released upon acceptance.
Authors: Liu Chang, Wang Dongbo, Liu liu, Zhao Zhixiao
Abstract: This study addresses the challenges in intelligent processing of Chinese ancient mathematical classics by constructing Guji_MATH, a benchmark for evaluating classical texts based on Suanjing Shishu. It systematically assesses the mathematical problem-solving capabilities of mainstream reasoning models under the unique linguistic constraints of classical Chinese. Through machine-assisted annotation and manual verification, 538 mathematical problems were extracted from 8 canonical texts, forming a structured dataset centered on the "Question-Answer-Solution" framework, supplemented by problem types and difficulty levels. Dual evaluation modes--closed-book (autonomous problem-solving) and open-book (reproducing classical solution methods)--were designed to evaluate the performance of six reasoning models on ancient Chinese mathematical problems. Results indicate that reasoning models can partially comprehend and solve these problems, yet their overall performance remains inferior to benchmarks on modern mathematical tasks. Enhancing models' classical Chinese comprehension and cultural knowledge should be prioritized for optimization. This study provides methodological support for mining mathematical knowledge from ancient texts and disseminating traditional culture, while offering new perspectives for evaluating cross-linguistic and cross-cultural capabilities of reasoning models.
Authors: Chaochen Gao, Xing Wu, Zijia Lin, Debing Zhang, Songlin Hu
Abstract: High-quality long-context instruction data is essential for aligning long-context large language models (LLMs). Despite the public release of models like Qwen and Llama, their long-context instruction data remains proprietary. Human annotation is costly and challenging, while template-based synthesis methods limit scale, diversity, and quality. We introduce LongMagpie, a self-synthesis framework that automatically generates large-scale long-context instruction data. Our key insight is that aligned long-context LLMs, when presented with a document followed by special tokens preceding a user turn, auto-regressively generate contextually relevant queries. By harvesting these document-query pairs and the model's responses, LongMagpie produces high-quality instructions without human effort. Experiments on HELMET, RULER, and Longbench v2 demonstrate that LongMagpie achieves leading performance on long-context tasks while maintaining competitive performance on short-context tasks, establishing it as a simple and effective approach for open, diverse, and scalable long-context instruction data synthesis.
Authors: Ramon Ferrer-i-Cancho
Abstract: Here we present a new class of optimality for coding systems. Members of that class are displaced linearly from optimal coding and thus exhibit Zipf's law, namely a power-law distribution of frequency ranks. Within that class, Zipf's law, the size-rank law and the size-probability law form a group-like structure. We identify human languages that are members of the class. All languages showing sufficient agreement with Zipf's law are potential members of the class. In contrast, there are communication systems in other species that cannot be members of that class for exhibiting an exponential distribution instead but dolphins and humpback whales might. We provide a new insight into plots of frequency versus rank in double logarithmic scale. For any system, a straight line in that scale indicates that the lengths of optimal codes under non-singular coding and under uniquely decodable encoding are displaced by a linear function whose slope is the exponent of Zipf's law. For systems under compression and constrained to be uniquely decodable, such a straight line may indicate that the system is coding close to optimality. We provide support for the hypothesis that Zipf's law originates from compression and define testable conditions for the emergence of Zipf's law in compressing systems.
Authors: Zitian Gao, Lynx Chen, Joey Zhou, Bryan Dai
Abstract: We trained 13,440 large language models and found that entropy minimization requires only a single unlabeled data and 10 steps optimization to achieve performance improvements comparable to or even greater than those obtained using thousands of data and carefully designed rewards in rule-based reinforcement learning. This striking result may prompt a rethinking of post-training paradigms for large language models. Our code is avaliable at https://github.com/zitian-gao/one-shot-em.
Authors: Mengru Wang, Ziwen Xu, Shengyu Mao, Shumin Deng, Zhaopeng Tu, Huajun Chen, Ningyu Zhang
Abstract: Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.
Authors: Qiuchen Wang, Ruixue Ding, Yu Zeng, Zehui Chen, Lin Chen, Shihang Wang, Pengjun Xie, Fei Huang, Feng Zhao
Abstract: Effectively retrieving, reasoning and understanding visually rich information remains a challenge for RAG methods. Traditional text-based methods cannot handle visual-related information. On the other hand, current vision-based RAG approaches are often limited by fixed pipelines and frequently struggle to reason effectively due to the insufficient activation of the fundamental capabilities of models. As RL has been proven to be beneficial for model reasoning, we introduce VRAG-RL, a novel RL framework tailored for complex reasoning across visually rich information. With this framework, VLMs interact with search engines, autonomously sampling single-turn or multi-turn reasoning trajectories with the help of visual perception tokens and undergoing continual optimization based on these samples. Our approach highlights key limitations of RL in RAG domains: (i) Prior Multi-modal RAG approaches tend to merely incorporate images into the context, leading to insufficient reasoning token allocation and neglecting visual-specific perception; and (ii) When models interact with search engines, their queries often fail to retrieve relevant information due to the inability to articulate requirements, thereby leading to suboptimal performance. To address these challenges, we define an action space tailored for visually rich inputs, with actions including cropping and scaling, allowing the model to gather information from a coarse-to-fine perspective. Furthermore, to bridge the gap between users' original inquiries and the retriever, we employ a simple yet effective reward that integrates query rewriting and retrieval performance with a model-based reward. Our VRAG-RL optimizes VLMs for RAG tasks using specially designed RL strategies, aligning the model with real-world applications. The code is available at https://github.com/Alibaba-NLP/VRAG.
Authors: Jintao Zhang, Zirui Liu, Mingyue Cheng, Shilong Zhang, Tingyue Pan, Qi Liu, Yanhu Xie
Abstract: Intraoperative hypotension (IOH) frequently occurs under general anesthesia and is strongly linked to adverse outcomes such as myocardial injury and increased mortality. Despite its significance, IOH prediction is hindered by event sparsity and the challenge of integrating static and dynamic data across diverse patients. In this paper, we propose \textbf{IOHFuseLM}, a multimodal language model framework. To accurately identify and differentiate sparse hypotensive events, we leverage a two-stage training strategy. The first stage involves domain adaptive pretraining on IOH physiological time series augmented through diffusion methods, thereby enhancing the model sensitivity to patterns associated with hypotension. Subsequently, task fine-tuning is performed on the original clinical dataset to further enhance the ability to distinguish normotensive from hypotensive states. To enable multimodal fusion for each patient, we align structured clinical descriptions with the corresponding physiological time series at the token level. Such alignment enables the model to capture individualized temporal patterns alongside their corresponding clinical semantics. In addition, we convert static patient attributes into structured text to enrich personalized information. Experimental evaluations on two intraoperative datasets demonstrate that IOHFuseLM outperforms established baselines in accurately identifying IOH events, highlighting its applicability in clinical decision support scenarios. Our code is publicly available to promote reproducibility at https://github.com/zjt-gpu/IOHFuseLM.
Authors: Alexander Gill, Abhilasha Ravichander, Ana Marasovi\'c
Abstract: Large language models (LLMs) are increasingly used for data generation. However, creating evaluation benchmarks raises the bar for this emerging paradigm. Benchmarks must target specific phenomena, penalize exploiting shortcuts, and be challenging. Through two case studies, we investigate whether LLMs can meet these demands by generating reasoning over-text benchmarks and comparing them to those created through careful crowdsourcing. Specifically, we evaluate both the validity and difficulty of LLM-generated versions of two high-quality reading comprehension datasets: CondaQA, which evaluates reasoning about negation, and DROP, which targets reasoning about quantities. We find that prompting LLMs can produce variants of these datasets that are often valid according to the annotation guidelines, at a fraction of the cost of the original crowdsourcing effort. However, we show that they are less challenging for LLMs than their human-authored counterparts. This finding sheds light on what may have been lost by generating evaluation data with LLMs, and calls for critically reassessing the immediate use of this increasingly prevalent approach to benchmark creation.
Authors: Pingjun Hong, Beiduo Chen, Siyao Peng, Marie-Catherine de Marneffe, Barbara Plank
Abstract: There is increasing evidence of Human Label Variation (HLV) in Natural Language Inference (NLI), where annotators assign different labels to the same premise-hypothesis pair. However, within-label variation--cases where annotators agree on the same label but provide divergent reasoning--poses an additional and mostly overlooked challenge. Several NLI datasets contain highlighted words in the NLI item as explanations, but the same spans on the NLI item can be highlighted for different reasons, as evidenced by free-text explanations, which offer a window into annotators' reasoning. To systematically understand this problem and gain insight into the rationales behind NLI labels, we introduce LITEX, a linguistically-informed taxonomy for categorizing free-text explanations. Using this taxonomy, we annotate a subset of the e-SNLI dataset, validate the taxonomy's reliability, and analyze how it aligns with NLI labels, highlights, and explanations. We further assess the taxonomy's usefulness in explanation generation, demonstrating that conditioning generation on LITEX yields explanations that are linguistically closer to human explanations than those generated using only labels or highlights. Our approach thus not only captures within-label variation but also shows how taxonomy-guided generation for reasoning can bridge the gap between human and model explanations more effectively than existing strategies.
Authors: Yize Cheng, Wenxiao Wang, Mazda Moayeri, Soheil Feizi
Abstract: Open benchmarks are essential for evaluating and advancing large language models, offering reproducibility and transparency. However, their accessibility makes them likely targets of test set contamination. In this work, we introduce DyePack, a framework that leverages backdoor attacks to identify models that used benchmark test sets during training, without requiring access to the loss, logits, or any internal details of the model. Like how banks mix dye packs with their money to mark robbers, DyePack mixes backdoor samples with the test data to flag models that trained on it. We propose a principled design incorporating multiple backdoors with stochastic targets, enabling exact false positive rate (FPR) computation when flagging every model. This provably prevents false accusations while providing strong evidence for every detected case of contamination. We evaluate DyePack on five models across three datasets, covering both multiple-choice and open-ended generation tasks. For multiple-choice questions, it successfully detects all contaminated models with guaranteed FPRs as low as 0.000073% on MMLU-Pro and 0.000017% on Big-Bench-Hard using eight backdoors. For open-ended generation tasks, it generalizes well and identifies all contaminated models on Alpaca with a guaranteed false positive rate of just 0.127% using six backdoors.
Authors: Seohyeong Lee, Eunwon Kim, Hwaran Lee, Buru Chang
Abstract: Human preference data plays a critical role in aligning large language models (LLMs) with human values. However, collecting such data is often expensive and inefficient, posing a significant scalability challenge. To address this, we introduce Alignment Data Map, a GPT-4o-assisted tool for analyzing and diagnosing preference data. Using GPT-4o as a proxy for LLM alignment, we compute alignment scores for LLM-generated responses to instructions from existing preference datasets. These scores are then used to construct an Alignment Data Map based on their mean and variance. Our experiments show that using only 33 percent of the data, specifically samples in the high-mean, low-variance region, achieves performance comparable to or better than using the entire dataset. This finding suggests that the Alignment Data Map can significantly improve data collection efficiency by identifying high-quality samples for LLM alignment without requiring explicit annotations. Moreover, the Alignment Data Map can diagnose existing preference datasets. Our analysis shows that it effectively detects low-impact or potentially misannotated samples. Source code is available online.
Authors: Beiduo Chen, Yang Janet Liu, Anna Korhonen, Barbara Plank
Abstract: The recent rise of reasoning-tuned Large Language Models (LLMs)--which generate chains of thought (CoTs) before giving the final answer--has attracted significant attention and offers new opportunities for gaining insights into human label variation, which refers to plausible differences in how multiple annotators label the same data instance. Prior work has shown that LLM-generated explanations can help align model predictions with human label distributions, but typically adopt a reverse paradigm: producing explanations based on given answers. In contrast, CoTs provide a forward reasoning path that may implicitly embed rationales for each answer option, before generating the answers. We thus propose a novel LLM-based pipeline enriched with linguistically-grounded discourse segmenters to extract supporting and opposing statements for each answer option from CoTs with improved accuracy. We also propose a rank-based HLV evaluation framework that prioritizes the ranking of answers over exact scores, which instead favor direct comparison of label distributions. Our method outperforms a direct generation method as well as baselines on three datasets, and shows better alignment of ranking methods with humans, highlighting the effectiveness of our approach.
Authors: Ziyin Zhang, Jiahao Xu, Zhiwei He, Tian Liang, Qiuzhi Liu, Yansi Li, Linfeng Song, Zhenwen Liang, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, Dong Yu
Abstract: Theorem proving serves as a major testbed for evaluating complex reasoning abilities in large language models (LLMs). However, traditional automated theorem proving (ATP) approaches rely heavily on formal proof systems that poorly align with LLMs' strength derived from informal, natural language knowledge acquired during pre-training. In this work, we propose DeepTheorem, a comprehensive informal theorem-proving framework exploiting natural language to enhance LLM mathematical reasoning. DeepTheorem includes a large-scale benchmark dataset consisting of 121K high-quality IMO-level informal theorems and proofs spanning diverse mathematical domains, rigorously annotated for correctness, difficulty, and topic categories, accompanied by systematically constructed verifiable theorem variants. We devise a novel reinforcement learning strategy (RL-Zero) explicitly tailored to informal theorem proving, leveraging the verified theorem variants to incentivize robust mathematical inference. Additionally, we propose comprehensive outcome and process evaluation metrics examining proof correctness and the quality of reasoning steps. Extensive experimental analyses demonstrate DeepTheorem significantly improves LLM theorem-proving performance compared to existing datasets and supervised fine-tuning protocols, achieving state-of-the-art accuracy and reasoning quality. Our findings highlight DeepTheorem's potential to fundamentally advance automated informal theorem proving and mathematical exploration.
Authors: Yuli Chen, Bo Cheng, Jiale Han, Yingying Zhang, Yingting Li, Shuhao Zhang
Abstract: Pruning has recently been widely adopted to reduce the parameter scale and improve the inference efficiency of Large Language Models (LLMs). Mainstream pruning techniques often rely on uniform layerwise pruning strategies, which can lead to severe performance degradation at high sparsity levels. Recognizing the varying contributions of different layers in LLMs, recent studies have shifted their focus toward non-uniform layerwise pruning. However, these approaches often rely on pre-defined values, which can result in suboptimal performance. To overcome these limitations, we propose a novel method called Dynamic Layerwise Pruning (DLP). This approach adaptively determines the relative importance of each layer by integrating model weights with input activation information, assigning pruning rates accordingly. Experimental results show that DLP effectively preserves model performance at high sparsity levels across multiple LLMs. Specifically, at 70% sparsity, DLP reduces the perplexity of LLaMA2-7B by 7.79 and improves the average accuracy by 2.7% compared to state-of-the-art methods. Moreover, DLP is compatible with various existing LLM compression techniques and can be seamlessly integrated into Parameter-Efficient Fine-Tuning (PEFT). We release the code at https://github.com/ironartisan/DLP to facilitate future research.
Authors: Haowei Yang, Haotian Lyu, Tianle Zhang, Dingzhou Wang, Yushang Zhao
Abstract: As e-commerce competition intensifies, balancing creative content with conversion effectiveness becomes critical. Leveraging LLMs' language generation capabilities, we propose a framework that integrates prompt engineering, multi-objective fine-tuning, and post-processing to generate marketing copy that is both engaging and conversion-driven. Our fine-tuning method combines sentiment adjustment, diversity enhancement, and CTA embedding. Through offline evaluations and online A/B tests across categories, our approach achieves a 12.5 % increase in CTR and an 8.3 % increase in CVR while maintaining content novelty. This provides a practical solution for automated copy generation and suggests paths for future multimodal, real-time personalization.
Authors: Zefan Cai, Wen Xiao, Hanshi Sun, Cheng Luo, Yikai Zhang, Ke Wan, Yucheng Li, Yeyang Zhou, Li-Wen Chang, Jiuxiang Gu, Zhen Dong, Anima Anandkumar, Abedelkadir Asi, Junjie Hu
Abstract: Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.
Authors: Celia Cintas, Miriam Rateike, Erik Miehling, Elizabeth Daly, Skyler Speakman
Abstract: We present a study on how and where personas -- defined by distinct sets of human characteristics, values, and beliefs -- are encoded in the representation space of large language models (LLMs). Using a range of dimension reduction and pattern recognition methods, we first identify the model layers that show the greatest divergence in encoding these representations. We then analyze the activations within a selected layer to examine how specific personas are encoded relative to others, including their shared and distinct embedding spaces. We find that, across multiple pre-trained decoder-only LLMs, the analyzed personas show large differences in representation space only within the final third of the decoder layers. We observe overlapping activations for specific ethical perspectives -- such as moral nihilism and utilitarianism -- suggesting a degree of polysemy. In contrast, political ideologies like conservatism and liberalism appear to be represented in more distinct regions. These findings help to improve our understanding of how LLMs internally represent information and can inform future efforts in refining the modulation of specific human traits in LLM outputs. Warning: This paper includes potentially offensive sample statements.
Authors: Shenghe Zheng, Qianjia Cheng, Junchi Yao, Mengsong Wu, Haonan He, Ning Ding, Yu Cheng, Shuyue Hu, Lei Bai, Dongzhan Zhou, Ganqu Cui, Peng Ye
Abstract: Large Language Models (LLMs) have achieved remarkable progress on advanced reasoning tasks such as mathematics and coding competitions. Meanwhile, physics, despite being both reasoning-intensive and essential to real-world understanding, received limited academic and industrial attention. This paper introduces PHYSICS, a dataset containing 16,568 high-quality physics problems spanning subjects and difficulty levels, to facilitate this issue. Specifically, PHYSICS is curated with exercises from over 100 textbooks through a carefully designed pipeline for quality control. It covers five major physics domains: Mechanics, Electromagnetism, Thermodynamics, Optics, and Modern Physics. It also spans a wide range of difficulty levels, from high school to graduate-level physics courses. To utilize the data for improving and evaluating the model's physical reasoning capabilities, we split the dataset into training and test sets, and provide reasoning paths generated by powerful reasoning models for the training data to facilitate model training. In addition, for the evaluation part, we find that existing evaluation frameworks exhibit biases in aspects such as units, simplification, and precision in physics domain. To balance efficiency and accuracy, we introduce a Rule+Model evaluation framework tailored to physics problems. Our evaluations on current state-of-the-art open-source and proprietary models highlight the limitations of current models in handling physics-related tasks. We hope that our dataset and evaluation methodology will jointly advance the development of LLMs in the field of physics.
Authors: Edward L. Wang, Tianyu Wang, Avanti Athreya, Vince Lyzinski, Carey E. Priebe
Abstract: Large language models (LLMs) are a powerful tool with the ability to match human capabilities and behavior in many settings. Retrieval-augmented generation (RAG) further allows LLMs to generate diverse output depending on the contents of their RAG database. This motivates their use in the social sciences to study human behavior between individuals when large-scale experiments are infeasible. However, LLMs depend on complex, computationally expensive algorithms. In this paper, we introduce interacting Gaussian mixture models (GMMs) as an alternative to similar frameworks using LLMs. We compare a simplified model of GMMs to select experimental simulations of LLMs whose updating and response depend on feedback from other LLMs. We find that interacting GMMs capture important features of the dynamics in interacting LLMs, and we investigate key similarities and differences between interacting LLMs and GMMs. We conclude by discussing the benefits of Gaussian mixture models, potential modifications, and future research directions.
Authors: Mohammad Javad Ranjbar Kalahroodi, Amirhossein Sheikholselami, Sepehr Karimi, Sepideh Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery
Abstract: Large Language Models (LLMs) have achieved remarkable performance on a wide range of NLP benchmarks, often surpassing human-level accuracy. However, their reliability in high-stakes domains such as medicine, particularly in low-resource languages, remains underexplored. In this work, we introduce PersianMedQA, a large-scale, expert-validated dataset of multiple-choice Persian medical questions, designed to evaluate LLMs across both Persian and English. We benchmark over 40 state-of-the-art models, including general-purpose, Persian fine-tuned, and medical LLMs, in zero-shot and chain-of-thought (CoT) settings. Our results show that closed-source general models (e.g., GPT-4.1) consistently outperform all other categories, achieving 83.3% accuracy in Persian and 80.7% in English, while Persian fine-tuned models such as Dorna underperform significantly (e.g., 35.9% in Persian), often struggling with both instruction-following and domain reasoning. We also analyze the impact of translation, showing that while English performance is generally higher, Persian responses are sometimes more accurate due to cultural and clinical contextual cues. Finally, we demonstrate that model size alone is insufficient for robust performance without strong domain or language adaptation. PersianMedQA provides a foundation for evaluating multilingual and culturally grounded medical reasoning in LLMs. The PersianMedQA dataset can be accessed at: https://huggingface.co/datasets/MohammadJRanjbar/PersianMedQA
URLs: https://huggingface.co/datasets/MohammadJRanjbar/PersianMedQA
Authors: Ahmed Elhady, Eneko Agirre, Mikel Artetxe
Abstract: Continued pretraining (CPT) is a popular approach to adapt existing large language models (LLMs) to new languages. When doing so, it is common practice to include a portion of English data in the mixture, but its role has not been carefully studied to date. In this work, we show that including English does not impact validation perplexity, yet it is critical for the emergence of downstream capabilities in the target language. We introduce a language-agnostic benchmark for in-context learning (ICL), which reveals catastrophic forgetting early on CPT when English is not included. This in turn damages the ability of the model to generalize to downstream prompts in the target language as measured by perplexity, even if it does not manifest in terms of accuracy until later in training, and can be tied to a big shift in the model parameters. Based on these insights, we introduce curriculum learning and exponential moving average (EMA) of weights as effective alternatives to mitigate the need for English. All in all, our work sheds light into the dynamics by which emergent abilities arise when doing CPT for language adaptation, and can serve as a foundation to design more effective methods in the future.
Authors: Yuxi Sun, Aoqi Zuo, Wei Gao, Jing Ma
Abstract: Large Language Models (LLMs) often exhibit knowledge disparities across languages. Encouraging LLMs to \textit{abstain} when faced with knowledge gaps is a promising strategy to reduce hallucinations in multilingual settings. Current abstention strategies for multilingual scenarios primarily rely on generating feedback in various languages using LLMs and performing self-reflection. However, these methods can be adversely impacted by inaccuracies and biases in the generated feedback. To address this, from a causal perspective, we introduce \textit{CausalAbstain}, a method that helps LLMs determine whether to utilize multiple generated feedback responses and how to identify the most useful ones. Extensive experiments demonstrate that \textit{CausalAbstain} effectively selects helpful feedback and enhances abstention decisions with interpretability in both native language (\textsc{Casual-native}) and multilingual (\textsc{Causal-multi}) settings, outperforming strong baselines on two benchmark datasets covering encyclopedic and commonsense knowledge QA tasks. Our code and data are open-sourced at https://github.com/peachch/CausalAbstain.
Authors: Running Yang, Wenlong Deng, Minghui Chen, Yuyin Zhou, Xiaoxiao Li
Abstract: Clinical tasks such as diagnosis and treatment require strong decision-making abilities, highlighting the importance of rigorous evaluation benchmarks to assess the reliability of large language models (LLMs). In this work, we introduce a knowledge-guided data augmentation framework that enhances the difficulty of clinical multiple-choice question (MCQ) datasets by generating distractors (i.e., incorrect choices that are similar to the correct one and may confuse existing LLMs). Using our KG-based pipeline, the generated choices are both clinically plausible and deliberately misleading. Our approach involves multi-step, semantically informed walks on a medical knowledge graph to identify distractor paths-associations that are medically relevant but factually incorrect-which then guide the LLM in crafting more deceptive distractors. We apply the designed knowledge graph guided distractor generation (KGGDG) pipline, to six widely used medical QA benchmarks and show that it consistently reduces the accuracy of state-of-the-art LLMs. These findings establish KGGDG as a powerful tool for enabling more robust and diagnostic evaluations of medical LLMs.
Authors: Haesung Pyun, Yoonah Park, Yohan Jo
Abstract: In dialogue state tracking (DST), in-context learning comprises a retriever that selects labeled dialogues as in-context examples and a DST model that uses these examples to infer the dialogue state of the query dialogue. Existing methods for constructing training data for retrievers suffer from three key limitations: (1) the synergistic effect of examples is not considered, (2) the linguistic characteristics of the query are not sufficiently factored in, and (3) scoring is not directly optimized for DST performance. Consequently, the retriever can fail to retrieve examples that would substantially improve DST performance. To address these issues, we present CombiSearch, a method that scores effective in-context examples based on their combinatorial impact on DST performance. Our evaluation on MultiWOZ shows that retrievers trained with CombiSearch surpass state-of-the-art models, achieving a 20x gain in data efficiency and generalizing well to the SGD dataset. Moreover, CombiSearch attains a 12% absolute improvement in the upper bound DST performance over traditional approaches when no retrieval errors are assumed. This significantly increases the headroom for practical DST performance while demonstrating that existing methods rely on suboptimal data for retriever training.
Authors: Li Zhang, Morgan Gray, Jaromir Savelka, Kevin D. Ashley
Abstract: Large Language Models (LLMs) demonstrate potential in complex legal tasks like argument generation, yet their reliability remains a concern. Building upon pilot work assessing LLM generation of 3-ply legal arguments using human evaluation, this paper introduces an automated pipeline to evaluate LLM performance on this task, specifically focusing on faithfulness (absence of hallucination), factor utilization, and appropriate abstention. We define hallucination as the generation of factors not present in the input case materials and abstention as the model's ability to refrain from generating arguments when instructed and no factual basis exists. Our automated method employs an external LLM to extract factors from generated arguments and compares them against the ground-truth factors provided in the input case triples (current case and two precedent cases). We evaluated eight distinct LLMs on three tests of increasing difficulty: 1) generating a standard 3-ply argument, 2) generating an argument with swapped precedent roles, and 3) recognizing the impossibility of argument generation due to lack of shared factors and abstaining. Our findings indicate that while current LLMs achieve high accuracy (over 90%) in avoiding hallucination on viable argument generation tests (Tests 1 & 2), they often fail to utilize the full set of relevant factors present in the cases. Critically, on the abstention test (Test 3), most models failed to follow instructions to stop, instead generating spurious arguments despite the lack of common factors. This automated pipeline provides a scalable method for assessing these crucial LLM behaviors, highlighting the need for improvements in factor utilization and robust abstention capabilities before reliable deployment in legal settings. Link: https://lizhang-aiandlaw.github.io/An-Automated-Pipeline-for-Evaluating-LLM-Generated-3-ply-Case-Based-Legal-Arguments/
Authors: Boheng Sheng, Jiacheng Yao, Meicong Zhang, Guoxiu He
Abstract: Large language models (LLMs) often struggle to accurately read and comprehend extremely long texts. Current methods for improvement typically rely on splitting long contexts into fixed-length chunks. However, fixed truncation risks separating semantically relevant content, leading to ambiguity and compromising accurate understanding. To overcome this limitation, we propose a straightforward approach for dynamically separating and selecting chunks of long context, facilitating a more streamlined input for LLMs. In particular, we compute semantic similarities between adjacent sentences, using lower similarities to adaptively divide long contexts into variable-length chunks. We further train a question-aware classifier to select sensitive chunks that are critical for answering specific questions. Experimental results on both single-hop and multi-hop question-answering benchmarks show that the proposed approach consistently outperforms strong baselines. Notably, it maintains robustness across a wide range of input lengths, handling sequences of up to 256k tokens. Our datasets and code are available at the following link: https://github.com/ECNU-Text-Computing/DCS
Authors: Keyuan Cheng, Zijian Kan, Zhixian He, Zhuoran Zhang, Muhammad Asif Ali, Ke Xu, Lijie Hu, Di Wang
Abstract: Knowledge Editing, which efficiently modifies the knowledge in large language models, has gathered great attention. Current benchmarks primarily use multi-hop question answering to assess and analyze newly injected or updated knowledge. However, we argue that these benchmarks fail to effectively evaluate how well the updated models apply this knowledge in real-life scenarios, particularly when questions require complex reasoning, involving one-to-many relationships or multi-step logical intersections. To fill in this gap, we introduce a new benchmark, COMPKE: Complex Question Answering under Knowledge Editing, which includes 11,924 complex questions that reflect real-life situations. We conduct an extensive evaluation of four knowledge editing methods on COMPKE, revealing that their effectiveness varies notably across different models. For instance, MeLLo attains an accuracy of 39.47 on GPT-4O-MINI, but this drops sharply to 3.83 on QWEN2.5-3B. We further investigate the underlying causes of these disparities from both methodological and model-specific perspectives. The datasets are available at https://github.com/kzjkzj666/CompKE.
Authors: Md Kowsher, Nusrat Jahan Prottasha, Shiyun Xu, Shetu Mohanto, Chen Chen, Ozlem Garibay, Niloofar Yousefi
Abstract: Bidirectional language models have better context understanding and perform better than unidirectional models on natural language understanding tasks, yet the theoretical reasons behind this advantage remain unclear. In this work, we investigate this disparity through the lens of the Information Bottleneck (IB) principle, which formalizes a trade-off between compressing input information and preserving task-relevant content. We propose FlowNIB, a dynamic and scalable method for estimating mutual information during training that addresses key limitations of classical IB approaches, including computational intractability and fixed trade-off schedules. Theoretically, we show that bidirectional models retain more mutual information and exhibit higher effective dimensionality than unidirectional models. To support this, we present a generalized framework for measuring representational complexity and prove that bidirectional representations are strictly more informative under mild conditions. We further validate our findings through extensive experiments across multiple models and tasks using FlowNIB, revealing how information is encoded and compressed throughout training. Together, our work provides a principled explanation for the effectiveness of bidirectional architectures and introduces a practical tool for analyzing information flow in deep language models.
Authors: Yongdong chi, Hanqing Wang, Zonghan Yang, Jian Yang, Xiao Yan, Yun Chen, Guanhua Chen
Abstract: Text-to-SQL transforms the user queries from natural language to executable SQL programs, enabling non-experts to interact with complex databases. Existing prompt-based methods craft meticulous text guidelines and examples to facilitate SQL generation, but their accuracy is hindered by the large semantic gap between the texts and the low-resource SQL programs. In this work, we propose Pi-SQL, which incorporates the high-resource Python program as a pivot to bridge between the natural language query and SQL program. In particular, Pi-SQL first generates Python programs that provide fine-grained step-by-step guidelines in their code blocks or comments, and then produces an SQL program following the guidance of each Python program. The final SQL program matches the reference Python program's query results and, through selection from candidates generated by different strategies, achieves superior execution speed, with a reward-based valid efficiency score up to 4.55 higher than the best-performing baseline. Extensive experiments demonstrate the effectiveness of Pi-SQL, which improves the execution accuracy of the best-performing baseline by up to 3.20.
Authors: Phan Anh Duong, Cat Luong, Divyesh Bommana, Tianyu Jiang
Abstract: Emotions manifest through physical experiences and bodily reactions, yet identifying such embodied emotions in text remains understudied. We present an embodied emotion classification dataset, CHEER-Ekman, extending the existing binary embodied emotion dataset with Ekman's six basic emotion categories. Using automatic best-worst scaling with large language models, we achieve performance superior to supervised approaches on our new dataset. Our investigation reveals that simplified prompting instructions and chain-of-thought reasoning significantly improve emotion recognition accuracy, enabling smaller models to achieve competitive performance with larger ones. Our dataset is publicly available at: https://github.com/menamerai/cheer-ekman.
Authors: Guitao Wang, Jinming Zhao, Hao Yang, Guilin Qi, Tongtong Wu, Gholamreza Haffari
Abstract: Rapid growth in speech data demands adaptive models, as traditional static methods fail to keep pace with dynamic and diverse speech information. We introduce continuous speech learning, a new set-up targeting at bridging the adaptation gap in current speech models. We use the encoder-decoder Whisper model to standardize speech tasks into a generative format. We integrate a learnable gated-fusion layer on the top of the encoder to dynamically select task-specific features for downstream tasks. Our approach improves accuracy significantly over traditional methods in six speech processing tasks, demonstrating gains in adapting to new speech tasks without full retraining.
Authors: Wenhao Liu, Zhenyi Lu, Xinyu Hu, Jierui Zhang, Dailin Li, Jiacheng Cen, Huilin Cao, Haiteng Wang, Yuhan Li, Kun Xie, Dandan Li, Pei Zhang, Chengbo Zhang, Yuxiang Ren, Xiaohong Huang, Yan Ma
Abstract: High-quality math datasets are crucial for advancing the reasoning abilities of large language models (LLMs). However, existing datasets often suffer from three key issues: outdated and insufficient challenging content, neglecting human-like reasoning, and limited reliability due to single-LLM generation. To address these, we introduce STORM-BORN, an ultra-challenging dataset of mathematical derivations sourced from cutting-edge academic papers, which includes dense human-like approximations and heuristic cues. To ensure the reliability and quality, we propose a novel human-in-the-loop, multi-agent data generation framework, integrating reasoning-dense filters, multi-agent collaboration, and human mathematicians' evaluations. We curated a set of 2,000 synthetic samples and deliberately selected the 100 most difficult problems. Even most advanced models like GPT-o1 solved fewer than 5% of them. Fine-tuning on STORM-BORN boosts accuracy by 7.84% (LLaMA3-8B) and 9.12% (Qwen2.5-7B). As AI approaches mathematician-level reasoning, STORM-BORN provides both a high-difficulty benchmark and a human-like reasoning training resource. Our code and dataset are publicly available at https://github.com/lwhere/STORM-BORN.
Authors: Pasunuti Prasanjith, Prathmesh B More, Anoop Kunchukuttan, Raj Dabre
Abstract: Retrieval-Augmented Generation (RAG) systems enable language models to access relevant information and generate accurate, well-grounded, and contextually informed responses. However, for Indian languages, the development of high-quality RAG systems is hindered by the lack of two critical resources: (1) evaluation benchmarks for retrieval and generation tasks, and (2) large-scale training datasets for multilingual retrieval. Most existing benchmarks and datasets are centered around English or high-resource languages, making it difficult to extend RAG capabilities to the diverse linguistic landscape of India. To address the lack of evaluation benchmarks, we create IndicMSMarco, a multilingual benchmark for evaluating retrieval quality and response generation in 13 Indian languages, created via manual translation of 1000 diverse queries from MS MARCO-dev set. To address the need for training data, we build a large-scale dataset of (question, answer, relevant passage) tuples derived from the Wikipedias of 19 Indian languages using state-of-the-art LLMs. Additionally, we include translated versions of the original MS MARCO dataset to further enrich the training data and ensure alignment with real-world information-seeking tasks. Resources are available here: https://huggingface.co/collections/ai4bharat/indicragsuite-683e7273cb2337208c8c0fcb
URLs: https://huggingface.co/collections/ai4bharat/indicragsuite-683e7273cb2337208c8c0fcb
Authors: Yile Liu, Ziwei Ma, Xiu Jiang, Jinglu Hu, Jing Chang, Liang Li
Abstract: With the rapid adoption of large language models (LLMs) in natural language processing, the ability to follow instructions has emerged as a key metric for evaluating their practical utility. However, existing evaluation methods often focus on single-language scenarios, overlooking the challenges and differences present in multilingual and cross-lingual contexts. To address this gap, we introduce MaXIFE: a comprehensive evaluation benchmark designed to assess instruction-following capabilities across 23 different languages with 1667 verifiable instruction tasks. MaXIFE integrates both Rule-Based Evaluation and Model-Based Evaluation, ensuring a balance of efficiency and accuracy. We applied MaXIFE to evaluate several leading commercial LLMs, establishing baseline results for future comparisons. By providing a standardized tool for multilingual instruction-following evaluation, MaXIFE aims to advance research and development in natural language processing.
Authors: Estelle Ruellan, Masarah Paquet-Clouston, Sebastian Garcia
Abstract: Ransomware-as-a-service (RaaS) is increasing the scale and complexity of ransomware attacks. Understanding the internal operations behind RaaS has been a challenge due to the illegality of such activities. The recent chat leak of the Conti RaaS operator, one of the most infamous ransomware operators on the international scene, offers a key opportunity to better understand the inner workings of such organizations. This paper analyzes the main topic discussions in the Conti chat leak using machine learning techniques such as Natural Language Processing (NLP) and Latent Dirichlet Allocation (LDA), as well as visualization strategies. Five discussion topics are found: 1) Business, 2) Technical, 3) Internal tasking/Management, 4) Malware, and 5) Customer Service/Problem Solving. Moreover, the distribution of topics among Conti members shows that only 4% of individuals have specialized discussions while almost all individuals (96%) are all-rounders, meaning that their discussions revolve around the five topics. The results also indicate that a significant proportion of Conti discussions are non-tech related. This study thus highlights that running such large RaaS operations requires a workforce skilled beyond technical abilities, with individuals involved in various tasks, from management to customer service or problem solving. The discussion topics also show that the organization behind the Conti RaaS oper5086933ator shares similarities with a large firm. We conclude that, although RaaS represents an example of specialization in the cybercrime industry, only a few members are specialized in one topic, while the rest runs and coordinates the RaaS operation.
Authors: Xinyue Zheng, Haowei Lin, Kaichen He, Zihao Wang, Zilong Zheng, Yitao Liang
Abstract: Developing AI agents capable of interacting with open-world environments to solve diverse tasks is a compelling challenge. However, evaluating such open-ended agents remains difficult, with current benchmarks facing scalability limitations. To address this, we introduce Minecraft Universe (MCU), a comprehensive evaluation framework set within the open-world video game Minecraft. MCU incorporates three key components: (1) an expanding collection of 3,452 composable atomic tasks that encompasses 11 major categories and 41 subcategories of challenges; (2) a task composition mechanism capable of generating infinite diverse tasks with varying difficulty; and (3) a general evaluation framework that achieves 91.5\% alignment with human ratings for open-ended task assessment. Empirical results reveal that even state-of-the-art foundation agents struggle with the increasing diversity and complexity of tasks. These findings highlight the necessity of MCU as a robust benchmark to drive progress in AI agent development within open-ended environments. Our evaluation code and scripts are available at https://github.com/CraftJarvis/MCU.
Authors: Mart van Baalen, Andrey Kuzmin, Ivan Koryakovskiy, Markus Nagel, Peter Couperus, Cedric Bastoul, Eric Mahurin, Tijmen Blankevoort, Paul Whatmough
Abstract: In this work we show that the size versus accuracy trade-off of neural network quantization can be significantly improved by increasing the quantization dimensionality. We propose the GPTVQ method, a new fast method for post-training vector quantization (VQ) that scales well to Large Language Models (LLMs). Our method interleaves quantization of one or more columns with updates to the remaining unquantized weights, using information from the Hessian of the per-layer output reconstruction MSE. Quantization codebooks are initialized using an efficient data-aware version of the EM algorithm. The codebooks are then updated, and further compressed by using integer quantization and SVD-based compression. GPTVQ establishes a new state-of-the art in the size vs accuracy trade-offs on a wide range of LLMs such as Llama-v2 and Mistral. Furthermore, our method is efficient: on a single H100 it takes between 3 and 11 hours to process a Llamav2-70B model, depending on quantization setting. Lastly, with on-device timings for VQ decompression on a mobile CPU we show that VQ leads to improved latency compared to using a 4-bit integer format.
Authors: Anselm Paulus, Arman Zharmagambetov, Chuan Guo, Brandon Amos, Yuandong Tian
Abstract: Large Language Models (LLMs) are vulnerable to jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires a time-consuming search for adversarial prompts, whereas automatic adversarial prompt generation often leads to semantically meaningless attacks that do not scale well. In this paper, we present a novel method that uses another LLM, called AdvPrompter, to generate human-readable adversarial prompts in seconds. AdvPrompter, which is trained using an alternating optimization algorithm, generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show highly competitive results on the AdvBench and HarmBench datasets, that also transfer to closed-source black-box LLMs. We also show that training on adversarial suffixes generated by AdvPrompter is a promising strategy for improving the robustness of LLMs to jailbreaking attacks.
Authors: Leshem Choshen, Yang Zhang, Jacob Andreas
Abstract: Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pretraining decisions involving optimizers, datasets, and model architectures. Despite the widespread use of scaling laws to model the dynamics of language model training, there has been little work on understanding how to best estimate and interpret them. We collect (and release) a large-scale dataset containing losses and downstream evaluations for 485 previously published pretrained models. We use these to estimate more than 1000 scaling laws, then derive a set of best practices for estimating scaling laws in new model families. We find that fitting scaling laws to intermediate checkpoints of training runs (and not just their final losses) substantially improves accuracy, and that -- all else equal -- estimates of performance are generally most accurate when derived from other models of similar sizes. However, because there is a significant degree of variability across model seeds, training multiple small models is sometimes more useful than training a single large one. Moreover, while different model families differ scaling behavior, they are often similar enough that a target model's behavior can be predicted from a single model with the same architecture, along with scaling parameter estimates derived from other model families.
Authors: Seongmin Lee (Polo), Hsiang Hsu (Polo), Chun-Fu Chen (Polo), Duen Horng (Polo), Chau
Abstract: LLM hallucination, where unfaithful text is generated, presents a critical challenge for LLMs' practical applications. Current detection methods often resort to external knowledge, LLM fine-tuning, or supervised training with large hallucination-labeled datasets. Moreover, these approaches do not distinguish between different types of hallucinations, which is crucial for enhancing detection performance. To address such limitations, we introduce hallucination probing, a new task that classifies LLM-generated text into three categories: aligned, misaligned, and fabricated. Driven by our novel discovery that perturbing key entities in prompts affects LLM's generation of these three types of text differently, we propose SHINE, a novel hallucination probing method that does not require external knowledge, supervised training, or LLM fine-tuning. SHINE is effective in hallucination probing across three modern LLMs, and achieves state-of-the-art performance in hallucination detection, outperforming seven competing methods across four datasets and four LLMs, underscoring the importance of probing for accurate detection.
Authors: Feng Chen, Chenhui Gou, Jing Liu, Yang Yang, Zhaoyang Li, Jiyuan Zhang, Zhenbang Sun, Bohan Zhuang, Qi Wu
Abstract: As multimodal large language models (MLLMs) advance rapidly, rigorous evaluation has become essential, providing further guidance for their development. In this work, we focus on a unified and robust evaluation of \textbf{vision perception} abilities, the foundational skill of MLLMs. We find that existing perception benchmarks, each focusing on different question types, domains, and evaluation metrics, introduce significant evaluation variance, complicating comprehensive assessments of perception abilities when relying on any single benchmark. To address this, we introduce \textbf{AbilityLens}, a unified benchmark designed to evaluate MLLMs in six key perception abilities (ranging from counting, OCR, to understanding structural data), focusing on both accuracy and stability, with each ability encompassing diverse types of questions, domains, and metrics. With the assistance of AbilityLens, we: (1) identify the strengths and weaknesses of current main-stream MLLMs, highlighting stability patterns and revealing a notable performance gap between state-of-the-art open-source and closed-source models; (2) uncover interesting ability conflict and early convergence phenomena during MLLM training; (3) reveal the primary reason of ability conflict is data mixing ratio and LLM model size; and (4) discuss the effectiveness of some straightforward strategies \eg, fine-tuning and model merging, to solve the ability conflict. The benchmark and online leaderboard is released in https://github.com/Chenfeng1271/AbilityLens.
Authors: Peter G. Brodeur, Thomas A. Buckley, Zahir Kanjee, Ethan Goh, Evelyn Bin Ling, Priyank Jain, Stephanie Cabral, Raja-Elie Abdulnour, Adrian D. Haimovich, Jason A. Freed, Andrew Olson, Daniel J. Morgan, Jason Hom, Robert Gallo, Liam G. McCoy, Haadi Mombini, Christopher Lucas, Misha Fotoohi, Matthew Gwiazdon, Daniele Restifo, Daniel Restrepo, Eric Horvitz, Jonathan Chen, Arjun K. Manrai, Adam Rodman
Abstract: A seminal paper published by Ledley and Lusted in 1959 introduced complex clinical diagnostic reasoning cases as the gold standard for the evaluation of expert medical computing systems, a standard that has held ever since. Here, we report the results of a physician evaluation of a large language model (LLM) on challenging clinical cases against a baseline of hundreds of physicians. We conduct five experiments to measure clinical reasoning across differential diagnosis generation, display of diagnostic reasoning, triage differential diagnosis, probabilistic reasoning, and management reasoning, all adjudicated by physician experts with validated psychometrics. We then report a real-world study comparing human expert and AI second opinions in randomly-selected patients in the emergency room of a major tertiary academic medical center in Boston, MA. We compared LLMs and board-certified physicians at three predefined diagnostic touchpoints: triage in the emergency room, initial evaluation by a physician, and admission to the hospital or intensive care unit. In all experiments--both vignettes and emergency room second opinions--the LLM displayed superhuman diagnostic and reasoning abilities, as well as continued improvement from prior generations of AI clinical decision support. Our study suggests that LLMs have achieved superhuman performance on general medical diagnostic and management reasoning, fulfilling the vision put forth by Ledley and Lusted, and motivating the urgent need for prospective trials.
Authors: Xiaoning Dong, Wenbo Hu, Wei Xu, Tianxing He
Abstract: Large language models (LLMs) have made significant advancements across various tasks, but their safety alignment remain a major concern. Exploring jailbreak prompts can expose LLMs' vulnerabilities and guide efforts to secure them. Existing methods primarily design sophisticated instructions for the LLM to follow, or rely on multiple iterations, which could hinder the performance and efficiency of jailbreaks. In this work, we propose a novel jailbreak paradigm, Simple Assistive Task Linkage (SATA), which can effectively circumvent LLM safeguards and elicit harmful responses. Specifically, SATA first masks harmful keywords within a malicious query to generate a relatively benign query containing one or multiple [MASK] special tokens. It then employs a simple assistive task such as a masked language model task or an element lookup by position task to encode the semantics of the masked keywords. Finally, SATA links the assistive task with the masked query to jointly perform the jailbreak. Extensive experiments show that SATA achieves state-of-the-art performance and outperforms baselines by a large margin. Specifically, on AdvBench dataset, with mask language model (MLM) assistive task, SATA achieves an overall attack success rate (ASR) of 85% and harmful score (HS) of 4.57, and with element lookup by position (ELP) assistive task, SATA attains an overall ASR of 76% and HS of 4.43.
Authors: Yang Xu, Yi Wang, Hengguan Huang, Hao Wang
Abstract: Understanding training dynamics and feature evolution is crucial for the mechanistic interpretability of large language models (LLMs). Although sparse autoencoders (SAEs) have been used to identify features within LLMs, a clear picture of how these features evolve during training remains elusive. In this study, we (1) introduce SAE-Track, a novel method for efficiently obtaining a continual series of SAEs, providing the foundation for a mechanistic study that covers (2) the semantic evolution of features, (3) the underlying processes of feature formation, and (4) the directional drift of feature vectors. Our work provides new insights into the dynamics of features in LLMs, enhancing our understanding of training mechanisms and feature evolution. For reproducibility, our code is available at https://github.com/Superposition09m/SAE-Track.
Authors: Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, Yongming Rao
Abstract: Recent advances in large language models, particularly following GPT-4o, have sparked increasing interest in developing omni-modal models capable of understanding more modalities. While some open-source alternatives have emerged, there is still a notable lag behind specialized single-modality models in performance. In this paper, we present Ola, an Omni-modal Language model that achieves competitive performance across image, video, and audio understanding compared to specialized counterparts, pushing the frontiers of the omni-modal language model to a large extent. We conduct a comprehensive exploration of architectural design, data curation, and training strategies essential for building a robust omni-modal model. Ola incorporates advanced visual understanding and audio recognition capabilities through several critical and effective improvements over mainstream baselines. Moreover, we rethink inter-modal relationships during omni-modal training, emphasizing cross-modal alignment with video as a central bridge, and propose a progressive training pipeline that begins with the most distinct modalities and gradually moves towards closer modality alignment. Extensive experiments demonstrate that Ola surpasses existing open omni-modal LLMs across all modalities while achieving highly competitive performance compared to state-of-the-art specialized models of similar sizes. We aim to make Ola a fully open omni-modal understanding solution to advance future research in this emerging field. Model weights, code, and data are open-sourced at https://github.com/Ola-Omni/Ola.
Authors: Gaurush Hiranandani, Haolun Wu, Subhojyoti Mukherjee, Sanmi Koyejo
Abstract: Many commercial Large Language Models (LLMs) are often closed-source, limiting developers to prompt tuning for aligning content generation with specific applications. While these models currently do not provide access to token logits, we argue that if such access were available, it would enable more powerful adaptation techniques beyond prompt engineering. In this paper, we propose a token-level probability reweighting framework that, given access to logits and a small amount of task-specific data, can effectively steer black-box LLMs toward application-specific content generation. Our approach views next-token prediction through the lens of supervised classification. We show that aligning black-box LLMs with task-specific data can be formulated as a label noise correction problem, leading to Plugin model -- an autoregressive probability reweighting model that operates solely on logits. We provide theoretical justification for why reweighting logits alone is sufficient for task adaptation. Extensive experiments with multiple datasets, LLMs, and reweighting models demonstrate the effectiveness of our method, advocating for broader access to token logits in closed-source models.
Authors: Shengguang Wu, Fan-Yun Sun, Kaiyue Wen, Nick Haber
Abstract: Recent studies have shown that Large Vision-Language Models (VLMs) tend to neglect image content and over-rely on language-model priors, resulting in errors in visually grounded tasks and hallucinations. We hypothesize that this issue arises because existing VLMs are not explicitly trained to generate texts that are accurately grounded in fine-grained image details. To enhance visual feedback during VLM training, we propose S-VCO (Symmetrical Visual Contrastive Optimization), a novel finetuning objective that steers the model toward capturing important visual details and aligning them with corresponding text tokens. To further facilitate this detailed alignment, we introduce MVC, a paired image-text dataset built by automatically filtering and augmenting visual counterfactual data to challenge the model with hard contrastive cases involving Minimal Visual Contrasts. Experiments show that our method consistently improves VLM performance across diverse benchmarks covering various abilities and domains, achieving up to a 22% reduction in hallucinations, and significant gains in vision-centric and general tasks. Notably, these improvements become increasingly pronounced in benchmarks with higher visual dependency. In short, S-VCO offers a significant enhancement of VLM's visually-dependent task performance while retaining or even improving the model's general abilities. We opensource our code at https://s-vco.github.io/
URLs: https://s-vco.github.io/
Authors: Moxin Li, Yuantao Zhang, Wenjie Wang, Wentao Shi, Zhuo Liu, Fuli Feng, Tat-Seng Chua
Abstract: Multi-Objective Alignment (MOA) aims to align LLMs' responses with multiple human preference objectives, with Direct Preference Optimization (DPO) emerging as a prominent approach. However, we find that DPO-based MOA approaches suffer from widespread preference conflicts in the data, where different objectives favor different responses. This results in conflicting optimization directions, hindering the optimization on the Pareto Front. To address this, we propose to construct Pareto-optimal responses to resolve preference conflicts. To efficiently obtain and utilize such responses, we propose a self-improving DPO framework that enables LLMs to self-generate and select Pareto-optimal responses for self-supervised preference alignment. Extensive experiments on two datasets demonstrate the superior Pareto Front achieved by our framework compared to various baselines. Code is available at https://github.com/zyttt-coder/SIPO.
Authors: Yang Yao, Xuan Tong, Ruofan Wang, Yixu Wang, Lujundong Li, Liang Liu, Yan Teng, Yingchun Wang
Abstract: Large Reasoning Models (LRMs) have significantly advanced beyond traditional Large Language Models (LLMs) with their exceptional logical reasoning capabilities, yet these improvements introduce heightened safety risks. When subjected to jailbreak attacks, their ability to generate more targeted and organized content can lead to greater harm. Although some studies claim that reasoning enables safer LRMs against existing LLM attacks, they overlook the inherent flaws within the reasoning process itself. To address this gap, we propose the first jailbreak attack targeting LRMs, exploiting their unique vulnerabilities stemming from the advanced reasoning capabilities. Specifically, we introduce a Chaos Machine, a novel component to transform attack prompts with diverse one-to-one mappings. The chaos mappings iteratively generated by the machine are embedded into the reasoning chain, which strengthens the variability and complexity and also promotes a more robust attack. Based on this, we construct the Mousetrap framework, which makes attacks projected into nonlinear-like low sample spaces with mismatched generalization enhanced. Also, due to the more competing objectives, LRMs gradually maintain the inertia of unpredictable iterative reasoning and fall into our trap. Success rates of the Mousetrap attacking o1-mini, Claude-Sonnet and Gemini-Thinking are as high as 96%, 86% and 98% respectively on our toxic dataset Trotter. On benchmarks such as AdvBench, StrongREJECT, and HarmBench, attacking Claude-Sonnet, well-known for its safety, Mousetrap can astonishingly achieve success rates of 87.5%, 86.58% and 93.13% respectively. Attention: This paper contains inappropriate, offensive and harmful content.
Authors: Jibang Wu, Chenghao Yang, Simon Mahns, Chaoqi Wang, Hao Zhu, Fei Fang, Haifeng Xu
Abstract: This paper develops an agentic framework that employs large language models (LLMs) to automate the generation of persuasive and grounded marketing content, using real estate listing descriptions as our focal application domain. Our method is designed to align the generated content with user preferences while highlighting useful factual attributes. This agent consists of three key modules: (1) Grounding Module, mimicking expert human behavior to predict marketable features; (2) Personalization Module, aligning content with user preferences; (3) Marketing Module, ensuring factual accuracy and the inclusion of localized features. We conduct systematic human-subject experiments in the domain of real estate marketing, with a focus group of potential house buyers. The results demonstrate that marketing descriptions generated by our approach are preferred over those written by human experts by a clear margin while maintaining the same level of factual accuracy. Our findings suggest a promising agentic approach to automate large-scale targeted marketing while ensuring factuality of content generation.
Authors: Qiuchen Wang, Ruixue Ding, Zehui Chen, Weiqi Wu, Shihang Wang, Pengjun Xie, Feng Zhao
Abstract: Understanding information from visually rich documents remains a significant challenge for traditional Retrieval-Augmented Generation (RAG) methods. Existing benchmarks predominantly focus on image-based question answering (QA), overlooking the fundamental challenges of efficient retrieval, comprehension, and reasoning within dense visual documents. To bridge this gap, we introduce ViDoSeek, a novel dataset designed to evaluate RAG performance on visually rich documents requiring complex reasoning. Based on it, we identify key limitations in current RAG approaches: (i) purely visual retrieval methods struggle to effectively integrate both textual and visual features, and (ii) previous approaches often allocate insufficient reasoning tokens, limiting their effectiveness. To address these challenges, we propose ViDoRAG, a novel multi-agent RAG framework tailored for complex reasoning across visual documents. ViDoRAG employs a Gaussian Mixture Model (GMM)-based hybrid strategy to effectively handle multi-modal retrieval. To further elicit the model's reasoning capabilities, we introduce an iterative agent workflow incorporating exploration, summarization, and reflection, providing a framework for investigating test-time scaling in RAG domains. Extensive experiments on ViDoSeek validate the effectiveness and generalization of our approach. Notably, ViDoRAG outperforms existing methods by over 10% on the competitive ViDoSeek benchmark. The code is available at https://github.com/Alibaba-NLP/ViDoRAG.
Authors: Eliahu Horwitz, Nitzan Kurer, Jonathan Kahana, Liel Amar, Yedid Hoshen
Abstract: Public model repositories now contain millions of models, yet most models remain undocumented and effectively lost. In this position paper, we advocate for charting the world's model population in a unified structure we call the Model Atlas: a graph that captures models, their attributes, and the weight transformations that connect them. The Model Atlas enables applications in model forensics, meta-ML research, and model discovery, challenging tasks given today's unstructured model repositories. However, because most models lack documentation, large atlas regions remain uncharted. Addressing this gap motivates new machine learning methods that treat models themselves as data, inferring properties such as functionality, performance, and lineage directly from their weights. We argue that a scalable path forward is to bypass the unique parameter symmetries that plague model weights. Charting all the world's models will require a community effort, and we hope its broad utility will rally researchers toward this goal.
Authors: Selim Jerad, Anej Svete, Jiaoda Li, Ryan Cotterell
Abstract: Understanding the expressive power of transformers has recently attracted attention, as it offers insights into their abilities and limitations. Many studies analyze unique hard attention transformers, where attention selects a single position that maximizes the attention scores. When multiple positions achieve the maximum score, either the rightmost or the leftmost of those is chosen. In this paper, we highlight the importance of this seeming triviality. Recently, finite-precision transformers with both leftmost- and rightmost-hard attention were shown to be equivalent to Linear Temporal Logic (LTL). We show that this no longer holds with only leftmost-hard attention -- in that case, they correspond to a \emph{strictly weaker} fragment of LTL. Furthermore, we show that models with leftmost-hard attention are equivalent to \emph{soft} attention, suggesting they may better approximate real-world transformers than right-attention models. These findings refine the landscape of transformer expressivity and underscore the role of attention directionality.
Authors: Deyu Cao, Samin Aref
Abstract: The growing use of large language models has raised environmental and economic concerns about their intensity of resource usage during inference. Serving these models to each user requires substantial energy and water for cooling. Model compression techniques like quantization can shrink large language models and make them more resource efficient at the cost of potential performance degradation. Quantization methods compress model size through replacing their high-precision parameters by quantized values of lower precision. Among existing methods, the ApiQ method achieves superior accuracy preservation at minimal memory and time overhead. We investigate two ideas to extend performance in ultra-low-bit quantization beyond ApiQ's level. First, we look into combining existing quantization-aware training techniques with ApiQ's partial training. We show that this does not outperform the baseline ApiQ method with limited training data and frozen weights. This leads to two key insights: (1) The substantial representational capacity that is gained through full retraining is unlikely to be feasible through partial training. (2) This gain may depend on using a large and diverse dataset in quantization-aware training. Second, through a novel approach informed by the two insights, we propose an ultra-low-bit quantization method that builds upon ApiQ and extends its performance without the need for full retraining. This publicly available method relies on a saliency-aware regularization term that prioritizes preserving the most impactful parameters during quantization. Our experiments on LLaMA 7B and 13B benchmarks demonstrate that our method reduces the ApiQ's accuracy degradation by 10.85\% and 7.54\% respectively.
Authors: Fangming Cui, Yonggang Zhang, Xuan Wang, Xinmei Tian, Jun Yu
Abstract: Recent developments in prompt learning of large Vision-Language Models (VLMs) have significantly improved performance in target-specific tasks. However, these prompting methods often struggle to tackle the target-unspecific or generalizable tasks effectively. It may be attributed to the fact that overfitting training causes the model to forget its general knowledge. The general knowledge has a strong promotion on target-unspecific tasks. To alleviate this issue, we propose a novel Features Matrix (FM) approach designed to enhance these models on target-unspecific tasks. Our method extracts and leverages general knowledge, shaping a Features Matrix (FM). Specifically, the FM captures the semantics of diverse inputs from a deep and fine perspective, preserving essential general knowledge, which mitigates the risk of overfitting. Representative evaluations demonstrate that: 1) the FM is compatible with existing frameworks as a generic and flexible module, and 2) the FM significantly showcases its effectiveness in enhancing target-unspecific tasks (base-to-novel generalization, domain generalization, and cross-dataset generalization), achieving state-of-the-art performance.
Authors: Wei Liu, Jiyuan Zhang, Binxiong Zheng, Yufeng Hu, Yingzhan Lin, Zengfeng Zeng
Abstract: End-to-end autonomous driving has advanced significantly, offering benefits such as system simplicity and stronger driving performance in both open-loop and closed-loop settings than conventional pipelines. However, existing frameworks still suffer from low success rates in closed-loop evaluations, highlighting their limitations in real-world deployment. In this paper, we introduce X-Driver, a unified multi-modal large language models(MLLMs) framework designed for closed-loop autonomous driving, leveraging Chain-of-Thought(CoT) and autoregressive modeling to enhance perception and decision-making. We validate X-Driver across multiple autonomous driving tasks using public benchmarks in CARLA simulation environment, including Bench2Drive[6]. Our experimental results demonstrate superior closed-loop performance, surpassing the current state-of-the-art(SOTA) while improving the interpretability of driving decisions. These findings underscore the importance of structured reasoning in end-to-end driving and establish X-Driver as a strong baseline for future research in closed-loop autonomous driving.
Authors: Junyang Wang, Haiyang Xu, Xi Zhang, Ming Yan, Ji Zhang, Fei Huang, Jitao Sang
Abstract: The exponential rise in mobile device usage necessitates streamlined automation for effective task management, yet many AI frameworks fall short due to inadequate operational expertise. While manually written knowledge can bridge this gap, it is often burdensome and inefficient. We introduce Mobile-Agent-V, an innovative framework that utilizes video as a guiding tool to effortlessly and efficiently inject operational knowledge into mobile automation processes. By deriving knowledge directly from video content, Mobile-Agent-V eliminates manual intervention, significantly reducing the effort and time required for knowledge acquisition. To rigorously evaluate this approach, we propose Mobile-Knowledge, a benchmark tailored to assess the impact of external knowledge on mobile agent performance. Our experimental findings demonstrate that Mobile-Agent-V enhances performance by 36% compared to existing methods, underscoring its effortless and efficient advantages in mobile automation.
Authors: Noah Amsel, David Persson, Christopher Musco, Robert M. Gower
Abstract: Computing the polar decomposition and the related matrix sign function, has been a well-studied problem in numerical analysis for decades. More recently, it has emerged as an important subroutine in deep learning, particularly within the Muon optimization framework. However, the requirements in this setting differ significantly from those of traditional numerical analysis. In deep learning, methods must be highly efficient and GPU-compatible, but high accuracy is often unnecessary. As a result, classical algorithms like Newton-Schulz (which suffers from slow initial convergence) and methods based on rational functions (which rely on QR decompositions or matrix inverses) are poorly suited to this context. In this work, we introduce Polar Express, a GPU-friendly algorithm for computing the polar decomposition. Like classical polynomial methods such as Newton-Schulz, our approach uses only matrix-matrix multiplications, making it GPU-compatible. Motivated by earlier work of Chen & Chow and Nakatsukasa & Freund, Polar Express adapts the polynomial update rule at each iteration by solving a minimax optimization problem, and we prove that it enjoys a strong worst-case optimality guarantee. This property ensures both rapid early convergence and fast asymptotic convergence. We also address finite-precision issues, making it stable in bfloat16 in practice. We apply Polar Express within the Muon optimization framework and show consistent improvements in validation loss on large-scale models such as GPT-2, outperforming recent alternatives across a range of learning rates.
Authors: Shikhhar Siingh, Abhinav Rawat, Chitta Baral, Vivek Gupta
Abstract: Publicly significant images from events hold valuable contextual information, crucial for journalism and education. However, existing methods often struggle to extract this relevance accurately. To address this, we introduce GETReason (Geospatial Event Temporal Reasoning), a framework that moves beyond surface-level image descriptions to infer deeper contextual meaning. We propose that extracting global event, temporal, and geospatial information enhances understanding of an image's significance. Additionally, we introduce GREAT (Geospatial Reasoning and Event Accuracy with Temporal Alignment), a new metric for evaluating reasoning-based image understanding. Our layered multi-agent approach, assessed using a reasoning-weighted metric, demonstrates that meaningful insights can be inferred, effectively linking images to their broader event context.
Authors: Qingzheng Wang, Jiancheng Sun, Yifan Peng, Shinji Watanabe
Abstract: Multilingual speech processing with self-supervised or supervised pre-trained Speech Foundation Models (SFM) has achieved strong performance on tasks like Language Identification (LID) and Automatic Speech Recognition (ASR). However, these models struggle with limited resources during fine-tuning. This paper enhances multilingual LID and ASR on ML-SUPERB 2.0 by exploring multiple strategies for adapting SFMs, including frozen upstream training, partial fine-tuning, and low-rank adaptation. Furthermore, we employ data augmentation to mitigate performance gaps in few-shot settings and introduce LID Connectionist Temporal Classification (CTC) loss for regularization. Our approach achieves a 14% relative improvement in LID accuracy and a 30% relative reduction in ASR CER over the baseline on ML-SUPERB 2.0, securing second place in the Interspeech 2025 ML-SUPERB 2.0 Challenge.
Authors: Yuchong Li, Xiaojun Zeng, Chihua Fang, Jian Yang, Fucang Jia, Lei Zhang
Abstract: Hepato-pancreato-biliary (HPB) disorders represent a global public health challenge due to their high morbidity and mortality. Although large language models (LLMs) have shown promising performance in general medical question-answering tasks, the current evaluation benchmarks are mostly derived from standardized examinations or manually designed questions, lacking HPB coverage and clinical cases. To address these issues, we systematically eatablish an HPB disease evaluation benchmark comprising 3,535 closed-ended multiple-choice questions and 337 open-ended real diagnosis cases, which encompasses all the 33 main categories and 465 subcategories of HPB diseases defined in the International Statistical Classification of Diseases, 10th Revision (ICD-10). The multiple-choice questions are curated from public datasets and synthesized data, and the clinical cases are collected from prestigious medical journals, case-sharing platforms, and collaborating hospitals. By evalauting commercial and open-source general and medical LLMs on our established benchmark, namely ClinBench-HBP, we find that while commercial LLMs perform competently on medical exam questions, they exhibit substantial performance degradation on HPB diagnosis tasks, especially on complex, inpatient clinical cases. Those medical LLMs also show limited generalizability to HPB diseases. Our results reveal the critical limitations of current LLMs in the domain of HPB diseases, underscoring the imperative need for future medical LLMs to handle real, complex clinical diagnostics rather than simple medical exam questions. The benchmark will be released at https://clinbench-hpb.github.io.
Authors: Xiaochen Wang, Zongyu Wu, Yuan Zhong, Xiang Zhang, Suhang Wang, Fenglong Ma
Abstract: Graph retrieval-augmented generation (GRAG) places high demands on graph-specific retrievers. However, existing retrievers often rely on language models pretrained on plain text, limiting their effectiveness due to domain misalignment and structure ignorance. To address these challenges, we propose GPR, a graph-based retriever pretrained directly on knowledge graphs. GPR aligns natural language questions with relevant subgraphs through LLM-guided graph augmentation and employs a structure-aware objective to learn fine-grained retrieval strategies. Experiments on two datasets, three LLM backbones, and five baselines show that GPR consistently improves both retrieval quality and downstream generation, demonstrating its effectiveness as a robust retrieval solution for GRAG.
Authors: Femi Bello, Anubrata Das, Fanzhi Zeng, Fangcong Yin, Liu Leqi
Abstract: It has been hypothesized that neural networks with similar architectures trained on similar data learn shared representations relevant to the learning task. We build on this idea by extending the conceptual framework where representations learned across models trained on the same data can be expressed as linear combinations of a \emph{universal} set of basis features. These basis features underlie the learning task itself and remain consistent across models, regardless of scale. From this framework, we propose the \textbf{Linear Representation Transferability (LRT)} Hypothesis -- that there exists an affine transformation between the representation spaces of different models. To test this hypothesis, we learn affine mappings between the hidden states of models of different sizes and evaluate whether steering vectors -- directions in hidden state space associated with specific model behaviors -- retain their semantic effect when transferred from small to large language models using the learned mappings. We find strong empirical evidence that such affine mappings can preserve steering behaviors. These findings suggest that representations learned by small models can be used to guide the behavior of large models, and that the LRT hypothesis may be a promising direction on understanding representation alignment across model scales.
Authors: Genta Indra Winata, David Anugraha, Emmy Liu, Alham Fikri Aji, Shou-Yi Hung, Aditya Parashar, Patrick Amadeus Irawan, Ruochen Zhang, Zheng-Xin Yong, Jan Christian Blaise Cruz, Niklas Muennighoff, Seungone Kim, Hanyang Zhao, Sudipta Kar, Kezia Erina Suryoraharjo, M. Farid Adilazuarda, En-Shiun Annie Lee, Ayu Purwarianti, Derry Tanti Wijaya, Monojit Choudhury
Abstract: High-quality datasets are fundamental to training and evaluating machine learning models, yet their creation-especially with accurate human annotations-remains a significant challenge. Many dataset paper submissions lack originality, diversity, or rigorous quality control, and these shortcomings are often overlooked during peer review. Submissions also frequently omit essential details about dataset construction and properties. While existing tools such as datasheets aim to promote transparency, they are largely descriptive and do not provide standardized, measurable methods for evaluating data quality. Similarly, metadata requirements at conferences promote accountability but are inconsistently enforced. To address these limitations, this position paper advocates for the integration of systematic, rubric-based evaluation metrics into the dataset review process-particularly as submission volumes continue to grow. We also explore scalable, cost-effective methods for synthetic data generation, including dedicated tools and LLM-as-a-judge approaches, to support more efficient evaluation. As a call to action, we introduce DataRubrics, a structured framework for assessing the quality of both human- and model-generated datasets. Leveraging recent advances in LLM-based evaluation, DataRubrics offers a reproducible, scalable, and actionable solution for dataset quality assessment, enabling both authors and reviewers to uphold higher standards in data-centric research. We also release code to support reproducibility of LLM-based evaluations at https://github.com/datarubrics/datarubrics.