Authors: Toyin Aguda, Erik Wilson, Allan Anzagira, Simerjot Kaur, Charese Smiley
Abstract: Large language models (LLMs) exhibit pronounced conservative bias in relation extraction tasks, frequently defaulting to No_Relation label when an appropriate option is unavailable. While this behavior helps prevent incorrect relation assignments, our analysis reveals that it also leads to significant information loss when reasoning is not explicitly included in the output. We systematically evaluate this trade-off across multiple prompts, datasets, and relation types, introducing the concept of Hobson's choice to capture scenarios where models opt for safe but uninformative labels over hallucinated ones. Our findings suggest that conservative bias occurs twice as often as hallucination. To quantify this effect, we use SBERT and LLM prompts to capture the semantic similarity between conservative bias behaviors in constrained prompts and labels generated from semi-constrained and open-ended prompts.
Authors: Jacob Dineen (Arizona State University), Aswin RRV (Arizona State University), Qin Liu (University of California Davis), Zhikun Xu (Arizona State University), Xiao Ye (Arizona State University), Ming Shen (Arizona State University), Zhaonan Li (Arizona State University), Shijie Lu (Arizona State University), Chitta Baral (Arizona State University), Muhao Chen (University of California Davis), Ben Zhou (Arizona State University)
Abstract: Alignment of large language models with explicit principles (such as helpfulness, honesty, and harmlessness) is crucial for ensuring safe and reliable AI systems. However, standard reward-based alignment methods typically collapse diverse feedback into a single scalar reward, entangling multiple objectives into one opaque training signal, which hinders interpretability. In this work, we introduce QA-LIGN, an automatic symbolic reward decomposition approach that preserves the structure of each constitutional principle within the reward mechanism. Instead of training a black-box reward model that outputs a monolithic score, QA-LIGN formulates principle-specific evaluation questions and derives separate reward components for each principle, making it a drop-in reward model replacement. Experiments aligning an uncensored large language model with a set of constitutional principles demonstrate that QA-LIGN offers greater transparency and adaptability in the alignment process. At the same time, our approach achieves performance on par with or better than a DPO baseline. Overall, these results represent a step toward more interpretable and controllable alignment of language models, achieved without sacrificing end-task performance.
Authors: Zefang Liu, Yinzhu Quan
Abstract: We introduce EconWebArena, a benchmark for evaluating autonomous agents on complex, multimodal economic tasks in realistic web environments. The benchmark comprises 360 curated tasks from 82 authoritative websites spanning domains such as macroeconomics, labor, finance, trade, and public policy. Each task challenges agents to navigate live websites, interpret structured and visual content, interact with real interfaces, and extract precise, time-sensitive data through multi-step workflows. We construct the benchmark by prompting multiple large language models (LLMs) to generate candidate tasks, followed by rigorous human curation to ensure clarity, feasibility, and source reliability. Unlike prior work, EconWebArena emphasizes fidelity to authoritative data sources and the need for grounded web-based economic reasoning. We evaluate a diverse set of state-of-the-art multimodal LLMs as web agents, analyze failure cases, and conduct ablation studies to assess the impact of visual grounding, plan-based reasoning, and interaction design. Our results reveal substantial performance gaps and highlight persistent challenges in grounding, navigation, and multimodal understanding, positioning EconWebArena as a rigorous testbed for economic web intelligence.
Authors: Muhammad Usman, Muhammad Ahmad, M. Shahiki Tash, Irina Gelbukh, Rolando Quintero Tellez, Grigori Sidorov
Abstract: Social media platforms are critical spaces for public discourse, shaping opinions and community dynamics, yet their widespread use has amplified harmful content, particularly hate speech, threatening online safety and inclusivity. While hate speech detection has been extensively studied in languages like English and Spanish, Urdu remains underexplored, especially using translation-based approaches. To address this gap, we introduce a trilingual dataset of 10,193 tweets in English (3,834 samples), Urdu (3,197 samples), and Spanish (3,162 samples), collected via keyword filtering, with a balanced distribution of 4,849 Hateful and 5,344 Not-Hateful labels. Our methodology leverages attention layers as a precursor to transformer-based models and large language models (LLMs), enhancing feature extraction for multilingual hate speech detection. For non-transformer models, we use TF-IDF for feature extraction. The dataset is benchmarked using state-of-the-art models, including GPT-3.5 Turbo and Qwen 2.5 72B, alongside traditional machine learning models like SVM and other transformers (e.g., BERT, RoBERTa). Three annotators, following rigorous guidelines, ensured high dataset quality, achieving a Fleiss' Kappa of 0.821. Our approach, integrating attention layers with GPT-3.5 Turbo and Qwen 2.5 72B, achieves strong performance, with macro F1 scores of 0.87 for English (GPT-3.5 Turbo), 0.85 for Spanish (GPT-3.5 Turbo), 0.81 for Urdu (Qwen 2.5 72B), and 0.88 for the joint multilingual model (Qwen 2.5 72B). These results reflect improvements of 8.75% in English (over SVM baseline 0.80), 8.97% in Spanish (over SVM baseline 0.78), 5.19% in Urdu (over SVM baseline 0.77), and 7.32% in the joint multilingual model (over SVM baseline 0.82). Our framework offers a robust solution for multilingual hate speech detection, fostering safer digital communities worldwide.
Authors: Lijing Zhu, Qizhen Lan, Qing Tian, Wenbo Sun, Li Yang, Lu Xia, Yixin Xie, Xi Xiao, Tiehang Duan, Cui Tao, Shuteng Niu
Abstract: Continual Knowledge Graph Embedding (CKGE) seeks to integrate new knowledge while preserving past information. However, existing methods struggle with efficiency and scalability due to two key limitations: (1) suboptimal knowledge preservation between snapshots caused by manually designed node/relation importance scores that ignore graph dependencies relevant to the downstream task, and (2) computationally expensive graph traversal for node/relation importance calculation, leading to slow training and high memory overhead. To address these limitations, we introduce ETT-CKGE (Efficient, Task-driven, Tokens for Continual Knowledge Graph Embedding), a novel task-guided CKGE method that leverages efficient task-driven tokens for efficient and effective knowledge transfer between snapshots. Our method introduces a set of learnable tokens that directly capture task-relevant signals, eliminating the need for explicit node scoring or traversal. These tokens serve as consistent and reusable guidance across snapshots, enabling efficient token-masked embedding alignment between snapshots. Importantly, knowledge transfer is achieved through simple matrix operations, significantly reducing training time and memory usage. Extensive experiments across six benchmark datasets demonstrate that ETT-CKGE consistently achieves superior or competitive predictive performance, while substantially improving training efficiency and scalability compared to state-of-the-art CKGE methods. The code is available at: https://github.com/lijingzhu1/ETT-CKGE/tree/main
Authors: Gerardo Aleman Manzanarez, Nora de la Cruz Arana, Jorge Garcia Flores, Yobany Garcia Medina, Raul Monroy, Nathalie Pernelle
Abstract: Automated story writing has been a subject of study for over 60 years. Large language models can generate narratively consistent and linguistically coherent short fiction texts. Despite these advancements, rigorous assessment of such outputs for literary merit - especially concerning aesthetic qualities - has received scant attention. In this paper, we address the challenge of evaluating AI-generated microfictions and argue that this task requires consideration of literary criteria across various aspects of the text, such as thematic coherence, textual clarity, interpretive depth, and aesthetic quality. To facilitate this, we present GrAImes: an evaluation protocol grounded in literary theory, specifically drawing from a literary perspective, to offer an objective framework for assessing AI-generated microfiction. Furthermore, we report the results of our validation of the evaluation protocol, as answered by both literature experts and literary enthusiasts. This protocol will serve as a foundation for evaluating automatically generated microfictions and assessing their literary value.
Authors: Li Weigang, Pedro Carvalho Brom
Abstract: The rapid expansion of English technical terminology presents a significant challenge to traditional expert-based standardization, particularly in rapidly developing areas such as artificial intelligence and quantum computing. Manual approaches face difficulties in maintaining consistent multilingual terminology. To address this, we introduce LLM-BT, a back-translation framework powered by large language models (LLMs) designed to automate terminology verification and standardization through cross-lingual semantic alignment. Our key contributions include: (1) term-level consistency validation: by performing English -> intermediate language -> English back-translation, LLM-BT achieves high term consistency across different models (such as GPT-4, DeepSeek, and Grok). Case studies demonstrate over 90 percent of terms are preserved either exactly or semantically; (2) multi-path verification workflow: we develop a novel pipeline described as Retrieve -> Generate -> Verify -> Optimize, which supports both serial paths (e.g., English -> Simplified Chinese -> Traditional Chinese -> English) and parallel paths (e.g., English -> Chinese / Portuguese -> English). BLEU scores and term-level accuracy indicate strong cross-lingual robustness, with BLEU scores exceeding 0.45 and Portuguese term accuracy reaching 100 percent; (3) back-translation as semantic embedding: we reinterpret back-translation as a form of dynamic semantic embedding that uncovers latent trajectories of meaning. In contrast to static embeddings, LLM-BT offers transparent, path-based embeddings shaped by the evolution of the models. This reframing positions back-translation as an active mechanism for multilingual terminology standardization, fostering collaboration between machines and humans - machines preserve semantic integrity, while humans provide cultural interpretation.
Authors: Chupei Wang (University of Virginia), Jiaqiu Vince Sun (New York University)
Abstract: Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly positioned just before the query, LLM retrieval accuracy declines log-linearly toward zero as interference accumulates; errors arise from retrieving previously overwritten values. Attempts to mitigate interference via prompt engineering (e.g., instructing models to ignore earlier input) yield limited success. These findings reveal a fundamental constraint on LLMs' ability to disentangle interference and flexibly manipulate information, suggesting a working memory bottleneck beyond mere context access. This calls for approaches that strengthen models' ability to suppress irrelevant content during retrieval.
Authors: Samra Zafar, Shaheer Minhas, Syed Ali Hassan Zaidi, Arfa Naeem, Zahra Ali
Abstract: Large language models(LLMs) like Gemini are becoming common tools for supporting student writing. But most of their feedback is based only on the final essay missing important context about how that text was written. In this paper, we explore whether using writing process data, collected through keystroke logging and periodic snapshots, can help LLMs give feedback that better reflects how learners think and revise while writing. We built a digital writing tool that captures both what students type and how their essays evolve over time. Twenty students used this tool to write timed essays, which were then evaluated in two ways: (i) LLM generated feedback using both the final essay and the full writing trace, and (ii) After the task, students completed surveys about how useful and relatable they found the feedback. Early results show that learners preferred the process-aware LLM feedback, finding it more in tune with their own thinking. We also found that certain types of edits, like adding new content or reorganizing paragraphs, aligned closely with higher scores in areas like coherence and elaboration. Our findings suggest that making LLMs more aware of the writing process can lead to feedback that feels more meaningful, personal, and supportive.
Authors: Yu-Ang Lee, Guan-Ting Yi, Mei-Yi Liu, Jui-Chao Lu, Guan-Bo Yang, Yun-Nung Chen
Abstract: Recent advancements in large language models (LLMs) and AI systems have led to a paradigm shift in the design and optimization of complex AI workflows. By integrating multiple components, compound AI systems have become increasingly adept at performing sophisticated tasks. However, as these systems grow in complexity, new challenges arise in optimizing not only individual components but also their interactions. While traditional optimization methods such as supervised fine-tuning (SFT) and reinforcement learning (RL) remain foundational, the rise of natural language feedback introduces promising new approaches, especially for optimizing non-differentiable systems. This paper provides a systematic review of recent progress in optimizing compound AI systems, encompassing both numerical and language-based techniques. We formalize the notion of compound AI system optimization, classify existing methods along several key dimensions, and highlight open research challenges and future directions in this rapidly evolving field. A list of surveyed papers is publicly available at https://github.com/MiuLab/AISysOpt-Survey.
Authors: Shashidhar Reddy Javaji, Yupeng Cao, Haohang Li, Yangyang Yu, Nikhil Muralidhar, Zining Zhu
Abstract: Large language models (LLMs) are increasingly being used for complex research tasks such as literature review, idea generation, and scientific paper analysis, yet their ability to truly understand and process the intricate relationships within complex research papers, such as the logical links between claims and supporting evidence remains largely unexplored. In this study, we present CLAIM-BENCH, a comprehensive benchmark for evaluating LLMs' capabilities in scientific claim-evidence extraction and validation, a task that reflects deeper comprehension of scientific argumentation. We systematically compare three approaches which are inspired by divide and conquer approaches, across six diverse LLMs, highlighting model-specific strengths and weaknesses in scientific comprehension. Through evaluation involving over 300 claim-evidence pairs across multiple research domains, we reveal significant limitations in LLMs' ability to process complex scientific content. Our results demonstrate that closed-source models like GPT-4 and Claude consistently outperform open-source counterparts in precision and recall across claim-evidence identification tasks. Furthermore, strategically designed three-pass and one-by-one prompting approaches significantly improve LLMs' abilities to accurately link dispersed evidence with claims, although this comes at increased computational cost. CLAIM-BENCH sets a new standard for evaluating scientific comprehension in LLMs, offering both a diagnostic tool and a path forward for building systems capable of deeper, more reliable reasoning across full-length papers.
Authors: Wanjing Anya Ma, Michael Flor, Zuowei Wang
Abstract: Inference making is an essential but complex skill in reading comprehension (RC). Some inferences require resolving references across sentences, and some rely on using prior knowledge to fill in the detail that is not explicitly written in the text. Diagnostic RC questions can help educators provide more effective and targeted reading instruction and interventions for school-age students. We introduce a taxonomy of inference types for RC and use it to analyze the distribution of items within a diagnostic RC item bank. Next, we present experiments using GPT-4o to generate bridging-inference RC items for given reading passages via few-shot prompting, comparing conditions with and without chain-of-thought prompts. Generated items were evaluated on three aspects: overall item quality, appropriate inference type, and LLM reasoning, achieving high inter-rater agreements above 0.90. Our results show that GPT-4o produced 93.8% good-quality questions suitable for operational use in grade 3-12 contexts; however, only 42.6% of the generated questions accurately matched the targeted inference type. We conclude that combining automatic item generation with human judgment offers a promising path toward scalable, high-quality diagnostic RC assessments.
Authors: Matteo Cargnelutti, Catherine Brobston, John Hess, Jack Cushman, Kristi Mukk, Aristana Scourtas, Kyle Courtney, Greg Leppert, Amanda Watson, Martha Whitehead, Jonathan Zittrain
Abstract: Large language models (LLMs) use data to learn about the world in order to produce meaningful correlations and predictions. As such, the nature, scale, quality, and diversity of the datasets used to train these models, or to support their work at inference time, have a direct impact on their quality. The rapid development and adoption of LLMs of varying quality has brought into focus the scarcity of publicly available, high-quality training data and revealed an urgent need to ground the stewardship of these datasets in sustainable practices with clear provenance chains. To that end, this technical report introduces Institutional Books 1.0, a large collection of public domain books originally digitized through Harvard Library's participation in the Google Books project, beginning in 2006. Working with Harvard Library, we extracted, analyzed, and processed these volumes into an extensively-documented dataset of historic texts. This analysis covers the entirety of Harvard Library's collection scanned as part of that project, originally spanning 1,075,899 volumes written in over 250 different languages for a total of approximately 250 billion tokens. As part of this initial release, the OCR-extracted text (original and post-processed) as well as the metadata (bibliographic, source, and generated) of the 983,004 volumes, or 242B tokens, identified as being in the public domain have been made available. This report describes this project's goals and methods as well as the results of the analyses we performed, all in service of making this historical collection more accessible and easier for humans and machines alike to filter, read and use.
Authors: Chenlong Wang, Yuanning Feng, Dongping Chen, Zhaoyang Chu, Ranjay Krishna, Tianyi Zhou
Abstract: Recent advances in large reasoning models have enabled complex, step-by-step reasoning but often introduce significant overthinking, resulting in verbose and redundant outputs that hinder efficiency. In this study, we examine whether explicit self-reflection, signaled by tokens such as "Wait" and "Hmm", is necessary for advanced reasoning. We propose NoWait, a simple yet effective approach that disables explicit self-reflection by suppressing these tokens during inference. Extensive experiments on ten benchmarks across textual, visual, and video reasoning tasks show that NoWait reduces chain-of-thought trajectory length by up to 27%-51% in five R1-style model series, without compromising model utility. NoWait thus offers a plug-and-play solution for efficient and utility-preserving multimodal reasoning.
Authors: Yuxuan Zhou, Xien Liu, Chenwei Yan, Chen Ning, Xiao Zhang, Boxun Li, Xiangling Fu, Shijin Wang, Guoping Hu, Yu Wang, Ji Wu
Abstract: Large language models (LLMs) have demonstrated remarkable performance on various medical benchmarks, but their capabilities across different cognitive levels remain underexplored. Inspired by Bloom's Taxonomy, we propose a multi-cognitive-level evaluation framework for assessing LLMs in the medical domain in this study. The framework integrates existing medical datasets and introduces tasks targeting three cognitive levels: preliminary knowledge grasp, comprehensive knowledge application, and scenario-based problem solving. Using this framework, we systematically evaluate state-of-the-art general and medical LLMs from six prominent families: Llama, Qwen, Gemma, Phi, GPT, and DeepSeek. Our findings reveal a significant performance decline as cognitive complexity increases across evaluated models, with model size playing a more critical role in performance at higher cognitive levels. Our study highlights the need to enhance LLMs' medical capabilities at higher cognitive levels and provides insights for developing LLMs suited to real-world medical applications.
Authors: Yiqun Sun, Qiang Huang, Anthony K. H. Tung, Jun Yu
Abstract: This position paper argues that the text embedding research community should move beyond surface meaning and embrace implicit semantics as a central modeling goal. Text embedding models have become foundational in modern NLP, powering a wide range of applications and drawing increasing research attention. Yet, much of this progress remains narrowly focused on surface-level semantics. In contrast, linguistic theory emphasizes that meaning is often implicit, shaped by pragmatics, speaker intent, and sociocultural context. Current embedding models are typically trained on data that lacks such depth and evaluated on benchmarks that reward the capture of surface meaning. As a result, they struggle with tasks requiring interpretive reasoning, speaker stance, or social meaning. Our pilot study highlights this gap, showing that even state-of-the-art models perform only marginally better than simplistic baselines on implicit semantics tasks. To address this, we call for a paradigm shift: embedding research should prioritize more diverse and linguistically grounded training data, design benchmarks that evaluate deeper semantic understanding, and explicitly frame implicit meaning as a core modeling objective, better aligning embeddings with real-world language complexity.
Authors: Li-Ming Zhan, Bo Liu, Zexin Lu, Chengqiang Xie, Jiannong Cao, Xiao-Ming Wu
Abstract: Inference-time steering aims to alter the response characteristics of large language models (LLMs) without modifying their underlying parameters. A critical step in this process is the identification of internal modules within LLMs that are associated with the target behavior. However, current approaches to module selection often depend on superficial cues or ad-hoc heuristics, which can result in suboptimal or unintended outcomes. In this work, we propose a principled causal-attribution framework for identifying behavior-relevant attention heads in transformers. For each head, we train a vector-quantized autoencoder (VQ-AE) on its attention activations, partitioning the latent space into behavior-relevant and behavior-irrelevant subspaces, each quantized with a shared learnable codebook. We assess the behavioral relevance of each head by quantifying the separability of VQ-AE encodings for behavior-aligned versus behavior-violating responses using a binary classification metric. This yields a behavioral relevance score that reflects each head discriminative capacity with respect to the target behavior, guiding both selection and importance weighting. Experiments on seven LLMs from two model families and five behavioral steering datasets demonstrate that our method enables more accurate inference-time interventions, achieving superior performance on the truthfulness-steering task. Furthermore, the heads selected by our approach exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.
Authors: Jash Rajesh Parekh, Pengcheng Jiang, Jiawei Han
Abstract: Understanding cause and effect relationships remains a formidable challenge for Large Language Models (LLMs), particularly in specialized domains where reasoning requires more than surface-level correlations. Retrieval-Augmented Generation (RAG) improves factual accuracy, but standard RAG pipelines treat evidence as flat context, lacking the structure required to model true causal dependencies. We introduce Causal-Chain RAG (CC-RAG), a novel approach that integrates zero-shot triple extraction and theme-aware graph chaining into the RAG pipeline, enabling structured multi-hop inference. Given a domain specific corpus, CC-RAG constructs a Directed Acyclic Graph (DAG) of
Authors: Zikai Xiao, Ziyang Wang, Wen Ma, Yan Zhang, Wei Shen, Yan Wang, Luqi Gong, Zuozhu Liu
Abstract: While Large Language Models (LLMs) support long contexts, they struggle with performance degradation within the context window. Current solutions incur prohibitive training costs, leaving statistical behaviors and cost-effective approaches underexplored. From the decoding perspective, we identify the Posterior Salience Attenuation (PSA) phenomenon, where the salience ratio correlates with long-text performance degradation. Notably, despite the attenuation, gold tokens still occupy high-ranking positions in the decoding space. Motivated by it, we propose the training-free Positional Contrastive Decoding (PCD) that contrasts the logits derived from long-aware attention with those from designed local-aware attention, enabling the model to focus on the gains introduced by large-scale short-to-long training. Through the analysis of long-term decay simulation, we demonstrate that PCD effectively alleviates attention score degradation. Experimental results show that PCD achieves state-of-the-art performance on long-context benchmarks.
Authors: Kevin Galim, Ethan Ewer, Wonjun Kang, Minjae Lee, Hyung Il Koo, Kangwook Lee
Abstract: Optimizing inference for long-context Large Language Models (LLMs) is increasingly important due to the quadratic compute and linear memory complexity of Transformers. Existing approximation methods, such as key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on rough predictions of token or KV pair importance. We propose a novel framework for approximate LLM inference that leverages small draft models to more accurately predict the importance of tokens and KV pairs. Specifically, we introduce two instantiations of our proposed framework: (i) SpecKV, which leverages a draft output to accurately assess the importance of each KV pair for more effective KV cache dropping, and (ii) SpecPC, which uses the draft model's attention activations to identify and discard unimportant prompt tokens. To the best of our knowledge, this is the first work to use draft models for approximate LLM inference acceleration, extending their utility beyond traditional lossless speculative decoding. We motivate our methods with theoretical and empirical analyses, and show a strong correlation between the attention patterns of draft and target models. Extensive experiments on long-context benchmarks show that our methods consistently achieve higher accuracy than existing baselines, while preserving the same improvements in memory usage, latency, and throughput. Our code is available at https://github.com/furiosa-ai/draft-based-approx-llm.
Authors: Tao Zou, Xinghua Zhang, Haiyang Yu, Minzheng Wang, Fei Huang, Yongbin Li
Abstract: With the development and widespread application of large language models (LLMs), the new paradigm of "Model as Product" is rapidly evolving, and demands higher capabilities to address complex user needs, often requiring precise workflow execution which involves the accurate understanding of multiple tasks. However, existing benchmarks focusing on single-task environments with limited constraints lack the complexity required to fully reflect real-world scenarios. To bridge this gap, we present the Extremely Complex Instruction Following Benchmark (EIFBENCH), meticulously crafted to facilitate a more realistic and robust evaluation of LLMs. EIFBENCH not only includes multi-task scenarios that enable comprehensive assessment across diverse task types concurrently, but also integrates a variety of constraints, replicating complex operational environments. Furthermore, we propose the Segment Policy Optimization (SegPO) algorithm to enhance the LLM's ability to accurately fulfill multi-task workflow. Evaluations on EIFBENCH have unveiled considerable performance discrepancies in existing LLMs when challenged with these extremely complex instructions. This finding underscores the necessity for ongoing optimization to navigate the intricate challenges posed by LLM applications.
Authors: Luel Hagos Beyene, Vivek Verma, Min Ma, Jesujoba O. Alabi, Fabian David Schmidt, Joyce Nakatumba-Nabende, David Ifeoluwa Adelani
Abstract: Large Language models (LLMs) have demonstrated impressive performance on a wide range of tasks, including in multimodal settings such as speech. However, their evaluation is often limited to English and a few high-resource languages. For low-resource languages, there is no standardized evaluation benchmark. In this paper, we address this gap by introducing mSTEB, a new benchmark to evaluate the performance of LLMs on a wide range of tasks covering language identification, text classification, question answering, and translation tasks on both speech and text modalities. We evaluated the performance of leading LLMs such as Gemini 2.0 Flash and GPT-4o (Audio) and state-of-the-art open models such as Qwen 2 Audio and Gemma 3 27B. Our evaluation shows a wide gap in performance between high-resource and low-resource languages, especially for languages spoken in Africa and Americas/Oceania. Our findings show that more investment is needed to address their under-representation in LLMs coverage.
Authors: Weiya Li, Junjie Chen, Bei Li, Boyang Liu, Zichen Wen, Nuanqiao Shan, Xiaoqian Liu, Anping Liu, Huajie Liu, Hu Song, Linfeng Zhang
Abstract: Machine translation has long been a central task in natural language processing. With the rapid advancement of large language models (LLMs), there has been remarkable progress in translation quality. However, fully realizing the translation potential of LLMs remains an open challenge. Recent studies have explored multi-agent systems to decompose complex translation tasks into collaborative subtasks, showing initial promise in enhancing translation quality through agent cooperation and specialization. Nevertheless, existing multi-agent translation frameworks largely neglect foundational insights from cognitive translation studies. These insights emphasize how human translators employ different cognitive strategies, such as balancing literal and free translation, refining expressions based on context, and iteratively evaluating outputs. To address this limitation, we propose a cognitively informed multi-agent framework called TACTIC, which stands for T ranslation A gents with Cognitive- T heoretic Interactive Collaboration. The framework comprises six functionally distinct agents that mirror key cognitive processes observed in human translation behavior. These include agents for drafting, refinement, evaluation, scoring, context reasoning, and external knowledge gathering. By simulating an interactive and theory-grounded translation workflow, TACTIC effectively leverages the full capacity of LLMs for high-quality translation. Experimental results on diverse language pairs from the FLORES-200 and WMT24 benchmarks show that our method consistently achieves state-of-the-art performance. Using DeepSeek-V3 as the base model, TACTIC surpasses GPT-4.1 by an average of +0.6 XCOMET and +1.18 COMETKIWI-23. Compared to DeepSeek-R1, it further improves by +0.84 XCOMET and +2.99 COMETKIWI-23. Code is available at https://github.com/weiyali126/TACTIC.
Authors: Ziyang Ma, Qingyue Yuan, Zhenglin Wang, Deyu Zhou
Abstract: Previous research has primarily focused on the cognitive error detection capabilities of Large Language Models (LLMs), often prompting them to analyze mistakes in reasoning chains. However, few studies have examined the meta-cognitive abilities of LLMs (e.g., their self-awareness of step errors), which are crucial for their reliability. While studies on LLM self-evaluation present some measures, such as perplexity, which can reflect the answer correctness and be viewed as the lens of meta-cognition, they lack step-level analysis and adaptation. This paper studies the evaluation of LLM meta-cognition using the current lenses and how to improve these lenses. Specifically, we propose AutoMeco, an Automated Meta-cognition Evaluation framework for benchmarking the existing lenses. Furthermore, a training-free Markovian Intrinsic Reward Adjustment strategy, MIRA, is proposed to boost current meta-cognition lenses. Experimental results on three mathematical reasoning datasets and three LLMs show the reasonableness of AutoMeco by comparing it with Best-of-N verification. Moreover, the meta-cognition ability of LLMs can be better evaluated using MIRA.
Authors: Jiaxiang Liu, Boxuan Xing, Chenhao Yuan, Chenxiang Zhang, Di Wu, Xiusheng Huang, Haida Yu, Chuhan Lang, Pengfei Cao, Jun Zhao, Kang Liu
Abstract: As large language models (LLMs) continue to advance, there is a growing urgency to enhance the interpretability of their internal knowledge mechanisms. Consequently, many interpretation methods have emerged, aiming to unravel the knowledge mechanisms of LLMs from various perspectives. However, current interpretation methods differ in input data formats and interpreting outputs. The tools integrating these methods are only capable of supporting tasks with specific inputs, significantly constraining their practical applications. To address these challenges, we present an open-source Knowledge Mechanisms Revealer&Interpreter (Know-MRI) designed to analyze the knowledge mechanisms within LLMs systematically. Specifically, we have developed an extensible core module that can automatically match different input data with interpretation methods and consolidate the interpreting outputs. It enables users to freely choose appropriate interpretation methods based on the inputs, making it easier to comprehensively diagnose the model's internal knowledge mechanisms from multiple perspectives. Our code is available at https://github.com/nlpkeg/Know-MRI. We also provide a demonstration video on https://youtu.be/NVWZABJ43Bs.
URLs: https://github.com/nlpkeg/Know-MRI., https://youtu.be/NVWZABJ43Bs.
Authors: Ziqi. Liu, Ziyang. Zhou, Mingxuan. Hu
Abstract: Large language model (LLM) have become mainstream methods in the field of sarcasm detection. However, existing LLM methods face challenges in irony detection, including: 1. single-perspective limitations, 2. insufficient comprehensive understanding, and 3. lack of interpretability. This paper introduces the Collaborative Agent Framework for Irony (CAF-I), an LLM-driven multi-agent system designed to overcome these issues. CAF-I employs specialized agents for Context, Semantics, and Rhetoric, which perform multidimensional analysis and engage in interactive collaborative optimization. A Decision Agent then consolidates these perspectives, with a Refinement Evaluator Agent providing conditional feedback for optimization. Experiments on benchmark datasets establish CAF-I's state-of-the-art zero-shot performance. Achieving SOTA on the vast majority of metrics, CAF-I reaches an average Macro-F1 of 76.31, a 4.98 absolute improvement over the strongest prior baseline. This success is attained by its effective simulation of human-like multi-perspective analysis, enhancing detection accuracy and interpretability.
Authors: Hern\'an Maina, Nicol\'as Wolovick, Luciana Benotti
Abstract: Training Large Language Models (LLMs) is costly in terms of energy, hardware, and annotated data, often resulting in a positionality rooted in predominant cultures and values (Santy et al., 2023). Domain adaptation has emerged as a promising strategy to better align models with diverse cultural and value contexts (Hershcovich et al., 2022), but its computational cost remains a significant barrier, particularly for research groups lacking access to large-scale infrastructure. In this paper, we evaluate how the use of different numerical precision formats and data parallelization strategies impacts both training speed (as a proxy to energy and hardware consumption) and model accuracy, with the goal of facilitating domain adaptation in low-resource environments. Our findings are relevant to any setting where energy efficiency, accessibility, or limited hardware availability are key concerns.
Authors: Jiujun He, Huazhen Lin
Abstract: Most existing structured pruning methods for Large Language Models (LLMs) require substantial computational and data resources for retraining to reestablish the corrupted correlations, making them prohibitively expensive. To address this, we propose a pruning framework for LLMs called Orthogonal decomposition and Linear Calibration (Olica), which eliminates the need for retraining. A key observation is that the multi-head attention (MHA) layer depends on two types of matrix products. By treating these matrix products as unified entities and applying principal component analysis (PCA), we extract the most important information to compress LLMs without sacrificing accuracy or disrupting their original structure. Consequently, retraining becomes unnecessary. A fast decomposition method is devised, reducing the complexity of PCA by a factor of the square of the number of attention heads. Additionally, to mitigate error accumulation problem caused by pruning the feed-forward network (FFN) layer, we introduce a linear calibration method to reconstruct the residual errors of pruned layers using low-rank matrices. By leveraging singular value decomposition (SVD) on the solution of the least-squares problem, these matrices are obtained without requiring retraining. Extensive experiments show that the proposed Olica is efficient in terms of data usage, GPU memory, and running time, while delivering superior performance across multiple benchmarks.
Authors: Fengjun Pan, Anh Tuan Luu, Xiaobao Wu
Abstract: Detecting harmful memes is essential for maintaining the integrity of online environments. However, current approaches often struggle with resource efficiency, flexibility, or explainability, limiting their practical deployment in content moderation systems. To address these challenges, we introduce U-CoT+, a novel framework for harmful meme detection. Instead of relying solely on prompting or fine-tuning multimodal models, we first develop a high-fidelity meme-to-text pipeline that converts visual memes into detail-preserving textual descriptions. This design decouples meme interpretation from meme classification, thus avoiding immediate reasoning over complex raw visual content and enabling resource-efficient harmful meme detection with general large language models (LLMs). Building on these textual descriptions, we further incorporate targeted, interpretable human-crafted guidelines to guide models' reasoning under zero-shot CoT prompting. As such, this framework allows for easy adaptation to different harmfulness detection criteria across platforms, regions, and over time, offering high flexibility and explainability. Extensive experiments on seven benchmark datasets validate the effectiveness of our framework, highlighting its potential for explainable and low-resource harmful meme detection using small-scale LLMs. Codes and data are available at: https://anonymous.4open.science/r/HMC-AF2B/README.md.
Authors: Chihiro Taguchi, Seiji Maekawa, Nikita Bhutani
Abstract: Retrieval-augmented generation (RAG) and long-context language models (LCLMs) both address context limitations of LLMs in open-domain question answering (QA). However, optimal external context to retrieve remains an open problem: fixing the retrieval size risks either wasting tokens or omitting key evidence. Existing adaptive methods like Self-RAG and Self-Route rely on iterative LLM prompting and perform well on factoid QA, but struggle with aggregation QA, where the optimal context size is both unknown and variable. We present Adaptive-$k$ retrieval, a simple and effective single-pass method that adaptively selects the number of passages based on the distribution of the similarity scores between the query and the candidate passages. It does not require model fine-tuning, extra LLM inferences or changes to existing retriever-reader pipelines. On both factoid and aggregation QA benchmarks, Adaptive-$k$ matches or outperforms fixed-$k$ baselines while using up to 10x fewer tokens than full-context input, yet still retrieves 70% of relevant passages. It improves accuracy across five LCLMs and two embedding models, highlighting that dynamically adjusting context size leads to more efficient and accurate QA.
Authors: Huixuan Zhang, Xiaojun Wan
Abstract: Text-to-image models often struggle to generate images that precisely match textual prompts. Prior research has extensively studied the evaluation of image-text alignment in text-to-image generation. However, existing evaluations primarily focus on agreement with human assessments, neglecting other critical properties of a trustworthy evaluation framework. In this work, we first identify two key aspects that a reliable evaluation should address. We then empirically demonstrate that current mainstream evaluation frameworks fail to fully satisfy these properties across a diverse range of metrics and models. Finally, we propose recommendations for improving image-text alignment evaluation.
Authors: Sumanth Manduru, Carlotta Domeniconi
Abstract: The rapid adoption of Small Language Models (SLMs) for on-device and resource-constrained deployments has outpaced our understanding of their ethical risks. To the best of our knowledge, we present the first large-scale audit of instruction-tuned SLMs spanning 0.5 to 5 billion parameters-an overlooked "middle tier" between BERT-class encoders and flagship LLMs. Our evaluation includes nine open-source models from the Qwen 2.5, LLaMA 3.2, Gemma 3, and Phi families. Using the BBQ benchmark under zero-shot prompting, we analyze both utility and fairness across ambiguous and disambiguated contexts. This evaluation reveals three key insights. First, competence and fairness need not be antagonistic: Phi models achieve F1 scores exceeding 90 percent while exhibiting minimal bias, showing that efficient and ethical NLP is attainable. Second, social bias varies significantly by architecture: Qwen 2.5 models may appear fair, but this often reflects vacuous neutrality, random guessing, or evasive behavior rather than genuine ethical alignment. In contrast, LLaMA 3.2 models exhibit stronger stereotypical bias, suggesting overconfidence rather than neutrality. Third, compression introduces nuanced trade-offs: 4-bit AWQ quantization improves F1 scores in ambiguous settings for LLaMA 3.2-3B but increases disability-related bias in Phi-4-Mini by over 7 percentage points. These insights provide practical guidance for the responsible deployment of SLMs in applications demanding fairness and efficiency, particularly benefiting small enterprises and resource-constrained environments.
Authors: Ashutosh Dwivedi, Siddhant Shivdutt Singh, Ashutosh Modi
Abstract: In recent years, researchers have started analyzing the cultural sensitivity of LLMs. In this respect, Etiquettes have been an active area of research. Etiquettes are region-specific and are an essential part of the culture of a region; hence, it is imperative to make LLMs sensitive to etiquettes. However, there needs to be more resources in evaluating LLMs for their understanding and bias with regard to etiquettes. In this resource paper, we introduce EtiCor++, a corpus of etiquettes worldwide. We introduce different tasks for evaluating LLMs for knowledge about etiquettes across various regions. Further, we introduce various metrics for measuring bias in LLMs. Extensive experimentation with LLMs shows inherent bias towards certain regions.
Authors: Xiao Wei, Xiaobao Wang, Ning Zhuang, Chenyang Wang, Longbiao Wang, Jianwu dang
Abstract: Intent detection aims to identify user intents from natural language inputs, where supervised methods rely heavily on labeled in-domain (IND) data and struggle with out-of-domain (OOD) intents, limiting their practical applicability. Generalized Intent Discovery (GID) addresses this by leveraging unlabeled OOD data to discover new intents without additional annotation. However, existing methods focus solely on clustering unsupervised data while neglecting domain adaptation. Therefore, we propose a consistency-driven prototype-prompting framework for GID from the perspective of integrating old and new knowledge, which includes a prototype-prompting framework for transferring old knowledge from external sources, and a hierarchical consistency constraint for learning new knowledge from target domains. We conducted extensive experiments and the results show that our method significantly outperforms all baseline methods, achieving state-of-the-art results, which strongly demonstrates the effectiveness and generalization of our methods. Our source code is publicly available at https://github.com/smileix/cpp.
Authors: Arie Cattan, Alon Jacovi, Ori Ram, Jonathan Herzig, Roee Aharoni, Sasha Goldshtein, Eran Ofek, Idan Szpektor, Avi Caciularu
Abstract: Retrieval Augmented Generation (RAG) is a commonly used approach for enhancing large language models (LLMs) with relevant and up-to-date information. However, the retrieved sources can often contain conflicting information and it remains unclear how models should address such discrepancies. In this work, we first propose a novel taxonomy of knowledge conflict types in RAG, along with the desired model behavior for each type. We then introduce CONFLICTS, a high-quality benchmark with expert annotations of conflict types in a realistic RAG setting. CONFLICTS is the first benchmark that enables tracking progress on how models address a wide range of knowledge conflicts. We conduct extensive experiments on this benchmark, showing that LLMs often struggle to appropriately resolve conflicts between sources. While prompting LLMs to explicitly reason about the potential conflict in the retrieved documents significantly improves the quality and appropriateness of their responses, substantial room for improvement in future research remains.
Authors: Divyaksh Shukla, Ritesh Baviskar, Dwijesh Gohil, Aniket Tiwari, Atul Shree, Ashutosh Modi
Abstract: Discourse parsing is an important task useful for NLU applications such as summarization, machine comprehension, and emotion recognition. The current discourse parsing datasets based on conversations consists of written English dialogues restricted to a single domain. In this resource paper, we introduce CoMuMDR: Code-mixed Multi-modal Multi-domain corpus for Discourse paRsing in conversations. The corpus (code-mixed in Hindi and English) has both audio and transcribed text and is annotated with nine discourse relations. We experiment with various SoTA baseline models; the poor performance of SoTA models highlights the challenges of multi-domain code-mixed corpus, pointing towards the need for developing better models for such realistic settings.
Authors: Xinyuan Wang, Dongjie Wang, Wangyang Ying, Haoyue Bai, Nanxu Gong, Sixun Dong, Kunpeng Liu, Yanjie Fu
Abstract: Reasoning is a key component of language understanding in Large Language Models. While Chain-of-Thought prompting enhances performance via explicit intermediate steps, it suffers from sufficient token overhead and a fixed reasoning trajectory, preventing step-wise refinement. Recent advances in latent reasoning address these limitations by refining internal reasoning processes directly in the model's latent space, without producing explicit outputs. However, a key challenge remains: how to effectively update reasoning embeddings during post-training to guide the model toward more accurate solutions. To overcome this challenge, we propose a lightweight post-training framework that refines latent reasoning trajectories using two novel strategies: 1) Contrastive reasoning feedback, which compares reasoning embeddings against strong and weak baselines to infer effective update directions via embedding enhancement; 2) Residual embedding refinement, which stabilizes updates by progressively integrating current and historical gradients, enabling fast yet controlled convergence. Extensive experiments and case studies are conducted on five reasoning benchmarks to demonstrate the effectiveness of the proposed framework. Notably, a 5\% accuracy gain on MathQA without additional training.
Authors: Tuukka T\"or\"o, Antti Suni, Juraj \v{S}imko
Abstract: Investigating linguistic relationships on a global scale requires analyzing diverse features such as syntax, phonology and prosody, which evolve at varying rates influenced by internal diversification, language contact, and sociolinguistic factors. Recent advances in machine learning (ML) offer complementary alternatives to traditional historical and typological approaches. Instead of relying on expert labor in analyzing specific linguistic features, these new methods enable the exploration of linguistic variation through embeddings derived directly from speech, opening new avenues for large-scale, data-driven analyses. This study employs embeddings from the fine-tuned XLS-R self-supervised language identification model voxlingua107-xls-r-300m-wav2vec, to analyze relationships between 106 world languages based on speech recordings. Using linear discriminant analysis (LDA), language embeddings are clustered and compared with genealogical, lexical, and geographical distances. The results demonstrate that embedding-based distances align closely with traditional measures, effectively capturing both global and local typological patterns. Challenges in visualizing relationships, particularly with hierarchical clustering and network-based methods, highlight the dynamic nature of language change. The findings show potential for scalable analyses of language variation based on speech embeddings, providing new perspectives on relationships among languages. By addressing methodological considerations such as corpus size and latent space dimensionality, this approach opens avenues for studying low-resource languages and bridging macro- and micro-level linguistic variation. Future work aims to extend these methods to underrepresented languages and integrate sociolinguistic variation for a more comprehensive understanding of linguistic diversity.
Authors: Yahan Li, Jifan Yao, John Bosco S. Bunyi, Adam C. Frank, Angel Hwang, Ruishan Liu
Abstract: Large language models (LLMs) are increasingly proposed for use in mental health support, yet their behavior in realistic counseling scenarios remains largely untested. We introduce CounselBench, a large-scale benchmark developed with 100 mental health professionals to evaluate and stress-test LLMs in single-turn counseling. The first component, CounselBench-EVAL, contains 2,000 expert evaluations of responses from GPT-4, LLaMA 3, Gemini, and online human therapists to real patient questions. Each response is rated along six clinically grounded dimensions, with written rationales and span-level annotations. We find that LLMs often outperform online human therapists in perceived quality, but experts frequently flag their outputs for safety concerns such as unauthorized medical advice. Follow-up experiments show that LLM judges consistently overrate model responses and overlook safety issues identified by human experts. To probe failure modes more directly, we construct CounselBench-Adv, an adversarial dataset of 120 expert-authored counseling questions designed to trigger specific model issues. Evaluation across 2,880 responses from eight LLMs reveals consistent, model-specific failure patterns. Together, CounselBench establishes a clinically grounded framework for benchmarking and improving LLM behavior in high-stakes mental health settings.
Authors: Liyan Xu, Zhenlin Su, Mo Yu, Jiangnan Li, Fandong Meng, Jie Zhou
Abstract: This work focuses on an observed limitation of text encoders: embeddings may not be able to recognize fine-grained entities or events within the semantics, resulting in failed dense retrieval on even simple cases. To examine such behaviors, we first introduce a new evaluation dataset in Chinese, named CapRetrieval, whose passages are image captions, and queries are phrases inquiring entities or events in various forms. Zero-shot evaluation suggests that encoders may fail on these fine-grained matching, regardless of training sources or model sizes. Aiming for enhancement, we proceed to finetune encoders with our proposed data generation strategies, which obtains the best performance on CapRetrieval. Within this process, we further identify an issue of granularity dilemma, a challenge for embeddings to express fine-grained salience while aligning with overall semantics. Our dataset, code and models in this work are publicly released at https://github.com/lxucs/CapRetrieval.
Authors: Shuzhou Yuan, Ercong Nie, Mario Tawfelis, Helmut Schmid, Hinrich Sch\"utze, Michael F\"arber
Abstract: Hate speech detection is a socially sensitive and inherently subjective task, with judgments often varying based on personal traits. While prior work has examined how socio-demographic factors influence annotation, the impact of personality traits on Large Language Models (LLMs) remains largely unexplored. In this paper, we present the first comprehensive study on the role of persona prompts in hate speech classification, focusing on MBTI-based traits. A human annotation survey confirms that MBTI dimensions significantly affect labeling behavior. Extending this to LLMs, we prompt four open-source models with MBTI personas and evaluate their outputs across three hate speech datasets. Our analysis uncovers substantial persona-driven variation, including inconsistencies with ground truth, inter-persona disagreement, and logit-level biases. These findings highlight the need to carefully define persona prompts in LLM-based annotation workflows, with implications for fairness and alignment with human values.
Authors: Minhae Oh, Jeonghye Kim, Nakyung Lee, Donggeon Seo, Taeuk Kim, Jungwoo Lee
Abstract: Scientific reasoning requires not only long-chain reasoning processes, but also knowledge of domain-specific terminologies and adaptation to updated findings. To deal with these challenges for scientific reasoning, we introduce RAISE, a step-by-step retrieval-augmented framework which retrieves logically relevant documents from in-the-wild corpus. RAISE is divided into three steps: problem decomposition, logical query generation, and logical retrieval. We observe that RAISE consistently outperforms other baselines on scientific reasoning benchmarks. We analyze that unlike other baselines, RAISE retrieves documents that are not only similar in terms of the domain knowledge, but also documents logically more relevant.
Authors: Son The Nguyen, Theja Tulabandhula
Abstract: Large language models (LLMs) are increasingly used for both open-ended and structured tasks, yet their inference-time behavior is still largely dictated by heuristic decoding strategies such as greedy search, sampling, or reranking. These methods provide limited control and do not explicitly optimize for task-specific objectives. We introduce MEMETRON, a task-agnostic framework that formulates LLM decoding as a discrete black-box optimization problem. MEMETRON leverages hybrid metaheuristic algorithms, GENETRON and ANNETRON, to search the response space, guided by reward models and contextual operations performed by the LLM itself. This approach enables efficient discovery of high-reward responses without requiring model retraining or gradient access. The framework is modular and generalizes across diverse tasks, requiring only a reward function and lightweight prompt templates. We evaluate our framework on the critical human preference alignment task and demonstrate that it significantly outperforms standard decoding and reranking methods, highlighting its potential to improve alignment without model retraining.
Authors: Mingyu Zheng, Zhifan Feng, Jia Wang, Lanrui Wang, Zheng Lin, Yang Hao, Weiping Wang
Abstract: Despite the commendable progress of recent LLM-based data synthesis methods, they face two limitations in generating table instruction tuning data. First, they can not thoroughly explore the vast input space of table understanding tasks, leading to limited data diversity. Second, they ignore the weaknesses in table understanding ability of the target LLM and blindly pursue the increase of data quantity, resulting in suboptimal data efficiency. In this paper, we introduce a progressive and weakness-guided data synthesis framework tailored for table instruction tuning, named TableDreamer, to mitigate the above issues. Specifically, we first synthesize diverse tables and related instructions as seed data, and then perform an iterative exploration of the input space under the guidance of the newly identified weakness data, which eventually serve as the final training data for fine-tuning the target LLM. Extensive experiments on 10 tabular benchmarks demonstrate the effectiveness of the proposed framework, which boosts the average accuracy of Llama3.1-8B-instruct by 11.62% (49.07% to 60.69%) with 27K GPT-4o synthetic data and outperforms state-of-the-art data synthesis baselines which use more training data. The code and data is available at https://github.com/SpursGoZmy/TableDreamer
Authors: Oumaima El Khettari, Solen Quiniou, Samuel Chaffron
Abstract: We explore a generative relation extraction (RE) pipeline tailored to the study of interactions in the intestinal microbiome, a complex and low-resource biomedical domain. Our method leverages summarization with large language models (LLMs) to refine context before extracting relations via instruction-tuned generation. Preliminary results on a dedicated corpus show that summarization improves generative RE performance by reducing noise and guiding the model. However, BERT-based RE approaches still outperform generative models. This ongoing work demonstrates the potential of generative methods to support the study of specialized domains in low-resources setting.
Authors: Yang Liu, Jiaqi Li, Zilong Zheng
Abstract: Rule-based reasoning has been acknowledged as one of the fundamental problems in reasoning, while deviations in rule formats, types, and complexity in real-world applications pose severe challenges. Recent studies have shown that large reasoning models (LRMs) have remarkable reasoning capabilities, and their performance is substantially enhanced by reinforcement learning (RL). However, it remains an open question whether small reasoning models (SRMs) can learn rule-based reasoning effectively with robust generalization across diverse tasks and domains. To address this, we introduce Reinforced Rule-based Reasoning, a.k.a. RuleReasoner, a simple yet effective method to conduct rule-based reasoning via a wide collection of curated tasks and a novel domain-aware dynamic sampling approach. Specifically, RuleReasoner resamples each training batch by updating the sampling weights of different domains based on historical rewards. This facilitates domain augmentation and flexible online learning schedules for RL, obviating the need for pre-hoc human-engineered mix-training recipes used in existing methods. Empirical evaluations on in-distribution (ID) and out-of-distribution (OOD) benchmarks reveal that RuleReasoner outperforms frontier LRMs by a significant margin ($\Delta$4.1% average points on eight ID tasks and $\Delta$10.4% average points on three OOD tasks over OpenAI-o1). Notably, our approach also exhibits higher computational efficiency compared to prior dynamic sampling methods for RL.
Authors: Soham Poddar, Paramita Koley, Janardan Misra, Sanjay Podder, Navveen Balani, Niloy Ganguly, Saptarshi Ghosh
Abstract: A significant portion of the energy consumed by Large Language Models (LLMs) arises from their inference processes; hence developing energy-efficient methods for inference is crucial. While several techniques exist for inference optimization, output compression remains relatively unexplored, with only a few preliminary efforts addressing this aspect. In this work, we first benchmark 12 decoder-only LLMs across 5 datasets, revealing that these models often produce responses that are substantially longer than necessary. We then conduct a comprehensive quality assessment of LLM responses, formally defining six information categories present in LLM responses. We show that LLMs often tend to include redundant or additional information besides the minimal answer. To address this issue of long responses by LLMs, we explore several simple and intuitive prompt-engineering strategies. Empirical evaluation shows that appropriate prompts targeting length reduction and controlling information content can achieve significant energy optimization between 25-60\% by reducing the response length while preserving the quality of LLM responses.
Authors: Ruiran Su, Jiasheng Si, Zhijiang Guo, Janet B. Pierrehumbert
Abstract: Scientific fact-checking has mostly focused on text and tables, overlooking scientific charts, which are key for presenting quantitative evidence and statistical reasoning. We introduce ClimateViz, the first large-scale benchmark for scientific fact-checking using expert-curated scientific charts. ClimateViz contains 49,862 claims linked to 2,896 visualizations, each labeled as support, refute, or not enough information. To improve interpretability, each example includes structured knowledge graph explanations covering trends, comparisons, and causal relations. We evaluate state-of-the-art multimodal language models, including both proprietary and open-source systems, in zero-shot and few-shot settings. Results show that current models struggle with chart-based reasoning: even the best systems, such as Gemini 2.5 and InternVL 2.5, reach only 76.2 to 77.8 percent accuracy in label-only settings, far below human performance (89.3 and 92.7 percent). Explanation-augmented outputs improve performance in some models. We released our dataset and code alongside the paper.
Authors: Hee Suk Yoon, Eunseop Yoon, Mark A. Hasegawa-Johnson, Sungwoong Kim, Chang D. Yoo
Abstract: We introduce ConfPO, a method for preference learning in Large Language Models (LLMs) that identifies and optimizes preference-critical tokens based solely on the training policy's confidence, without requiring any auxiliary models or compute. Unlike prior Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO), which uniformly adjust all token probabilities regardless of their relevance to preference, ConfPO focuses optimization on the most impactful tokens. This targeted approach improves alignment quality while mitigating overoptimization (i.e., reward hacking) by using the KL divergence budget more efficiently. In contrast to recent token-level methods that rely on credit-assignment models or AI annotators, raising concerns about scalability and reliability, ConfPO is simple, lightweight, and model-free. Experimental results on challenging alignment benchmarks, including AlpacaEval 2 and Arena-Hard, demonstrate that ConfPO consistently outperforms uniform DAAs across various LLMs, delivering better alignment with zero additional computational overhead.
Authors: Fariz Ikhwantri, Dusica Marijan
Abstract: Ensuring complex systems meet regulations typically requires checking the validity of assurance cases through a claim-argument-evidence framework. Some challenges in this process include the complicated nature of legal and technical texts, the need for model explanations, and limited access to assurance case data. We propose a compliance detection approach based on Natural Language Inference (NLI): EXplainable CompLiance detection with Argumentative Inference of Multi-hop reasoning (EXCLAIM). We formulate the claim-argument-evidence structure of an assurance case as a multi-hop inference for explainable and traceable compliance detection. We address the limited number of assurance cases by generating them using large language models (LLMs). We introduce metrics that measure the coverage and structural consistency. We demonstrate the effectiveness of the generated assurance case from GDPR requirements in a multi-hop inference task as a case study. Our results highlight the potential of NLI-based approaches in automating the regulatory compliance process.
Authors: Mehedi Hasan Bijoy, Dejan Porjazovski, Tam\'as Gr\'osz, Mikko Kurimo
Abstract: Speech Emotion Recognition (SER) is crucial for improving human-computer interaction. Despite strides in monolingual SER, extending them to build a multilingual system remains challenging. Our goal is to train a single model capable of multilingual SER by distilling knowledge from multiple teacher models. To address this, we introduce a novel language-aware multi-teacher knowledge distillation method to advance SER in English, Finnish, and French. It leverages Wav2Vec2.0 as the foundation of monolingual teacher models and then distills their knowledge into a single multilingual student model. The student model demonstrates state-of-the-art performance, with a weighted recall of 72.9 on the English dataset and an unweighted recall of 63.4 on the Finnish dataset, surpassing fine-tuning and knowledge distillation baselines. Our method excels in improving recall for sad and neutral emotions, although it still faces challenges in recognizing anger and happiness.
Authors: Nelvin Tan, Zian Seng, Liang Zhang, Yu-Ching Shih, Dong Yang, Amol Salunkhe
Abstract: Large language models (LLMs) have shown impressive capabilities on numerous natural language processing tasks. However, LLMs still struggle with numerical question answering for financial documents that include tabular and textual data. Recent works have showed the effectiveness of critic agents (i.e., self-correction) for this task given oracle labels. Building upon this framework, this paper examines the effectiveness of the traditional critic agent when oracle labels are not available, and show, through experiments, that this critic agent's performance deteriorates in this scenario. With this in mind, we present an improved critic agent, along with the calculator agent which outperforms the previous state-of-the-art approach (program-of-thought) and is safer. Furthermore, we investigate how our agents interact with each other, and how this interaction affects their performance.
Authors: Dror Kris Markus, Fabrizio Gilardi, Daria Stetsenko
Abstract: As artificial intelligence (AI) systems become deeply embedded in everyday life, calls to align AI development with ethical and societal values have intensified. Interdisciplinary collaboration is often championed as a key pathway for fostering such engagement. Yet it remains unclear whether interdisciplinary research teams are actually leading this shift in practice. This study analyzes over 100,000 AI-related papers published on ArXiv between 2014 and 2024 to examine how ethical values and societal concerns are integrated into technical AI research. We develop a classifier to identify societal content and measure the extent to which research papers express these considerations. We find a striking shift: while interdisciplinary teams remain more likely to produce societally-oriented research, computer science-only teams now account for a growing share of the field's overall societal output. These teams are increasingly integrating societal concerns into their papers and tackling a wide range of domains - from fairness and safety to healthcare and misinformation. These findings challenge common assumptions about the drivers of societal AI and raise important questions. First, what are the implications for emerging understandings of AI safety and governance if most societally-oriented research is being undertaken by exclusively technical teams? Second, for scholars in the social sciences and humanities: in a technical field increasingly responsive to societal demands, what distinctive perspectives can we still offer to help shape the future of AI?
Authors: Muhammad Anwar, Mishca de Costa, Issam Hammad, Daniel Lau
Abstract: This paper introduces a domain-specific Large Language Model for nuclear applications, built from the publicly accessible Essential CANDU textbook. Drawing on a compact Transformer-based architecture, the model is trained on a single GPU to protect the sensitive data inherent in nuclear operations. Despite relying on a relatively small dataset, it shows encouraging signs of capturing specialized nuclear vocabulary, though the generated text sometimes lacks syntactic coherence. By focusing exclusively on nuclear content, this approach demonstrates the feasibility of in-house LLM solutions that align with rigorous cybersecurity and data confidentiality standards. Early successes in text generation underscore the model's utility for specialized tasks, while also revealing the need for richer corpora, more sophisticated preprocessing, and instruction fine-tuning to enhance domain accuracy. Future directions include extending the dataset to cover diverse nuclear subtopics, refining tokenization to reduce noise, and systematically evaluating the model's readiness for real-world applications in nuclear domain.
Authors: Muhammad Anwar, Daniel Lau, Mishca de Costa, Issam Hammad
Abstract: The nuclear industry possesses a wealth of valuable information locked away in unstructured text data. This data, however, is not readily usable for advanced Large Language Model (LLM) applications that require clean, structured question-answer pairs for tasks like model training, fine-tuning, and evaluation. This paper explores how synthetic data generation can bridge this gap, enabling the development of robust LLMs for the nuclear domain. We discuss the challenges of data scarcity and privacy concerns inherent in the nuclear industry and how synthetic data provides a solution by transforming existing text data into usable Q&A pairs. This approach leverages LLMs to analyze text, extract key information, generate relevant questions, and evaluate the quality of the resulting synthetic dataset. By unlocking the potential of LLMs in the nuclear industry, synthetic data can pave the way for improved information retrieval, enhanced knowledge sharing, and more informed decision-making in this critical sector.
Authors: Pradyoth Hegde, Santosh Kesiraju, Jan \v{S}vec, \v{S}imon Sedl\'a\v{c}ek, Bolaji Yusuf, Old\v{r}ich Plchot, Deepak K T, Jan \v{C}ernock\'y
Abstract: This study explores the application of in-context learning (ICL) to the dialogue state tracking (DST) problem and investigates the factors that influence its effectiveness. We use a sentence embedding based k-nearest neighbour method to retrieve the suitable demonstrations for ICL. The selected demonstrations, along with the test samples, are structured within a template as input to the LLM. We then conduct a systematic study to analyse the impact of factors related to demonstration selection and prompt context on DST performance. This work is conducted using the MultiWoZ2.4 dataset and focuses primarily on the OLMo-7B-instruct, Mistral-7B-Instruct-v0.3, and Llama3.2-3B-Instruct models. Our findings provide several useful insights on in-context learning abilities of LLMs for dialogue state tracking.
Authors: Mishca de Costa, Muhammad Anwar, Dave Mercier, Mark Randall, Issam Hammad
Abstract: Retrieving operational data from nuclear power plants requires exceptional accuracy and transparency due to the criticality of the decisions it supports. Traditionally, natural language to SQL (NL-to-SQL) approaches have been explored for querying such data. While NL-to-SQL promises ease of use, it poses significant risks: end-users cannot easily validate generated SQL queries, and legacy nuclear plant databases -- often complex and poorly structured -- complicate query generation due to decades of incremental modifications. These challenges increase the likelihood of inaccuracies and reduce trust in the approach. In this work, we propose an alternative paradigm: leveraging function-calling large language models (LLMs) to address these challenges. Instead of directly generating SQL queries, we define a set of pre-approved, purpose-specific functions representing common use cases. Queries are processed by invoking these functions, which encapsulate validated SQL logic. This hybrid approach mitigates the risks associated with direct NL-to-SQL translations by ensuring that SQL queries are reviewed and optimized by experts before deployment. While this strategy introduces the upfront cost of developing and maintaining the function library, we demonstrate how NL-to-SQL tools can assist in the initial generation of function code, allowing experts to focus on validation rather than creation. Our study includes a performance comparison between direct NL-to-SQL generation and the proposed function-based approach, highlighting improvements in accuracy and maintainability. This work underscores the importance of balancing user accessibility with operational safety and provides a novel, actionable framework for robust data retrieval in critical systems.
Authors: Ahmed Hasanaath, Aisha Alansari, Ahmed Ashraf, Chafik Salmane, Hamzah Luqman, Saad Ezzini
Abstract: Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks-boosting sentiment analysis from 35.3% to 87.5% and paraphrase detection from 56.1% to 87.0%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://anonymous.4open.science/r/AraReasoner41299
Authors: Francisco Vargas, Alejandro Gonz\'alez Coene, Gaston Escalante, Exequiel Lob\'on, Manuel Pulido
Abstract: The extraction of information about traffic accidents from legal documents is crucial for quantifying insurance company costs. Extracting entities such as percentages of physical and/or psychological disability and the involved compensation amounts is a challenging process, even for experts, due to the subtle arguments and reasoning in the court decision. A two-step procedure is proposed: first, segmenting the document identifying the most relevant segments, and then extracting the entities. For text segmentation, two methodologies are compared: a classic method based on regular expressions and a second approach that divides the document into blocks of n-tokens, which are then vectorized using multilingual models for semantic searches (text-embedding-ada-002/MiniLM-L12-v2 ). Subsequently, large language models (LLaMA-2 7b, 70b, LLaMA-3 8b, and GPT-4 Turbo) are applied with prompting to the selected segments for entity extraction. For the LLaMA models, fine-tuning is performed using LoRA. LLaMA-2 7b, even with zero temperature, shows a significant number of hallucinations in extractions which are an important contention point for named entity extraction. This work shows that these hallucinations are substantially reduced after finetuning the model. The performance of the methodology based on segment vectorization and subsequent use of LLMs significantly surpasses the classic method which achieves an accuracy of 39.5%. Among open-source models, LLaMA-2 70B with finetuning achieves the highest accuracy 79.4%, surpassing its base version 61.7%. Notably, the base LLaMA-3 8B model already performs comparably to the finetuned LLaMA-2 70B model, achieving 76.6%, highlighting the rapid progress in model development. Meanwhile, GPT-4 Turbo achieves the highest accuracy at 86.1%.
Authors: Flavio D'Intino, Hans-Peter Hutter
Abstract: Swiss German is a low-resource language represented by diverse dialects that differ significantly from Standard German and from each other, lacking a standardized written form. As a result, transcribing Swiss German involves translating into Standard German. Existing datasets have been collected in controlled environments, yielding effective speech-to-text (STT) models, but these models struggle with spontaneous conversational speech. This paper, therefore, introduces the new SRB-300 dataset, a 300-hour annotated speech corpus featuring real-world long-audio recordings from 39 Swiss German radio and TV stations. It captures spontaneous speech across all major Swiss dialects recorded in various realistic environments and overcomes the limitation of prior sentence-level corpora. We fine-tuned multiple OpenAI Whisper models on the SRB-300 dataset, achieving notable enhancements over previous zero-shot performance metrics. Improvements in word error rate (WER) ranged from 19% to 33%, while BLEU scores increased between 8% and 40%. The best fine-tuned model, large-v3, achieved a WER of 17.1% and a BLEU score of 74.8. This advancement is crucial for developing effective and robust STT systems for Swiss German and other low-resource languages in real-world contexts.
Authors: Danush Khanna, Krishna Kumar, Basab Ghosh, Vinija Jain, Vasu Sharma, Aman Chadha, Amitava Das
Abstract: Adversarial threats against LLMs are escalating faster than current defenses can adapt. We expose a critical geometric blind spot in alignment: adversarial prompts exploit latent camouflage, embedding perilously close to the safe representation manifold while encoding unsafe intent thereby evading surface level defenses like Direct Preference Optimization (DPO), which remain blind to the latent geometry. We introduce ALKALI, the first rigorously curated adversarial benchmark and the most comprehensive to date spanning 9,000 prompts across three macro categories, six subtypes, and fifteen attack families. Evaluation of 21 leading LLMs reveals alarmingly high Attack Success Rates (ASRs) across both open and closed source models, exposing an underlying vulnerability we term latent camouflage, a structural blind spot where adversarial completions mimic the latent geometry of safe ones. To mitigate this vulnerability, we introduce GRACE - Geometric Representation Aware Contrastive Enhancement, an alignment framework coupling preference learning with latent space regularization. GRACE enforces two constraints: latent separation between safe and adversarial completions, and adversarial cohesion among unsafe and jailbreak behaviors. These operate over layerwise pooled embeddings guided by a learned attention profile, reshaping internal geometry without modifying the base model, and achieve up to 39% ASR reduction. Moreover, we introduce AVQI, a geometry aware metric that quantifies latent alignment failure via cluster separation and compactness. AVQI reveals when unsafe completions mimic the geometry of safe ones, offering a principled lens into how models internally encode safety. We make the code publicly available at https://anonymous.4open.science/r/alkali-B416/README.md.
URLs: https://anonymous.4open.science/r/alkali-B416/README.md.
Authors: Hiba Khey, Amine Lakhder, Salma Rouichi, Imane El Ghabi, Kamal Hejjaoui, Younes En-nahli, Fahd Kalloubi, Moez Amri
Abstract: The rapid advancement of transformer-based language models has catalyzed breakthroughs in biomedical and clinical natural language processing; however, plant science remains markedly underserved by such domain-adapted tools. In this work, we present PlantBert, a high-performance, open-source language model specifically tailored for extracting structured knowledge from plant stress-response literature. Built upon the DeBERTa architecture-known for its disentangled attention and robust contextual encoding-PlantBert is fine-tuned on a meticulously curated corpus of expert-annotated abstracts, with a primary focus on lentil (Lens culinaris) responses to diverse abiotic and biotic stressors. Our methodology combines transformer-based modeling with rule-enhanced linguistic post-processing and ontology-grounded entity normalization, enabling PlantBert to capture biologically meaningful relationships with precision and semantic fidelity. The underlying corpus is annotated using a hierarchical schema aligned with the Crop Ontology, encompassing molecular, physiological, biochemical, and agronomic dimensions of plant adaptation. PlantBert exhibits strong generalization capabilities across entity types and demonstrates the feasibility of robust domain adaptation in low-resource scientific fields. By providing a scalable and reproducible framework for high-resolution entity recognition, PlantBert bridges a critical gap in agricultural NLP and paves the way for intelligent, data-driven systems in plant genomics, phenomics, and agronomic knowledge discovery. Our model is publicly released to promote transparency and accelerate cross-disciplinary innovation in computational plant science.
Authors: Elias Horner, Cristinel Mateis, Guido Governatori, Agata Ciabattoni
Abstract: We present a novel approach to the automated semantic analysis of legal texts using large language models (LLMs), targeting their transformation into formal representations in Defeasible Deontic Logic (DDL). We propose a structured pipeline that segments complex normative language into atomic snippets, extracts deontic rules, and evaluates them for syntactic and semantic coherence. Our methodology is evaluated across various LLM configurations, including prompt engineering strategies, fine-tuned models, and multi-stage pipelines, focusing on legal norms from the Australian Telecommunications Consumer Protections Code. Empirical results demonstrate promising alignment between machine-generated and expert-crafted formalizations, showing that LLMs - particularly when prompted effectively - can significantly contribute to scalable legal informatics.
Authors: Antonios Dimakis (Archimedes, Athena Research Center, Greece, Department of Informatics and Telecommunications, NKUA), John Pavlopoulos (Archimedes, Athena Research Center, Greece, Department of Informatics, Athens University of Economics and Business, Greece), Antonios Anastasopoulos (Archimedes, Athena Research Center, Greece, Department of Computer Science, George Mason University)
Abstract: Natural language understanding systems struggle with low-resource languages, including many dialects of high-resource ones. Dialect-to-standard normalization attempts to tackle this issue by transforming dialectal text so that it can be used by standard-language tools downstream. In this study, we tackle this task by introducing a new normalization method that combines rule-based linguistically informed transformations and large language models (LLMs) with targeted few-shot prompting, without requiring any parallel data. We implement our method for Greek dialects and apply it on a dataset of regional proverbs, evaluating the outputs using human annotators. We then use this dataset to conduct downstream experiments, finding that previous results regarding these proverbs relied solely on superficial linguistic information, including orthographic artifacts, while new observations can still be made through the remaining semantics.
Authors: Zeyu Leo Liu, Greg Durrett, Eunsol Choi
Abstract: Knowledge editing techniques for large language models (LLMs) can inject knowledge that is later reproducible verbatim, but they fall short on propagating that knowledge: models cannot answer questions that require reasoning with the injected knowledge. We present a hypernetwork-based approach for knowledge propagation, named PropMEND, where we meta-learn how to modify gradients of a language modeling loss to encourage injected information to propagate. Our approach extends the meta-objective of MEND [29] so that gradient updates on knowledge are transformed to enable answering multi-hop questions involving that knowledge. We show improved performance on the RippleEdit dataset, showing almost 2x accuracy on challenging multi-hop questions whose answers are not explicitly stated in the injected fact. We further introduce a new dataset, Controlled RippleEdit, to evaluate the generalization of our hypernetwork, testing knowledge propagation along relations and entities unseen during hypernetwork training. PropMEND still outperforms existing approaches in unseen entity-relation pairs, yet the performance gap decreases substantially, suggesting future work in propagating knowledge to a wide range of relations.
Authors: Andrew Shin
Abstract: While large language models (LLMs) have achieved remarkable performance in various tasks including mathematical reasoning, their development typically demands prohibitive computational resources. Recent advancements have reduced costs for training capable models, yet even these approaches rely on high-end hardware clusters. In this paper, we demonstrate that a single average gaming GPU can train a solid mathematical reasoning model, by integrating reinforcement learning and memory optimization techniques. Specifically, we train a 1.5B parameter mathematical reasoning model on RTX 3080 Ti of 16GB memory that achieves comparable or better performance on mathematical reasoning benchmarks than models several times larger, in resource-constrained environments. Our results challenge the paradigm that state-of-the-art mathematical reasoning necessitates massive infrastructure, democratizing access to high-performance AI research. https://github.com/shinandrew/YouronMath.
Authors: Qinggang Zhang, Zhishang Xiang, Yilin Xiao, Le Wang, Junhui Li, Xinrun Wang, Jinsong Su
Abstract: Large language models (LLMs) augmented with retrieval systems have demonstrated significant potential in handling knowledge-intensive tasks. However, these models often struggle with unfaithfulness issues, generating outputs that either ignore the retrieved context or inconsistently blend it with the LLM`s parametric knowledge. This issue is particularly severe in cases of knowledge conflict, where the retrieved context conflicts with the model`s parametric knowledge. While existing faithful RAG approaches enforce strict context adherence through well-designed prompts or modified decoding strategies, our analysis reveals a critical limitation: they achieve faithfulness by forcibly suppressing the model`s parametric knowledge, which undermines the model`s internal knowledge structure and increases the risk of misinterpreting the context. To this end, this paper proposes FaithfulRAG, a novel framework that resolves knowledge conflicts by explicitly modeling discrepancies between the model`s parametric knowledge and retrieved context. Specifically, FaithfulRAG identifies conflicting knowledge at the fact level and designs a self-thinking process, allowing LLMs to reason about and integrate conflicting facts before generating responses. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. The code is available at https:// github.com/DeepLearnXMU/Faithful-RAG
Authors: Clara Lachenmaier, Judith Sieker, Sina Zarrie{\ss}
Abstract: Communication among humans relies on conversational grounding, allowing interlocutors to reach mutual understanding even when they do not have perfect knowledge and must resolve discrepancies in each other's beliefs. This paper investigates how large language models (LLMs) manage common ground in cases where they (don't) possess knowledge, focusing on facts in the political domain where the risk of misinformation and grounding failure is high. We examine the ability of LLMs to answer direct knowledge questions and loaded questions that presuppose misinformation. We evaluate whether loaded questions lead LLMs to engage in active grounding and correct false user beliefs, in connection to their level of knowledge and their political bias. Our findings highlight significant challenges in LLMs' ability to engage in grounding and reject false user beliefs, raising concerns about their role in mitigating misinformation in political discourse.
Authors: Marek Kadl\v{c}\'ik, Michal \v{S}tef\'anik, Timothee Mickus, Michal Spiegel, Josef Kucha\v{r}
Abstract: Pretrained language models (LMs) are prone to arithmetic errors. Existing work showed limited success in probing numeric values from models' representations, indicating that these errors can be attributed to the inherent unreliability of distributionally learned embeddings in representing exact quantities. However, we observe that previous probing methods are inadequate for the emergent structure of learned number embeddings with sinusoidal patterns. In response, we propose a novel probing technique that decodes numeric values from input embeddings with near-perfect accuracy across a range of open-source LMs. This proves that after the sole pre-training, LMs represent numbers with remarkable precision. Finally, we find that the embeddings' preciseness judged by our probe's accuracy explains a large portion of LM's errors in elementary arithmetic, and show that aligning the embeddings with the pattern discovered by our probe can mitigate these errors.
Authors: Yuan Guo, Tingjia Miao, Zheng Wu, Pengzhou Cheng, Ming Zhou, Zhuosheng Zhang
Abstract: Autonomous agents powered by multimodal large language models have been developed to facilitate task execution on mobile devices. However, prior work has predominantly focused on atomic tasks -- such as shot-chain execution tasks and single-screen grounding tasks -- while overlooking the generalization to compositional tasks, which are indispensable for real-world applications. This work introduces UI-NEXUS, a comprehensive benchmark designed to evaluate mobile agents on three categories of compositional operations: Simple Concatenation, Context Transition, and Deep Dive. UI-NEXUS supports interactive evaluation in 20 fully controllable local utility app environments, as well as 30 online Chinese and English service apps. It comprises 100 interactive task templates with an average optimal step count of 14.05. Experimental results across a range of mobile agents with agentic workflow or agent-as-a-model show that UI-NEXUS presents significant challenges. Specifically, existing agents generally struggle to balance performance and efficiency, exhibiting representative failure modes such as under-execution, over-execution, and attention drift, causing visible atomic-to-compositional generalization gap. Inspired by these findings, we propose AGENT-NEXUS, a lightweight and efficient scheduling system to tackle compositional mobile tasks. AGENT-NEXUS extrapolates the abilities of existing mobile agents by dynamically decomposing long-horizon tasks to a series of self-contained atomic subtasks. AGENT-NEXUS achieves 24% to 40% task success rate improvement for existing mobile agents on compositional operation tasks within the UI-NEXUS benchmark without significantly sacrificing inference overhead. The demo video, dataset, and code are available on the project page at https://ui-nexus.github.io.
Authors: Satu Hopponen, Tomi Kinnunen, Alexandre Nikolaev, Rosa Gonz\'alez Hautam\"aki, Lauri Tavi, Einar Meister
Abstract: We introduce a new FROST-EMA (Finnish and Russian Oral Speech Dataset of Electromagnetic Articulography) corpus. It consists of 18 bilingual speakers, who produced speech in their native language (L1), second language (L2), and imitated L2 (fake foreign accent). The new corpus enables research into language variability from phonetic and technological points of view. Accordingly, we include two preliminary case studies to demonstrate both perspectives. The first case study explores the impact of L2 and imitated L2 on the performance of an automatic speaker verification system, while the second illustrates the articulatory patterns of one speaker in L1, L2, and a fake accent.
Authors: Yuejiao Wang, Xianmin Gong, Xixin Wu, Patrick Wong, Hoi-lam Helene Fung, Man Wai Mak, Helen Meng
Abstract: Early detection is crucial for timely intervention aimed at preventing and slowing the progression of neurocognitive disorder (NCD), a common and significant health problem among the aging population. Recent evidence has suggested that language-related functional magnetic resonance imaging (fMRI) may be a promising approach for detecting cognitive decline and early NCD. In this paper, we proposed a novel, naturalistic language-related fMRI task for this purpose. We examined the effectiveness of this task among 97 non-demented Chinese older adults from Hong Kong. The results showed that machine-learning classification models based on fMRI features extracted from the task and demographics (age, gender, and education year) achieved an average area under the curve of 0.86 when classifying participants' cognitive status (labeled as NORMAL vs DECLINE based on their scores on a standard neurcognitive test). Feature localization revealed that the fMRI features most frequently selected by the data-driven approach came primarily from brain regions associated with language processing, such as the superior temporal gyrus, middle temporal gyrus, and right cerebellum. The study demonstrated the potential of the naturalistic language-related fMRI task for early detection of aging-related cognitive decline and NCD.
Authors: Theo Zhang, Madurya Suresh, Anne S. Warlaumont, Kasia Hitczenko, Alejandrina Cristia, Margaret Cychosz
Abstract: Speech technology systems struggle with many downstream tasks for child speech due to small training corpora and the difficulties that child speech pose. We apply a novel dataset, SpeechMaturity, to state-of-the-art transformer models to address a fundamental classification task: identifying child vocalizations. Unlike previous corpora, our dataset captures maximally ecologically-valid child vocalizations across an unprecedented sample, comprising children acquiring 25+ languages in the U.S., Bolivia, Vanuatu, Papua New Guinea, Solomon Islands, and France. The dataset contains 242,004 labeled vocalizations, magnitudes larger than previous work. Models were trained to distinguish between cry, laughter, mature (consonant+vowel), and immature speech (just consonant or vowel). Models trained on the dataset outperform state-of-the-art models trained on previous datasets, achieved classification accuracy comparable to humans, and were robust across rural and urban settings.
Authors: Lei Zhang, Jiaxi Yang, Min Yang, Jian Yang, Mouxiang Chen, Jiajun Zhang, Zeyu Cui, Binyuan Hui, Junyang Lin
Abstract: We introduce **SWE-Flow**, a novel data synthesis framework grounded in Test-Driven Development (TDD). Unlike existing software engineering data that rely on human-submitted issues, **SWE-Flow** automatically infers incremental development steps directly from unit tests, which inherently encapsulate high-level requirements. The core of **SWE-Flow** is the construction of a Runtime Dependency Graph (RDG), which precisely captures function interactions, enabling the generation of a structured, step-by-step *development schedule*. At each step, **SWE-Flow** produces a partial codebase, the corresponding unit tests, and the necessary code modifications, resulting in fully verifiable TDD tasks. With this approach, we generated 16,061 training instances and 2,020 test instances from real-world GitHub projects, creating the **SWE-Flow-Eval** benchmark. Our experiments show that fine-tuning open model on this dataset significantly improves performance in TDD-based coding. To facilitate further research, we release all code, datasets, models, and Docker images at [Github](https://github.com/Hambaobao/SWE-Flow).
Authors: Hakyung Sung, Gyu-Ho Shin, Chanyoung Lee, You Kyung Sung, Boo Kyung Jung
Abstract: The present study extends recent work on Universal Dependencies annotations for second-language (L2) Korean by introducing a semi-automated framework that identifies morphosyntactic constructions from XPOS sequences and aligns those constructions with corresponding UPOS categories. We also broaden the existing L2-Korean corpus by annotating 2,998 new sentences from argumentative essays. To evaluate the impact of XPOS-UPOS alignments, we fine-tune L2-Korean morphosyntactic analysis models on datasets both with and without these alignments, using two NLP toolkits. Our results indicate that the aligned dataset not only improves consistency across annotation layers but also enhances morphosyntactic tagging and dependency-parsing accuracy, particularly in cases of limited annotated data.
Authors: Jianing Qi, Xi Ye, Hao Tang, Zhigang Zhu, Eunsol Choi
Abstract: Scaling test-time compute brings substantial performance gains for large language models (LLMs). By sampling multiple answers and heuristically aggregate their answers (e.g., either through majority voting or using verifiers to rank the answers), one can achieve consistent performance gains in math domains. In this paper, we propose a new way to leverage such multiple sample set. We train a compact LLM, called Sample Set Aggregator (SSA), that takes a concatenated sequence of multiple samples and output the final answer, optimizing it for the answer accuracy with reinforcement learning. Experiments on multiple reasoning datasets show that SSA outperforms other test-time scaling methods such as reward model-based re-ranking. Our approach also shows a promising generalization ability, across sample set sizes, base model families and scales, and tasks. By separating LLMs to generate answers and LLMs to analyze and aggregate sampled answers, our approach can work with the outputs from premier black box models easily and efficiently.
Authors: Hakyung Sung, Karla Csuros, Min-Chang Sung
Abstract: This study examines the lexical and syntactic interventions of human and LLM proofreading aimed at improving overall intelligibility in identical second language writings, and evaluates the consistency of outcomes across three LLMs (ChatGPT-4o, Llama3.1-8b, Deepseek-r1-8b). Findings show that both human and LLM proofreading enhance bigram lexical features, which may contribute to better coherence and contextual connectedness between adjacent words. However, LLM proofreading exhibits a more generative approach, extensively reworking vocabulary and sentence structures, such as employing more diverse and sophisticated vocabulary and incorporating a greater number of adjective modifiers in noun phrases. The proofreading outcomes are highly consistent in major lexical and syntactic features across the three models.
Authors: Haozhen Zhang, Tao Feng, Jiaxuan You
Abstract: The rapid emergence of diverse large language models (LLMs) has spurred the development of LLM routers that assign user queries to the most suitable model. However, existing LLM routers typically perform a single-round, one-to-one mapping (\textit{i.e.}, assigning each query to a single model in isolation), which limits their capability to tackle complex tasks that demand the complementary strengths of multiple LLMs. In this paper, we present \textbf{Router-R1}, a reinforcement learning (RL)-based framework that formulates multi-LLM routing and aggregation as a sequential decision process. Router-R1 instantiates the router itself as a capable LLM, leveraging its reasoning ability to interleave "think" actions (internal deliberation) with "route" actions (dynamic model invocation), and integrates each response into its evolving context. To guide learning, we employ a lightweight rule-based reward comprising format rewards, final outcome rewards, and a novel cost reward for performance and cost trade-off optimization, opening a pathway toward optimizing performance-cost tradeoffs via RL. Router-R1 also conditions only on simple model descriptors such as pricing, latency, and example performance, enabling strong generalization to unseen model selection. Experiments on seven general and multi-hop QA benchmarks show that Router-R1 outperforms over several strong baselines, achieving superior performance while maintaining robust generalization and cost management.Code is available at https://github.com/ulab-uiuc/Router-R1.
Authors: Yaniv Nikankin, Dana Arad, Yossi Gandelsman, Yonatan Belinkov
Abstract: Vision-Language models (VLMs) show impressive abilities to answer questions on visual inputs (e.g., counting objects in an image), yet demonstrate higher accuracies when performing an analogous task on text (e.g., counting words in a text). We investigate this accuracy gap by identifying and comparing the \textit{circuits} - the task-specific computational sub-graphs - in different modalities. We show that while circuits are largely disjoint between modalities, they implement relatively similar functionalities: the differences lie primarily in processing modality-specific data positions (an image or a text sequence). Zooming in on the image data representations, we observe they become aligned with the higher-performing analogous textual representations only towards later layers, too late in processing to effectively influence subsequent positions. To overcome this, we patch the representations of visual data tokens from later layers back into earlier layers. In experiments with multiple tasks and models, this simple intervention closes a third of the performance gap between the modalities, on average. Our analysis sheds light on the multi-modal performance gap in VLMs and suggests a training-free approach for reducing it.
Authors: Chenxi Liu, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, Tianyi Zhou, Heng Huang
Abstract: The task adaptation and alignment of Large Multimodal Models (LMMs) have been significantly advanced by instruction tuning and further strengthened by recent preference optimization. Yet, most LMMs still suffer from severe modality imbalance during reasoning, i.e., outweighing language prior biases over visual inputs, which bottlenecks their generalization to downstream tasks and causes hallucinations. However, existing preference optimization approaches for LMMs do not focus on restraining the internal biases of their Large Language Model (LLM) backbones when curating the training data. Moreover, they heavily rely on offline data and lack the capacity to explore diverse responses adaptive to dynamic distributional shifts during training. Meanwhile, Group Relative Policy Optimization (GRPO), a recent method using online-generated data and verified rewards to improve reasoning capabilities, remains largely underexplored in LMM alignment. In this paper, we propose a novel preference learning framework, Modality-Balancing Preference Optimization (MBPO), to address the modality imbalance in LMMs. MBPO constructs a more effective offline preference dataset by generating hard negatives, i.e., rejected responses misled by LLM biases due to limited usage of visual information, through adversarial perturbation of input images. Moreover, MBPO leverages the easy-to-verify nature of close-ended tasks to generate online responses with verified rewards. GRPO is then employed to train the model with offline-online hybrid data. Extensive experiments demonstrate that MBPO can enhance LMM performance on challenging vision-language tasks and effectively reduce hallucinations.
Authors: Abdellah Ghassel, Ian Robinson, Gabriel Tanase, Hal Cooper, Bryan Thompson, Zhen Han, Vassilis N. Ioannidis, Soji Adeshina, Huzefa Rangwala
Abstract: Retrieval-Augmented Generation (RAG) grounds large language models in external evidence, yet it still falters when answers must be pieced together across semantically distant documents. We close this gap with the Hierarchical Lexical Graph (HLG), a three-tier index that (i) traces every atomic proposition to its source, (ii) clusters propositions into latent topics, and (iii) links entities and relations to expose cross-document paths. On top of HLG we build two complementary, plug-and-play retrievers: StatementGraphRAG, which performs fine-grained entity-aware beam search over propositions for high-precision factoid questions, and TopicGraphRAG, which selects coarse topics before expanding along entity links to supply broad yet relevant context for exploratory queries. Additionally, existing benchmarks lack the complexity required to rigorously evaluate multi-hop summarization systems, often focusing on single-document queries or limited datasets. To address this, we introduce a synthetic dataset generation pipeline that curates realistic, multi-document question-answer pairs, enabling robust evaluation of multi-hop retrieval systems. Extensive experiments across five datasets demonstrate that our methods outperform naive chunk-based RAG achieving an average relative improvement of 23.1% in retrieval recall and correctness. Open-source Python library is available at https://github.com/awslabs/graphrag-toolkit.
Authors: Hanbing Liu, Lang Cao, Yuanyi Ren, Mengyu Zhou, Haoyu Dong, Xiaojun Ma, Shi Han, Dongmei Zhang
Abstract: Large language models have demonstrated impressive reasoning capabilities, yet they often suffer from inefficiencies due to unnecessarily verbose or redundant outputs. While many works have explored reinforcement learning (RL) to enhance reasoning abilities, most primarily focus on improving accuracy, with limited attention to reasoning efficiency. Some existing approaches introduce direct length-based rewards to encourage brevity, but this often leads to noticeable drops in accuracy. In this paper, we propose Bingo, an RL framework that advances length-based reward design to boost efficient reasoning. Bingo incorporates two key mechanisms: a significance-aware length reward, which gradually guides the model to reduce only insignificant tokens, and a dynamic length reward, which initially encourages elaborate reasoning for hard questions but decays over time to improve overall efficiency. Experiments across multiple reasoning benchmarks show that Bingo improves both accuracy and efficiency. It outperforms the vanilla reward and several other length-based reward baselines in RL, achieving a favorable trade-off between accuracy and efficiency. These results underscore the potential of training LLMs explicitly for efficient reasoning.
Authors: Yifei Li, Hanane Nour Moussa, Ziru Chen, Shijie Chen, Botao Yu, Mingyi Xue, Benjamin Burns, Tzu-Yao Chiu, Vishal Dey, Zitong Lu, Chen Wei, Qianheng Zhang, Tianyu Zhang, Song Gao, Xuhui Huang, Xia Ning, Nesreen K. Ahmed, Ali Payani, Huan Sun
Abstract: Despite long-standing efforts in accelerating scientific discovery with AI, building AI co-scientists remains challenging due to limited high-quality data for training and evaluation. To tackle this data scarcity issue, we present AutoSDT, an automatic pipeline that collects high-quality coding tasks in real-world data-driven discovery workflows. AutoSDT leverages the coding capabilities and parametric knowledge of LLMs to search for diverse sources, select ecologically valid tasks, and synthesize accurate task instructions and code solutions. Using our pipeline, we construct AutoSDT-5K, a dataset of 5,404 coding tasks for data-driven discovery that covers four scientific disciplines and 756 unique Python packages. To the best of our knowledge, AutoSDT-5K is the only automatically collected and the largest open dataset for data-driven scientific discovery. Expert feedback on a subset of 256 tasks shows the effectiveness of AutoSDT: 93% of the collected tasks are ecologically valid, and 92.2% of the synthesized programs are functionally correct. Trained on AutoSDT-5K, the Qwen2.5-Coder-Instruct LLM series, dubbed AutoSDT-Coder, show substantial improvement on two challenging data-driven discovery benchmarks, ScienceAgentBench and DiscoveryBench. Most notably, AutoSDT-Coder-32B reaches the same level of performance as GPT-4o on ScienceAgentBench with a success rate of 7.8%, doubling the performance of its base model. On DiscoveryBench, it lifts the hypothesis matching score to 8.1, bringing a 17.4% relative improvement and closing the gap between open-weight models and GPT-4o.
Authors: Wenlong Meng, Shuguo Fan, Chengkun Wei, Min Chen, Yuwei Li, Yuanchao Zhang, Zhikun Zhang, Wenzhi Chen
Abstract: In this paper, we introduce GradEscape, the first gradient-based evader designed to attack AI-generated text (AIGT) detectors. GradEscape overcomes the undifferentiable computation problem, caused by the discrete nature of text, by introducing a novel approach to construct weighted embeddings for the detector input. It then updates the evader model parameters using feedback from victim detectors, achieving high attack success with minimal text modification. To address the issue of tokenizer mismatch between the evader and the detector, we introduce a warm-started evader method, enabling GradEscape to adapt to detectors across any language model architecture. Moreover, we employ novel tokenizer inference and model extraction techniques, facilitating effective evasion even in query-only access. We evaluate GradEscape on four datasets and three widely-used language models, benchmarking it against four state-of-the-art AIGT evaders. Experimental results demonstrate that GradEscape outperforms existing evaders in various scenarios, including with an 11B paraphrase model, while utilizing only 139M parameters. We have successfully applied GradEscape to two real-world commercial AIGT detectors. Our analysis reveals that the primary vulnerability stems from disparity in text expression styles within the training data. We also propose a potential defense strategy to mitigate the threat of AIGT evaders. We open-source our GradEscape for developing more robust AIGT detectors.
Authors: Andrew Z. Wang, Songwei Ge, Tero Karras, Ming-Yu Liu, Yogesh Balaji
Abstract: Both text-to-image generation and large language models (LLMs) have made significant advancements. However, many text-to-image models still employ the somewhat outdated T5 and CLIP as their text encoders. In this work, we investigate the effectiveness of using modern decoder-only LLMs as text encoders for text-to-image diffusion models. We build a standardized training and evaluation pipeline that allows us to isolate and evaluate the effect of different text embeddings. We train a total of 27 text-to-image models with 12 different text encoders to analyze the critical aspects of LLMs that could impact text-to-image generation, including the approaches to extract embeddings, different LLMs variants, and model sizes. Our experiments reveal that the de facto way of using last-layer embeddings as conditioning leads to inferior performance. Instead, we explore embeddings from various layers and find that using layer-normalized averaging across all layers significantly improves alignment with complex prompts. Most LLMs with this conditioning outperform the baseline T5 model, showing enhanced performance in advanced visio-linguistic reasoning skills.
Authors: Ken Gu, Zhihan Zhang, Kate Lin, Yuwei Zhang, Akshay Paruchuri, Hong Yu, Mehran Kazemi, Kumar Ayush, A. Ali Heydari, Maxwell A. Xu, Girish Narayanswamy, Yun Liu, Ming-Zher Poh, Yuzhe Yang, Mark Malhotra, Shwetak Patel, Hamid Palangi, Xuhai Xu, Daniel McDuff, Tim Althoff, Xin Liu
Abstract: Language models (LMs) are increasingly being deployed to perform autonomous data analyses. However, their data awareness -- the ability to recognize, reason over, and appropriately handle data artifacts such as missing values, outliers, and logical inconsistencies -- remains underexplored. These artifacts are especially common in real-world tabular data and, if mishandled, can significantly compromise the validity of analytical conclusions. To address this gap, we present RADAR, a benchmark for systematically evaluating data-aware reasoning on tabular data. We develop a framework to simulate data artifacts via programmatic perturbations to enable targeted evaluation of model behavior. RADAR comprises 2980 table query pairs, grounded in real-world data spanning 9 domains and 5 data artifact types. In addition to evaluating artifact handling, RADAR systematically varies table size to study how reasoning performance holds when increasing table size. Our evaluation reveals that, despite decent performance on tables without data artifacts, frontier models degrade significantly when data artifacts are introduced, exposing critical gaps in their capacity for robust, data-aware analysis. Designed to be flexible and extensible, RADAR supports diverse perturbation types and controllable table sizes, offering a valuable resource for advancing tabular reasoning.
Authors: Yaswanth Chittepu, Blossom Metevier, Will Schwarzer, Austin Hoag, Scott Niekum, Philip S. Thomas
Abstract: Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs a two-step process to find safe solutions. In the first step, it optimizes the reward function under an intentionally pessimistic version of the cost constraint. In the second step, the trained model undergoes a safety test to verify whether its performance stays within an upper-confidence bound of the actual cost constraint. We provide a theoretical analysis of HC-RLHF, including proof that it will not return an unsafe solution with a probability greater than a user-specified threshold. For our empirical analysis, we apply HC-RLHF to align three different language models (Qwen2-1.5B, Qwen2.5-3B, and LLaMa3.2-3B) with human preferences. Our results demonstrate that HC-RLHF produces safe models with high probability and can improve harmlessness and helpfulness compared to previous methods.
Authors: Subba Reddy Oota, Khushbu Pahwa, Prachi Jindal, Satya Sai Srinath Namburi, Maneesh Singh, Tanmoy Chakraborty, Bapi S. Raju, Manish Gupta
Abstract: Recent voxel-wise multimodal brain encoding studies have shown that multimodal large language models (MLLMs) exhibit a higher degree of brain alignment compared to unimodal models in both unimodal and multimodal stimulus settings. More recently, instruction-tuned multimodal models have shown to generate task-specific representations that align strongly with brain activity. However, prior work evaluating the brain alignment of MLLMs has primarily focused on unimodal settings or relied on non-instruction-tuned multimodal models for multimodal stimuli. To address this gap, we investigated brain alignment, that is, measuring the degree of predictivity of neural activity recorded while participants were watching naturalistic movies (video along with audio) with representations derived from MLLMs. We utilized instruction-specific embeddings from six video and two audio instruction-tuned MLLMs. Experiments with 13 video task-specific instructions show that instruction-tuned video MLLMs significantly outperform non-instruction-tuned multimodal (by 15%) and unimodal models (by 20%). Our evaluation of MLLMs for both video and audio tasks using language-guided instructions shows clear disentanglement in task-specific representations from MLLMs, leading to precise differentiation of multimodal functional processing in the brain. We also find that MLLM layers align hierarchically with the brain, with early sensory areas showing strong alignment with early layers, while higher-level visual and language regions align more with middle to late layers. These findings provide clear evidence for the role of task-specific instructions in improving the alignment between brain activity and MLLMs, and open new avenues for mapping joint information processing in both the systems. We make the code publicly available [https://github.com/subbareddy248/mllm_videos].
Authors: Xie Yi, Zhanke Zhou, Chentao Cao, Qiyu Niu, Tongliang Liu, Bo Han
Abstract: Multi-agent frameworks can substantially boost the reasoning power of large language models (LLMs), but they typically incur heavy computational costs and lack convergence guarantees. To overcome these challenges, we recast multi-LLM coordination as an incomplete-information game and seek a Bayesian Nash equilibrium (BNE), in which each agent optimally responds to its probabilistic beliefs about the strategies of others. We introduce Efficient Coordination via Nash Equilibrium (ECON), a hierarchical reinforcement-learning paradigm that marries distributed reasoning with centralized final output. Under ECON, each LLM independently selects responses that maximize its expected reward, conditioned on its beliefs about co-agents, without requiring costly inter-agent exchanges. We mathematically prove that ECON attains a markedly tighter regret bound than non-equilibrium multi-agent schemes. Empirically, ECON outperforms existing multi-LLM approaches by 11.2% on average across six benchmarks spanning complex reasoning and planning tasks. Further experiments demonstrate ECON's ability to flexibly incorporate additional models, confirming its scalability and paving the way toward larger, more powerful multi-LLM ensembles. The code is publicly available at: https://github.com/tmlr-group/ECON.
Authors: Zhanke Zhou, Xiao Feng, Zhaocheng Zhu, Jiangchao Yao, Sanmi Koyejo, Bo Han
Abstract: While existing benchmarks probe the reasoning abilities of large language models (LLMs) across diverse domains, they predominantly assess passive reasoning, providing models with all the information needed to reach a solution. By contrast, active reasoning-where an LLM must interact with external systems to acquire missing evidence or data-has received little systematic attention. To address this shortfall, we present AR-Bench, a novel benchmark designed explicitly to evaluate an LLM's active reasoning skills. AR-Bench comprises three task families-detective cases, situation puzzles, and guessing numbers-that together simulate real-world, agentic scenarios and measure performance across commonsense, logical, and symbolic reasoning challenges. Empirical evaluation on AR-Bench demonstrates that contemporary LLMs exhibit pronounced difficulties with active reasoning: they frequently fail to acquire or leverage the information needed to solve tasks. This gap highlights a stark divergence between their passive and active reasoning abilities. Moreover, ablation studies indicate that even advanced strategies, such as tree-based searching or post-training approaches, yield only modest gains and fall short of the levels required for real-world deployment. Collectively, these findings highlight the critical need to advance methodology for active reasoning, e.g., incorporating interactive learning, real-time feedback loops, and environment-aware objectives for training. The benchmark is publicly available at: https://github.com/tmlr-group/AR-Bench.
Authors: Huixuan Zhang, Junzhe Zhang, Xiaojun Wan
Abstract: With the rapid development of text-to-vision generation diffusion models, classifier-free guidance has emerged as the most prevalent method for conditioning. However, this approach inherently requires twice as many steps for model forwarding compared to unconditional generation, resulting in significantly higher costs. While previous study has introduced the concept of adaptive guidance, it lacks solid analysis and empirical results, making previous method unable to be applied to general diffusion models. In this work, we present another perspective of applying adaptive guidance and propose Step AG, which is a simple, universally applicable adaptive guidance strategy. Our evaluations focus on both image quality and image-text alignment. whose results indicate that restricting classifier-free guidance to the first several denoising steps is sufficient for generating high-quality, well-conditioned images, achieving an average speedup of 20% to 30%. Such improvement is consistent across different settings such as inference steps, and various models including video generation models, highlighting the superiority of our method.
Authors: Yurun Yuan, Tengyang Xie
Abstract: Leveraging more test-time computation has proven to be an effective way to boost the reasoning capabilities of large language models (LLMs). Among various methods, the verify-and-improve paradigm stands out for enabling dynamic solution exploration and feedback incorporation. However, existing approaches often suffer from restricted feedback spaces and lack of coordinated training of different parties, leading to suboptimal performance. To address this, we model this multi-turn refinement process as a Markov Decision Process and introduce DPSDP (Direct Policy Search by Dynamic Programming), a reinforcement learning algorithm that trains an actor-critic LLM system to iteratively refine answers via direct preference learning on self-generated data. Theoretically, DPSDP can match the performance of any policy within the training distribution. Empirically, we instantiate DPSDP with various base models and show improvements on both in- and out-of-distribution benchmarks. For example, on benchmark MATH 500, majority voting over five refinement steps increases first-turn accuracy from 58.2% to 63.2% with Ministral-based models. An ablation study further confirms the benefits of multi-agent collaboration and out-of-distribution generalization.
Authors: Edoardo Cetin, Tianyu Zhao, Yujin Tang
Abstract: Training reasoning language models (LMs) with reinforcement learning (RL) for one-hot correctness inherently relies on the LM being able to explore and solve its task with some chance at initialization. Furthermore, a key use case of reasoning LMs is to act as teachers for distilling new students and cold-starting future RL iterations rather than being deployed themselves. From these considerations, we introduce a new framework that avoids RL's exploration challenge by training a new class of Reinforcement-Learned Teachers (RLTs) focused on yielding the most effective downstream distillation. RLTs are prompted with both the question and solution to each problem, and tasked to simply "connect-the-dots" with detailed explanations tailored for their students. We train RLTs with dense rewards obtained by feeding each explanation to the student and testing its understanding of the problem's solution. In practice, the raw outputs of a 7B RLT provide higher final performance on competition and graduate-level tasks than existing distillation and cold-starting pipelines that collect and postprocess the reasoning traces of orders of magnitude larger LMs. Furthermore, RLTs maintain their effectiveness when training larger students and when applied zero-shot to out-of-distribution tasks, unlocking new levels of efficiency and re-usability for the RL reasoning framework.
Authors: Peng-Yuan Wang, Tian-Shuo Liu, Chenyang Wang, Yi-Di Wang, Shu Yan, Cheng-Xing Jia, Xu-Hui Liu, Xin-Wei Chen, Jia-Cheng Xu, Ziniu Li, Yang Yu
Abstract: Mathematical reasoning has long represented one of the most fundamental and challenging frontiers in artificial intelligence research. In recent years, large language models (LLMs) have achieved significant advances in this area. This survey examines the development of mathematical reasoning abilities in LLMs through two high-level cognitive phases: comprehension, where models gain mathematical understanding via diverse pretraining strategies, and answer generation, which has progressed from direct prediction to step-by-step Chain-of-Thought (CoT) reasoning. We review methods for enhancing mathematical reasoning, ranging from training-free prompting to fine-tuning approaches such as supervised fine-tuning and reinforcement learning, and discuss recent work on extended CoT and "test-time scaling". Despite notable progress, fundamental challenges remain in terms of capacity, efficiency, and generalization. To address these issues, we highlight promising research directions, including advanced pretraining and knowledge augmentation techniques, formal reasoning frameworks, and meta-generalization through principled learning paradigms. This survey tries to provide some insights for researchers interested in enhancing reasoning capabilities of LLMs and for those seeking to apply these techniques to other domains.
Authors: Waiss Azizian, Michael Kirchhof, Eugene Ndiaye, Louis Bethune, Michal Klein, Pierre Ablin, Marco Cuturi
Abstract: Large Language Models (LLMs) have demonstrated impressive generalization capabilities across various tasks, but their claim to practical relevance is still mired by concerns on their reliability. Recent works have proposed examining the activations produced by an LLM at inference time to assess whether its answer to a question is correct. Some works claim that a "geometry of truth" can be learned from examples, in the sense that the activations that generate correct answers can be distinguished from those leading to mistakes with a linear classifier. In this work, we underline a limitation of these approaches: we observe that these "geometries of truth" are intrinsically task-dependent and fail to transfer across tasks. More precisely, we show that linear classifiers trained across distinct tasks share little similarity and, when trained with sparsity-enforcing regularizers, have almost disjoint supports. We show that more sophisticated approaches (e.g., using mixtures of probes and tasks) fail to overcome this limitation, likely because activation vectors commonly used to classify answers form clearly separated clusters when examined across tasks.
Authors: \v{S}imon Sedl\'a\v{c}ek, Bolaji Yusuf, J\'an \v{S}vec, Pradyoth Hegde, Santosh Kesiraju, Old\v{r}ich Plchot, Jan \v{C}ernock\'y
Abstract: In this work, we approach spoken Dialogue State Tracking (DST) by bridging the representation spaces of speech encoders and LLMs via a small connector module, with a focus on fully open-sourced and open-data components (WavLM-large, OLMo). We focus on ablating different aspects of such systems including full/LoRA adapter fine-tuning, the effect of agent turns in the dialogue history, as well as fuzzy matching-based output post-processing, which greatly improves performance of our systems on named entities in the dialogue slot values. We conduct our experiments on the SpokenWOZ dataset, and additionally utilize the Speech-Aware MultiWOZ dataset to augment our training data. Ultimately, our best-performing WavLM + connector + OLMo-1B aligned models achieve state of the art on the SpokenWOZ test set (34.66% JGA), and our system with Gemma-2-9B-instruct further surpasses this result, reaching 42.17% JGA on SpokenWOZ test.
Authors: Kongcheng Zhang, Qi Yao, Shunyu Liu, Yingjie Wang, Baisheng Lai, Jieping Ye, Mingli Song, Dacheng Tao
Abstract: Recent advances of Reinforcement Learning (RL) have highlighted its potential in complex reasoning tasks, yet effective training often relies on external supervision, which limits the broader applicability. In this work, we propose a novel self-rewarding reinforcement learning framework to enhance Large Language Model (LLM) reasoning by leveraging the consistency of intermediate reasoning states across different reasoning trajectories. Our key insight is that correct responses often exhibit consistent trajectory patterns in terms of model likelihood: their intermediate reasoning states tend to converge toward their own final answers (high consistency) with minimal deviation toward other candidates (low volatility). Inspired by this observation, we introduce CoVo, an intrinsic reward mechanism that integrates Consistency and Volatility via a robust vector-space aggregation strategy, complemented by a curiosity bonus to promote diverse exploration. CoVo enables LLMs to perform RL in a self-rewarding manner, offering a scalable pathway for learning to reason without external supervision. Extensive experiments on diverse reasoning benchmarks show that CoVo achieves performance comparable to or even surpassing supervised RL. Our code is available at https://github.com/sastpg/CoVo.
Authors: Issa Sugiura, Takashi Ishida, Taro Makino, Chieko Tazuke, Takanori Nakagawa, Kosuke Nakago, David Ha
Abstract: Financial analysis presents complex challenges that could leverage large language model (LLM) capabilities. However, the scarcity of challenging financial datasets, particularly for Japanese financial data, impedes academic innovation in financial analytics. As LLMs advance, this lack of accessible research resources increasingly hinders their development and evaluation in this specialized domain. To address this gap, we introduce EDINET-Bench, an open-source Japanese financial benchmark designed to evaluate the performance of LLMs on challenging financial tasks including accounting fraud detection, earnings forecasting, and industry prediction. EDINET-Bench is constructed by downloading annual reports from the past 10 years from Japan's Electronic Disclosure for Investors' NETwork (EDINET) and automatically assigning labels corresponding to each evaluation task. Our experiments reveal that even state-of-the-art LLMs struggle, performing only slightly better than logistic regression in binary classification for fraud detection and earnings forecasting. These results highlight significant challenges in applying LLMs to real-world financial applications and underscore the need for domain-specific adaptation. Our dataset, benchmark construction code, and evaluation code is publicly available to facilitate future research in finance with LLMs.
Authors: Yuni Susanti, Michael F\"arber
Abstract: Inferring causal relationships between variable pairs is crucial for understanding multivariate interactions in complex systems. Knowledge-based causal discovery -- which involves inferring causal relationships by reasoning over the metadata of variables (e.g., names or textual context) -- offers a compelling alternative to traditional methods that rely on observational data. However, existing methods using Large Language Models (LLMs) often produce unstable and inconsistent results, compromising their reliability for causal inference. To address this, we introduce a novel approach that integrates Knowledge Graphs (KGs) with LLMs to enhance knowledge-based causal discovery. Our approach identifies informative metapath-based subgraphs within KGs and further refines the selection of these subgraphs using Learning-to-Rank-based models. The top-ranked subgraphs are then incorporated into zero-shot prompts, improving the effectiveness of LLMs in inferring the causal relationship. Extensive experiments on biomedical and open-domain datasets demonstrate that our method outperforms most baselines by up to 44.4 points in F1 scores, evaluated across diverse LLMs and KGs. Our code and datasets are available on GitHub: https://github.com/susantiyuni/path-to-causality
Authors: Irene Testini, Jos\'e Hern\'andez-Orallo, Lorenzo Pacchiardi
Abstract: Data science aims to extract insights from data to support decision-making processes. Recently, Large Language Models (LLMs) are increasingly used as assistants for data science, by suggesting ideas, techniques and small code snippets, or for the interpretation of results and reporting. Proper automation of some data-science activities is now promised by the rise of LLM agents, i.e., AI systems powered by an LLM equipped with additional affordances--such as code execution and knowledge bases--that can perform self-directed actions and interact with digital environments. In this paper, we survey the evaluation of LLM assistants and agents for data science. We find (1) a dominant focus on a small subset of goal-oriented activities, largely ignoring data management and exploratory activities; (2) a concentration on pure assistance or fully autonomous agents, without considering intermediate levels of human-AI collaboration; and (3) an emphasis on human substitution, therefore neglecting the possibility of higher levels of automation thanks to task transformation.
Authors: Shravan Nayak, Mehar Bhatia, Xiaofeng Zhang, Verena Rieser, Lisa Anne Hendricks, Sjoerd van Steenkiste, Yash Goyal, Karolina Sta\'nczak, Aishwarya Agrawal
Abstract: The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit as well as implicit cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that T2I models not only fail to meet the more challenging implicit expectations but also the less challenging explicit expectations. Across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we demonstrate that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, providing actionable directions for developing more culturally informed T2I models and evaluation methodologies.
Authors: David Acuna, Ximing Lu, Jaehun Jung, Hyunwoo Kim, Amlan Kar, Sanja Fidler, Yejin Choi
Abstract: Recent research in vision-language models (VLMs) has centered around the possibility of equipping them with implicit long-form chain-of-thought reasoning -- akin to the success observed in language models -- via distillation and reinforcement learning. But what about the non-reasoning models already trained and deployed across the internet? Should we simply abandon them, or is there hope for a search mechanism that can elicit hidden knowledge and induce long reasoning traces -- without any additional training or supervision? In this paper, we explore this possibility using a Monte Carlo Tree Search (MCTS)-inspired algorithm, which injects subquestion-subanswer pairs into the model's output stream. We show that framing reasoning as a search process -- where subquestions act as latent decisions within a broader inference trajectory -- helps the model "connect the dots" between fragmented knowledge and produce extended reasoning traces in non-reasoning models. We evaluate our method across three benchmarks and observe consistent improvements. Notably, our approach yields a 2% overall improvement on MMMU-PRO, including a significant 9% gain in Liberal Arts.
Authors: Ailin Huang, Bingxin Li, Bruce Wang, Boyong Wu, Chao Yan, Chengli Feng, Heng Wang, Hongyu Zhou, Hongyuan Wang, Jingbei Li, Jianjian Sun, Joanna Wang, Mingrui Chen, Peng Liu, Ruihang Miao, Shilei Jiang, Tian Fei, Wang You, Xi Chen, Xuerui Yang, Yechang Huang, Yuxiang Zhang, Zheng Ge, Zheng Gong, Zhewei Huang, Zixin Zhang, Bin Wang, Bo Li, Buyun Ma, Changxin Miao, Changyi Wan, Chen Xu, Dapeng Shi, Dingyuan Hu, Enle Liu, Guanzhe Huang, Gulin Yan, Hanpeng Hu, Haonan Jia, Jiahao Gong, Jiaoren Wu, Jie Wu, Jie Yang, Junzhe Lin, Kaixiang Li, Lei Xia, Longlong Gu, Ming Li, Nie Hao, Ranchen Ming, Shaoliang Pang, Siqi Liu, Song Yuan, Tiancheng Cao, Wen Li, Wenqing He, Xu Zhao, Xuelin Zhang, Yanbo Yu, Yinmin Zhong, Yu Zhou, Yuanwei Liang, Yuanwei Lu, Yuxiang Yang, Zidong Yang, Zili Zhang, Binxing Jiao, Heung-Yeung Shum, Jiansheng Chen, Jing Li, Xiangyu Zhang, Xinhao Zhang, Yibo Zhu, Daxin Jiang, Shuchang Zhou, Chen Hu
Abstract: Large Audio-Language Models (LALMs) have significantly advanced intelligent human-computer interaction, yet their reliance on text-based outputs limits their ability to generate natural speech responses directly, hindering seamless audio interactions. To address this, we introduce Step-Audio-AQAA, a fully end-to-end LALM designed for Audio Query-Audio Answer (AQAA) tasks. The model integrates a dual-codebook audio tokenizer for linguistic and semantic feature extraction, a 130-billion-parameter backbone LLM and a neural vocoder for high-fidelity speech synthesis. Our post-training approach employs interleaved token-output of text and audio to enhance semantic coherence and combines Direct Preference Optimization (DPO) with model merge to improve performance. Evaluations on the StepEval-Audio-360 benchmark demonstrate that Step-Audio-AQAA excels especially in speech control, outperforming the state-of-art LALMs in key areas. This work contributes a promising solution for end-to-end LALMs and highlights the critical role of token-based vocoder in enhancing overall performance for AQAA tasks.
Authors: Xiao Liang, Zhong-Zhi Li, Yeyun Gong, Yang Wang, Hengyuan Zhang, Yelong Shen, Ying Nian Wu, Weizhu Chen
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for training large language models (LLMs) on complex reasoning tasks, such as mathematical problem solving. A prerequisite for the scalability of RLVR is a high-quality problem set with precise and verifiable answers. However, the scarcity of well-crafted human-labeled math problems and limited-verification answers in existing distillation-oriented synthetic datasets limit their effectiveness in RL. Additionally, most problem synthesis strategies indiscriminately expand the problem set without considering the model's capabilities, leading to low efficiency in generating useful questions. To mitigate this issue, we introduce a Self-aware Weakness-driven problem Synthesis framework (SwS) that systematically identifies model deficiencies and leverages them for problem augmentation. Specifically, we define weaknesses as questions that the model consistently fails to learn through its iterative sampling during RL training. We then extract the core concepts from these failure cases and synthesize new problems to strengthen the model's weak areas in subsequent augmented training, enabling it to focus on and gradually overcome its weaknesses. Without relying on external knowledge distillation, our framework enables robust generalization byempowering the model to self-identify and address its weaknesses in RL, yielding average performance gains of 10.0% and 7.7% on 7B and 32B models across eight mainstream reasoning benchmarks.
Authors: Amrith Setlur, Matthew Y. R. Yang, Charlie Snell, Jeremy Greer, Ian Wu, Virginia Smith, Max Simchowitz, Aviral Kumar
Abstract: Test-time scaling offers a promising path to improve LLM reasoning by utilizing more compute at inference time; however, the true promise of this paradigm lies in extrapolation (i.e., improvement in performance on hard problems as LLMs keep "thinking" for longer, beyond the maximum token budget they were trained on). Surprisingly, we find that most existing reasoning models do not extrapolate well. We show that one way to enable extrapolation is by training the LLM to perform in-context exploration: training the LLM to effectively spend its test time budget by chaining operations (such as generation, verification, refinement, etc.), or testing multiple hypotheses before it commits to an answer. To enable in-context exploration, we identify three key ingredients as part of our recipe e3: (1) chaining skills that the base LLM has asymmetric competence in, e.g., chaining verification (easy) with generation (hard), as a way to implement in-context search; (2) leveraging "negative" gradients from incorrect traces to amplify exploration during RL, resulting in longer search traces that chains additional asymmetries; and (3) coupling task difficulty with training token budget during training via a specifically-designed curriculum to structure in-context exploration. Our recipe e3 produces the best known 1.7B model according to AIME'25 and HMMT'25 scores, and extrapolates to 2x the training token budget. Our e3-1.7B model not only attains high pass@1 scores, but also improves pass@k over the base model.
Authors: Dianyi Wang, Wei Song, Yikun Wang, Siyuan Wang, Kaicheng Yu, Zhongyu Wei, Jiaqi Wang
Abstract: Typical large vision-language models (LVLMs) apply autoregressive supervision solely to textual sequences, without fully incorporating the visual modality into the learning process. This results in three key limitations: (1) an inability to utilize images without accompanying captions, (2) the risk that captions omit critical visual details, and (3) the challenge that certain vision-centric content cannot be adequately conveyed through text. As a result, current LVLMs often prioritize vision-to-language alignment while potentially overlooking fine-grained visual information. While some prior works have explored autoregressive image generation, effectively leveraging autoregressive visual supervision to enhance image understanding remains an open challenge. In this paper, we introduce Autoregressive Semantic Visual Reconstruction (ASVR), which enables joint learning of visual and textual modalities within a unified autoregressive framework. We show that autoregressively reconstructing the raw visual appearance of images does not enhance and may even impair multimodal understanding. In contrast, autoregressively reconstructing the semantic representation of images consistently improves comprehension. Notably, we find that even when models are given continuous image features as input, they can effectively reconstruct discrete semantic tokens, resulting in stable and consistent improvements across a wide range of multimodal understanding benchmarks. Our approach delivers significant performance gains across varying data scales (556k-2M) and types of LLM bacbones. Specifically, ASVR improves LLaVA-1.5 by 5% in average scores across 14 multimodal benchmarks. The code is available at https://github.com/AlenjandroWang/ASVR.
Authors: Effi Levi, Shaul R. Shenhav
Abstract: We propose a novel method to conceptually decompose an existing annotation into separate levels, allowing the analysis of inter-annotators disagreement in each level separately. We suggest two distinct strategies in order to actualize this approach: a theoretically-driven one, in which the researcher defines a decomposition based on prior knowledge of the annotation task, and an exploration-based one, in which many possible decompositions are inductively computed and presented to the researcher for interpretation and evaluation. Utilizing a recently constructed dataset for narrative analysis as our use-case, we apply each of the two strategies to demonstrate the potential of our approach in testing hypotheses regarding the sources of annotation disagreements, as well as revealing latent structures and relations within the annotation task. We conclude by suggesting how to extend and generalize our approach, as well as use it for other purposes.
Authors: Zican Dong, Tianyi Tang, Junyi Li, Wayne Xin Zhao
Abstract: Modeling long texts has been an essential technique in the field of natural language processing (NLP). With the ever-growing number of long documents, it is important to develop effective modeling methods that can process and analyze such texts. However, long texts pose important research challenges for existing text models, with more complex semantics and special characteristics. In this paper, we provide an overview of the recent advances on long texts modeling based on Transformer models. Firstly, we introduce the formal definition of long text modeling. Then, as the core content, we discuss how to process long input to satisfy the length limitation and design improved Transformer architectures to effectively extend the maximum context length. Following this, we discuss how to adapt Transformer models to capture the special characteristics of long texts. Finally, we describe four typical applications involving long text modeling and conclude this paper with a discussion of future directions. Our survey intends to provide researchers with a synthesis and pointer to related work on long text modeling.
Authors: Razan Baltaji, Saurabh Pujar, Louis Mandel, Martin Hirzel, Luca Buratti, Lav Varshney
Abstract: Large language models (LLMs) have achieved state-of-the-art performance in various software engineering tasks, including error detection, clone detection, and code translation, primarily leveraging high-resource programming languages like Python and Java. However, many critical languages, such as COBOL, as well as emerging languages, such as Rust and Swift, remain low-resource due to limited openly available code. This scarcity hampers the training and effectiveness of LLMs for these languages, increasing software maintenance costs and stifling innovation. Addressing this gap, we investigate the potential of transfer learning to enhance LLM performance on low-resource programming languages by leveraging data from high-resource counterparts. Our extensive empirical study evaluates transferability across 10 to 41 programming languages and five key tasks: code generation, clone detection, code repair, solution domain classification, and error detection. Additionally, we develop a performance prediction model to guess the best source languages for a given target and task, and analyze the features that influence transfer performance. We further replicate a representative subset of experiments with a larger model to test the generalizability of our conclusions to contemporary large-scale LLMs. Our findings demonstrate that cross-lingual transfer significantly outperforms zero-shot learning, with effectiveness varying based on both source and target languages. Furthermore, our model reliably predicts successful transfer sources by considering linguistic and dataset-specific features, offering practical guidance for data acquisition and model training. This work contributes to the development of LLM-driven tools for low-resource programming languages and provides insights into the characteristics that facilitate transfer across language pairs.
Authors: Risto Luukkonen, Jonathan Burdge, Elaine Zosa, Aarne Talman, Ville Komulainen, V\"ain\"o Hatanp\"a\"a, Peter Sarlin, Sampo Pyysalo
Abstract: The pretraining of state-of-the-art large language models now requires trillions of words of text, which is orders of magnitude more than available for the vast majority of languages. While including text in more than one language is an obvious way to acquire more pretraining data, multilinguality is often seen as a curse, and most model training efforts continue to focus near-exclusively on individual large languages. We believe that multilinguality can be a blessing: when the lack of training data is a constraint for effectively training larger models for a target language, augmenting the dataset with other languages can offer a way to improve over the capabilities of monolingual models for that language. In this study, we introduce Poro 34B, a 34 billion parameter model trained for 1 trillion tokens of Finnish, English, and programming languages, and demonstrate that a multilingual training approach can produce a model that substantially advances over the capabilities of existing models for Finnish and excels in translation, while also achieving competitive performance in its class for English and programming languages. We release the model parameters, scripts, and data under open licenses at https://huggingface.co/LumiOpen/Poro-34B.
Authors: Yuhao Dan, Jie Zhou, Qin Chen, Junfeng Tian, Liang He
Abstract: Personalized large language models (LLMs) have attracted great attention in many applications, such as emotional support and role-playing. However, existing works primarily focus on modeling explicit character profiles, while ignoring the underlying personality traits that truly shape behaviors and decision-making, hampering the development of more anthropomorphic and psychologically-grounded AI systems. In this paper, we explore the modeling of Big Five personality traits, which is the most widely used trait theory in psychology, and propose P-React, a mixture of experts (MoE)-based personalized LLM. Particularly, we integrate a Personality Specialization Loss (PSL) to better capture individual trait expressions, providing a more nuanced and psychologically grounded personality simulacrum. To facilitate research in this field, we curate OCEAN-Chat, a high-quality, human-verified dataset designed to train LLMs in expressing personality traits across diverse topics. Extensive experiments demonstrate the effectiveness of P-React in maintaining consistent and real personality.
Authors: Xin Su, Man Luo, Kris W Pan, Tien Pei Chou, Vasudev Lal, Phillip Howard
Abstract: Multimodal retrieval augmented generation (RAG) plays a crucial role in domains such as knowledge-based visual question answering (KB-VQA), where external knowledge is needed to answer a question. However, existing multimodal LLMs (MLLMs) are not designed for context-augmented generation, limiting their effectiveness in such tasks. While synthetic data generation has recently gained attention for training MLLMs, its application for context-augmented generation remains underexplored. To address this gap, we introduce SK-VQA, a large-scale synthetic multimodal dataset containing over 2 million visual question-answer pairs, each associated with context documents containing information necessary to determine the final answer. Compared to previous datasets, SK-VQA contains 11x more unique questions, exhibits greater domain diversity, and covers a broader spectrum of image sources. Through human evaluations, we confirm the high quality of the generated question-answer pairs and their contextual relevance. Extensive experiments show that SK-VQA serves both as a challenging KB-VQA benchmark and as an effective training resource for adapting MLLMs to context-augmented generation. Our results further indicate that models trained on SK-VQA demonstrate enhanced generalization in both context-aware VQA and multimodal RAG settings. SK-VQA is publicly available via Hugging Face Hub.
Authors: Daniel B. Hier, S. Ilyas Munzir, Anne Stahlfeld, Tayo Obafemi-Ajayi, Michael D. Carrithers
Abstract: High-throughput phenotyping automates the mapping of patient signs to standardized ontology concepts and is essential for precision medicine. This study evaluates the automation of phenotyping of clinical summaries from the Online Mendelian Inheritance in Man (OMIM) database using large language models. Due to their rich phenotype data, these summaries can be surrogates for physician notes. We conduct a performance comparison of GPT-4 and GPT-3.5-Turbo. Our results indicate that GPT-4 surpasses GPT-3.5-Turbo in identifying, categorizing, and normalizing signs, achieving concordance with manual annotators comparable to inter-rater agreement. Despite some limitations in sign normalization, the extensive pre-training of GPT-4 results in high performance and generalizability across several phenotyping tasks while obviating the need for manually annotated training data. Large language models are expected to be the dominant method for automating high-throughput phenotyping of clinical text.
Authors: Mingyu Cui, Yifan Yang, Jiajun Deng, Jiawen Kang, Shujie Hu, Tianzi Wang, Zhaoqing Li, Shiliang Zhang, Xie Chen, Xunying Liu
Abstract: Self-supervised learning (SSL) based discrete speech representations are highly compact and domain adaptable. In this paper, SSL discrete speech features extracted from WavLM models are used as additional cross-utterance acoustic context features in Zipformer-Transducer ASR systems. The efficacy of replacing Fbank features with discrete token features for modelling either cross-utterance contexts (from preceding and future segments), or current utterance's internal contexts alone, or both at the same time, are demonstrated thoroughly on the Gigaspeech 1000-hr corpus. The best Zipformer-Transducer system using discrete tokens based cross-utterance context features outperforms the baseline using utterance internal context only with statistically significant word error rate (WER) reductions of 0.32% to 0.41% absolute (2.78% to 3.54% relative) on the dev and test data. The lowest published WER of 11.15% and 11.14% were obtained on the dev and test sets. Our work is open-source and publicly available at https://github.com/open-creator/icefall/tree/master/egs/gigaspeech/Context\_ASR.
URLs: https://github.com/open-creator/icefall/tree/master/egs/gigaspeech/Context\_ASR.
Authors: Nizar Habash, Hanada Taha-Thomure, Khalid N. Elmadani, Zeina Zeino, Abdallah Abushmaes
Abstract: This paper presents the annotation guidelines of the Balanced Arabic Readability Evaluation Corpus (BAREC), a large-scale resource for fine-grained sentence-level readability assessment in Arabic. BAREC includes 69,441 sentences (1M+ words) labeled across 19 levels, from kindergarten to postgraduate. Based on the Taha/Arabi21 framework, the guidelines were refined through iterative training with native Arabic-speaking educators. We highlight key linguistic, pedagogical, and cognitive factors in determining readability and report high inter-annotator agreement: Quadratic Weighted Kappa 81.8% (substantial/excellent agreement) in the last annotation phase. We also benchmark automatic readability models across multiple classification granularities (19-, 7-, 5-, and 3-level). The corpus and guidelines are publicly available.
Authors: Yoichi Ishibashi, Taro Yano, Masafumi Oyamada
Abstract: Large Language Models (LLMs) have achieved remarkable capabilities, yet their improvement methods remain fundamentally constrained by human design. We present Self-Developing, a framework that enables LLMs to autonomously discover, implement, and refine their own improvement algorithms. Our approach employs an iterative cycle where a seed model generates algorithmic candidates as executable code, evaluates their effectiveness, and uses Direct Preference Optimization to recursively improve increasingly sophisticated improvement strategies. We demonstrate this framework through model merging, a practical technique for combining specialized models. Self-Developing successfully discovered novel merging algorithms that outperform existing human-designed algorithms. On mathematical reasoning benchmarks, the autonomously discovered algorithms improve the seed model's GSM8k performance by 6\% and exceed human-designed approaches like Task Arithmetic by 4.3\%. Remarkably, these algorithms exhibit strong generalization, achieving 7.4\% gains on out-of-domain models without re-optimization. Our findings demonstrate that LLMs can transcend their training to invent genuinely novel optimization techniques. This capability represents a crucial step toward a new era where LLMs not only solve problems but autonomously develop the methodologies for their own advancement.
Authors: Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Zuozhu Liu
Abstract: The growing use of large language model (LLM)-based chatbots has raised concerns about fairness. Fairness issues in LLMs can lead to severe consequences, such as bias amplification, discrimination, and harm to marginalized communities. While existing fairness benchmarks mainly focus on single-turn dialogues, multi-turn scenarios, which in fact better reflect real-world conversations, present greater challenges due to conversational complexity and potential bias accumulation. In this paper, we propose a comprehensive fairness benchmark for LLMs in multi-turn dialogue scenarios, \textbf{FairMT-Bench}. Specifically, we formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. To ensure coverage of diverse bias types and attributes, we draw from existing fairness datasets and employ our template to construct a multi-turn dialogue dataset, \texttt{FairMT-10K}. For evaluation, GPT-4 is applied, alongside bias classifiers including Llama-Guard-3 and human validation to ensure robustness. Experiments and analyses on \texttt{FairMT-10K} reveal that in multi-turn dialogue scenarios, current LLMs are more likely to generate biased responses, and there is significant variation in performance across different tasks and models. Based on this, we curate a challenging dataset, \texttt{FairMT-1K}, and test 15 current state-of-the-art (SOTA) LLMs on this dataset. The results show the current state of fairness in LLMs and showcase the utility of this novel approach for assessing fairness in more realistic multi-turn dialogue contexts, calling for future work to focus on LLM fairness improvement and the adoption of \texttt{FairMT-1K} in such efforts.
Authors: Eunwon Kim, Chanho Park, Buru Chang
Abstract: Shared memories between two individuals strengthen their bond and are crucial for facilitating their ongoing conversations. This study aims to make long-term dialogue more engaging by leveraging these shared memories. To this end, we introduce a new long-term dialogue dataset named SHARE, constructed from movie scripts, which are a rich source of shared memories among various relationships. Our dialogue dataset contains the summaries of persona information and events of two individuals, as explicitly revealed in their conversation, along with implicitly extractable shared memories. We also introduce EPISODE, a long-term dialogue framework based on SHARE that utilizes shared experiences between individuals. Through experiments using SHARE, we demonstrate that shared memories between two individuals make long-term dialogues more engaging and sustainable, and that EPISODE effectively manages shared memories during dialogue. Our dataset and code are available at https://github.com/e1kim/SHARE.
Authors: Yuqi Zhou, Sunhao Dai, Zhanshuo Cao, Xiao Zhang, Jun Xu
Abstract: Text embeddings from PLM-based models enable a wide range of applications, yet their performance often degrades on longer texts. In this paper, we introduce a phenomenon we call Length Collapse, where embeddings of longer texts tend to cluster together. This clustering results in a distributional inconsistency between the embeddings of short and long texts. We further investigate how these differences contribute to the performance decline observed with longer texts across various downstream tasks. Through a rigorous theoretical analysis of the self-attention mechanism, which acts as a low-pass filter in PLM-based models, we demonstrate that as text length increases, the strength of low-pass filtering intensifies, causing embeddings to retain more low-frequency components. As a result, input token features become more similar, leading to clustering and ultimately the collapse of embeddings for longer texts. To address this issue, we propose a simple method, TempScale, which mitigates the Length Collapse phenomenon. By narrowing the gap in low-pass filtering rates between long and short texts, TempScale ensures more consistent embeddings across different text lengths. This approach leads to performance improvements of 0.94% on MTEB and 1.10% on LongEmbed, which focuses specifically on long-context retrieval, providing strong evidence for the validity of our analysis. The source code is available at https://github.com/Yuqi-Zhou/Length_Collapse.
Authors: Alessandro Trevisan, Harry Giddens, Sarah Dillon, Alan F. Blackwell
Abstract: What can we learn about language from studying how it is used by ChatGPT and other large language model (LLM)-based chatbots? In this paper, we analyse the distinctive character of language generated by ChatGPT, in relation to questions raised by natural language processing pioneer, and student of Wittgenstein, Margaret Masterman. Following frequent complaints that LLM-based chatbots produce "slop," or even "bullshit," in the sense of Frankfurt's popular monograph On Bullshit, we conduct an empirical study to contrast the language of 1,000 scientific publications with typical text generated by ChatGPT. We then explore whether the same language features can be detected in two well-known contexts of social dysfunction: George Orwell's critique of political speech, and David Graeber's characterisation of bullshit jobs. Using simple hypothesis-testing methods, we demonstrate that a statistical model of sloppy bullshit can reliably relate the Frankfurtian artificial bullshit of ChatGPT to the political and workplace functions of bullshit as observed in natural human language.
Authors: Tim Vieira, Ben LeBrun, Mario Giulianelli, Juan Luis Gastaldi, Brian DuSell, John Terilla, Timothy J. O'Donnell, Ryan Cotterell
Abstract: Modern language models are internally -- and mathematically -- distributions over $\it{token}$ strings rather than $\it{character}$ strings, posing numerous challenges for programmers building user applications on top of them. For example, if a prompt is specified as a character string, it must be tokenized before passing it to the token-level language model. Thus, the tokenizer and consequent processing are very sensitive to the specification of the prompt (e.g., whether the prompt ends with a space or not). This paper presents algorithms for converting token-level language models to character-level ones. We present both exact and approximate algorithms. In the empirical portion of the paper, we benchmark the practical runtime and approximation quality. Across four publicly available language models, we find that -- even with a small computation budget -- our method is able to accurately approximate the character-level distribution at reasonably fast speeds, and that a significant improvement in the language model's compression rate (bits/byte) is achieved.
Authors: Ariel Gera, Odellia Boni, Yotam Perlitz, Roy Bar-Haim, Lilach Eden, Asaf Yehudai
Abstract: Given the rapid progress of generative AI, there is a pressing need to systematically compare and choose between the numerous models and configurations available. The scale and versatility of such evaluations make the use of LLM-based judges a compelling solution for this challenge. Crucially, this approach requires first to validate the quality of the LLM judge itself. Previous work has focused on instance-based assessment of LLM judges, where a judge is evaluated over a set of responses, or response pairs, while being agnostic to their source systems. We argue that this setting overlooks critical factors affecting system-level ranking, such as a judge's positive or negative bias towards certain systems. To address this gap, we conduct the first large-scale study of LLM judges as system rankers. System scores are generated by aggregating judgment scores over multiple system outputs, and the judge's quality is assessed by comparing the resulting system ranking to a human-based ranking. Beyond overall judge assessment, our analysis provides a fine-grained characterization of judge behavior, including their decisiveness and bias.
Authors: Jinghan He, Kuan Zhu, Haiyun Guo, Junfeng Fang, Zhenglin Hua, Yuheng Jia, Ming Tang, Tat-Seng Chua, Jinqiao Wang
Abstract: Large vision-language models (LVLMs) have made substantial progress in integrating large language models (LLMs) with visual inputs, enabling advanced multimodal reasoning. Despite their success, a persistent challenge is hallucination-where generated text fails to accurately reflect visual content-undermining both accuracy and reliability. Existing methods focus on alignment training or decoding refinements but primarily address symptoms at the generation stage without probing the underlying causes. In this work, we investigate the internal mechanisms driving hallucination in LVLMs, with an emphasis on the multi-head attention module. Specifically, we introduce Vision-aware Head Divergence (VHD), a metric that quantifies the sensitivity of attention head outputs to visual context. Based on this, our findings reveal the presence of vision-aware attention heads that are more attuned to visual information; however, the model's overreliance on its prior language patterns is closely related to hallucinations. Building on these insights, we propose Vision-aware Head Reinforcement (VHR), a training-free approach to mitigate hallucination by enhancing the role of vision-aware attention heads. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches in mitigating hallucinations, while maintaining high efficiency with negligible additional time overhead.
Authors: Haoran Ye, Tianze Zhang, Yuhang Xie, Liyuan Zhang, Yuanyi Ren, Xin Zhang, Guojie Song
Abstract: Values are core drivers of individual and collective perception, cognition, and behavior. Value systems, such as Schwartz's Theory of Basic Human Values, delineate the hierarchy and interplay among these values, enabling cross-disciplinary investigations into decision-making and societal dynamics. Recently, the rise of Large Language Models (LLMs) has raised concerns regarding their elusive intrinsic values. Despite growing efforts in evaluating, understanding, and aligning LLM values, a psychologically grounded LLM value system remains underexplored. This study addresses the gap by introducing the Generative Psycho-Lexical Approach (GPLA), a scalable, adaptable, and theoretically informed method for constructing value systems. Leveraging GPLA, we propose a psychologically grounded five-factor value system tailored for LLMs. For systematic validation, we present three benchmarking tasks that integrate psychological principles with cutting-edge AI priorities. Our results reveal that the proposed value system meets standard psychological criteria, better captures LLM values, improves LLM safety prediction, and enhances LLM alignment, when compared to the canonical Schwartz's values.
Authors: Paul Youssef, Zhixue Zhao, Daniel Braun, J\"org Schl\"otterer, Christin Seifert
Abstract: Large Language Models (LLMs) contain large amounts of facts about the world. These facts can become outdated over time, which has led to the development of knowledge editing methods (KEs) that can change specific facts in LLMs with limited side effects. This position paper argues that editing LLMs poses serious safety risks that have been largely overlooked. First, we note the fact that KEs are widely available, computationally inexpensive, highly performant, and stealthy makes them an attractive tool for malicious actors. Second, we discuss malicious use cases of KEs, showing how KEs can be easily adapted for a variety of malicious purposes. Third, we highlight vulnerabilities in the AI ecosystem that allow unrestricted uploading and downloading of updated models without verification. Fourth, we argue that a lack of social and institutional awareness exacerbates this risk, and discuss the implications for different stakeholders. We call on the community to (i) research tamper-resistant models and countermeasures against malicious model editing, and (ii) actively engage in securing the AI ecosystem.
Authors: Bowen Jin, Jinsung Yoon, Zhen Qin, Ziqi Wang, Wei Xiong, Yu Meng, Jiawei Han, Sercan O. Arik
Abstract: Large Language Models (LLMs) have revolutionized artificial intelligence with capabilities in reasoning, coding, and communication, driving innovation across industries. Their true potential depends on effective alignment to ensure correct, trustworthy and ethical behavior, addressing challenges like misinformation, hallucinations, bias and misuse. While existing Reinforcement Learning (RL)-based alignment methods are notoriously complex, direct optimization approaches offer a simpler alternative. In this work, we introduce a novel direct optimization approach for LLM alignment by drawing on established Information Retrieval (IR) principles. We present a systematic framework that bridges LLM alignment and IR methodologies, mapping LLM generation and reward models to IR's retriever-reranker paradigm. Building on this foundation, we propose LLM Alignment as Retriever Preference Optimization (LarPO), a new alignment method that enhances overall alignment quality. Extensive experiments validate LarPO's effectiveness with 38.9 % and 13.7 % averaged improvement on AlpacaEval2 and MixEval-Hard respectively. Our work opens new avenues for advancing LLM alignment by integrating IR foundations, offering a promising direction for future research.
Authors: Simone Clemente, Zied Ben Houidi, Alexis Huet, Dario Rossi, Giulio Franzese, Pietro Michiardi
Abstract: Through systematic empirical investigation, we uncover a fundamental and concerning property of Large Language Models: while they can safely learn facts that don't contradict their knowledge, attempting to update facts with contradictory information triggers catastrophic corruption of unrelated knowledge. Unlike humans, who naturally resist contradictory information, these models indiscriminately accept contradictions, leading to devastating interference, destroying up to 80% of unrelated knowledge even when learning as few as 10-100 contradicting facts. To understand whether this interference could be mitigated through selective plasticity, we experiment with targeted network updates, distinguishing between previously used (stubborn) and rarely used (plastic) neurons. We uncover another asymmetry: while sparing frequently-used neurons significantly improves retention of existing knowledge for non-contradictory updates (98% vs 93% with standard updates), contradictory updates trigger catastrophic interference regardless of targeting strategy. This effect which persists across tested model scales (GPT-2 to GPT-J-6B), suggests a fundamental limitation in how neural networks handle contradictions. Finally, we demonstrate that contradictory information can be reliably detected (95%+ accuracy) using simple model features, offering a potential protective mechanism. These findings motivate new architectures that can, like humans, naturally resist contradictions rather than allowing destructive overwrites.
Authors: Nadav Timor, Jonathan Mamou, Daniel Korat, Moshe Berchansky, Gaurav Jain, Oren Pereg, Moshe Wasserblat, David Harel
Abstract: Accelerating the inference of large language models (LLMs) is a critical challenge in generative AI. Speculative decoding (SD) methods offer substantial efficiency gains by generating multiple tokens using a single target forward pass. However, existing SD approaches require the drafter and target models to share the same vocabulary, thus limiting the pool of possible drafters, often necessitating the training of a drafter from scratch. We present three new SD methods that remove this shared-vocabulary constraint. All three methods preserve the target distribution (i.e., they are lossless) and work with off-the-shelf models without requiring additional training or modifications. Empirically, on summarization, programming, and long-context tasks, our algorithms demonstrate significant speedups of up to 2.8x over standard autoregressive decoding. By enabling any off-the-shelf model to serve as a drafter and requiring no retraining, this work substantially broadens the applicability of the SD framework in practice.
Authors: Wenlong Meng, Zhenyuan Guo, Lenan Wu, Chen Gong, Wenyan Liu, Weixian Li, Chengkun Wei, Wenzhi Chen
Abstract: Large Language Models (LLMs) pose significant privacy risks, potentially leaking training data due to implicit memorization. Existing privacy attacks primarily focus on membership inference attacks (MIAs) or data extraction attacks, but reconstructing specific personally identifiable information (PII) in LLMs' training data remains challenging. In this paper, we propose R.R. (Recollect and Rank), a novel two-step privacy stealing attack that enables attackers to reconstruct PII entities from scrubbed training data where the PII entities have been masked. In the first stage, we introduce a prompt paradigm named recollection, which instructs the LLM to repeat a masked text but fill in masks. Then we can use PII identifiers to extract recollected PII candidates. In the second stage, we design a new criterion to score each PII candidate and rank them. Motivated by membership inference, we leverage the reference model as a calibration to our criterion. Experiments across three popular PII datasets demonstrate that the R.R. achieves better PII identification performance than baselines. These results highlight the vulnerability of LLMs to PII leakage even when training data has been scrubbed. We release our code and datasets at GitHub.
Authors: Lionel Wong, Ayman Ali, Raymond Xiong, Shannon Zeijang Shen, Yoon Kim, Monica Agrawal
Abstract: Patients have long sought health information online, and increasingly, they are turning to generative AI to answer their health-related queries. Given the high stakes of the medical domain, techniques like retrieval-augmented generation and citation grounding have been widely promoted as methods to reduce hallucinations and improve the accuracy of AI-generated responses and have been widely adopted into search engines. This paper argues that even when these methods produce literally accurate content drawn from source documents sans hallucinations, they can still be highly misleading. Patients may derive significantly different interpretations from AI-generated outputs than they would from reading the original source material, let alone consulting a knowledgeable clinician. Through a large-scale query analysis on topics including disputed diagnoses and procedure safety, we support our argument with quantitative and qualitative evidence of the suboptimal answers resulting from current systems. In particular, we highlight how these models tend to decontextualize facts, omit critical relevant sources, and reinforce patient misconceptions or biases. We propose a series of recommendations -- such as the incorporation of communication pragmatics and enhanced comprehension of source documents -- that could help mitigate these issues and extend beyond the medical domain.
Authors: Mengqiao Liu, Tevin Wang, Cassandra A. Cohen, Sarah Li, Chenyan Xiong
Abstract: Which large language model (LLM) is better? Every evaluation tells a story, but what do users really think about current LLMs? This paper presents CLUE, an LLM-powered interviewer that conducts in-the-moment user experience interviews, right after users interact with LLMs, and automatically gathers insights about user opinions from massive interview logs. We conduct a study with thousands of users to understand user opinions on mainstream LLMs, recruiting users to first chat with a target LLM and then be interviewed by CLUE. Our experiments demonstrate that CLUE captures interesting user opinions, e.g., the bipolar views on the displayed reasoning process of DeepSeek-R1 and demands for information freshness and multi-modality. Our code and data are at https://github.com/cxcscmu/LLM-Interviewer.
Authors: Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, Se-Young Yun
Abstract: Chain-of-thought (CoT) reasoning has enabled large language models (LLMs) to utilize additional computation through intermediate tokens to solve complex tasks. However, we posit that typical reasoning traces contain many redundant tokens, incurring extraneous inference costs. Upon examination of the output distribution of current LLMs, we find evidence on their latent ability to reason more concisely, relative to their default behavior. To elicit this capability, we propose simple fine-tuning methods which leverage self-generated concise reasoning paths obtained by best-of-N sampling and few-shot conditioning, in task-specific settings. Our combined method achieves a 30% reduction in output tokens on average, across five model families on GSM8K and MATH, while maintaining average accuracy. By exploiting the fundamental stochasticity and in-context learning capabilities of LLMs, our self-training approach robustly elicits concise reasoning on a wide range of models, including those with extensive post-training. Code is available at https://github.com/TergelMunkhbat/concise-reasoning
Authors: Jacqueline R. M. A. Maasch, Alihan H\"uy\"uk, Xinnuo Xu, Aditya V. Nori, Javier Gonzalez
Abstract: Causal reasoning and compositional reasoning are two core aspirations in AI. Measuring the extent of these behaviors requires principled evaluation methods. We explore a unified perspective that considers both behaviors simultaneously, termed compositional causal reasoning (CCR): the ability to infer how causal measures compose and, equivalently, how causal quantities propagate through graphs. We instantiate a framework for the systematic evaluation of CCR for the average treatment effect and the probability of necessity and sufficiency. As proof of concept, we demonstrate CCR evaluation for language models in the LLama, Phi, and GPT families. On a math word problem, our framework revealed a range of taxonomically distinct error patterns. CCR errors increased with the complexity of causal paths for all models except o1.
Authors: Wenjie Qiu, Yi-Chen Li, Xuqin Zhang, Tianyi Zhang, Yihang Zhang, Zongzhang Zhang, Yang Yu
Abstract: Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models operate at a coarse-grained level, requiring the generation of a complete response to obtain a reward value. The sparse reward may present challenges for downstream reinforcement learning. While recent efforts have attempted to learn token-level reward models, the lack of explicit semantic information makes it difficult to model the credit of every individual token. In this paper, we propose assigning scores to every sentence, introducing an intermediate-grained reward model. By segmenting the complete response into sentences and applying differential operations to reward output at the start and end positions of each sentence, we can effectively model the rewards of sentences. Moreover, a novel attention mechanism is introduced to aggregate the scores of all sentences into a response-level score, which allows it to be trained using the Bradley-Terry model. On common benchmarks, our method outperforms the response-level reward model by 2.7% on RewardBench (for reward modeling evaluation) and surpasses all baselines on AlpacaEval (for alignment evaluation).
Authors: Nevidu Jayatilleke, Ruvan Weerasinghe
Abstract: Automatic patent summarization approaches that help in the patent analysis and comprehension procedure are in high demand due to the colossal growth of innovations. The development of natural language processing (NLP), text mining, and deep learning has notably amplified the efficacy of text summarization models for abundant types of documents. Summarizing patent text remains a pertinent challenge due to the labyrinthine writing style of these documents, which includes technical and legal intricacies. Additionally, these patent document contents are considerably lengthier than archetypal documents, which complicates the process of extracting pertinent information for summarization. Embodying extractive and abstractive text summarization methodologies into a hybrid framework, this study proposes a system for efficiently creating abstractive summaries of patent records. The procedure involves leveraging the LexRank graph-based algorithm to retrieve the important sentences from input parent texts, then utilizing a Bidirectional Auto-Regressive Transformer (BART) model that has been fine-tuned using Low-Ranking Adaptation (LoRA) for producing text summaries. This is accompanied by methodical testing and evaluation strategies. Furthermore, the author employed certain meta-learning techniques to achieve Domain Generalization (DG) of the abstractive component across multiple patent fields.
Authors: Chatrine Qwaider, Bashar Alhafni, Kirill Chirkunov, Nizar Habash, Ted Briscoe
Abstract: Automated Essay Scoring (AES) plays a crucial role in assessing language learners' writing quality, reducing grading workload, and providing real-time feedback. The lack of annotated essay datasets inhibits the development of Arabic AES systems. This paper leverages Large Language Models (LLMs) and Transformer models to generate synthetic Arabic essays for AES. We prompt an LLM to generate essays across the Common European Framework of Reference (CEFR) proficiency levels and introduce and compare two approaches to error injection. We create a dataset of 3,040 annotated essays with errors injected using our two methods. Additionally, we develop a BERT-based Arabic AES system calibrated to CEFR levels. Our experimental results demonstrate the effectiveness of our synthetic dataset in improving Arabic AES performance. We make our code and data publicly available.
Authors: Fabian Retkowski, Maike Z\"ufle, Andreas Sudmann, Dinah Pfau, Shinji Watanabe, Jan Niehues, Alexander Waibel
Abstract: Speech summarization has become an essential tool for efficiently managing and accessing the growing volume of spoken and audiovisual content. However, despite its increasing importance, speech summarization remains loosely defined. The field intersects with several research areas, including speech recognition, text summarization, and specific applications like meeting summarization. This survey not only examines existing datasets and evaluation protocols, which are crucial for assessing the quality of summarization approaches, but also synthesizes recent developments in the field, highlighting the shift from traditional systems to advanced models like fine-tuned cascaded architectures and end-to-end solutions. In doing so, we surface the ongoing challenges, such as the need for realistic evaluation benchmarks, multilingual datasets, and long-context handling.
Authors: Mohamed Gado, Towhid Taliee, Muhammad Memon, Dmitry Ignatov, Radu Timofte
Abstract: Visual storytelling is an interdisciplinary field combining computer vision and natural language processing to generate cohesive narratives from sequences of images. This paper presents a novel approach that leverages recent advancements in multimodal models, specifically adapting transformer-based architectures and large multimodal models, for the visual storytelling task. Leveraging the large-scale Visual Storytelling (VIST) dataset, our VIST-GPT model produces visually grounded, contextually appropriate narratives. We address the limitations of traditional evaluation metrics, such as BLEU, METEOR, ROUGE, and CIDEr, which are not suitable for this task. Instead, we utilize RoViST and GROOVIST, novel reference-free metrics designed to assess visual storytelling, focusing on visual grounding, coherence, and non-redundancy. These metrics provide a more nuanced evaluation of narrative quality, aligning closely with human judgment.
Authors: Zhiting Fan, Ruizhe Chen, Zuozhu Liu
Abstract: Identifying bias in LLM-generated content is a crucial prerequisite for ensuring fairness in LLMs. Existing methods, such as fairness classifiers and LLM-based judges, face limitations related to difficulties in understanding underlying intentions and the lack of criteria for fairness judgment. In this paper, we introduce BiasGuard, a novel bias detection tool that explicitly analyzes inputs and reasons through fairness specifications to provide accurate judgments. BiasGuard is implemented through a two-stage approach: the first stage initializes the model to explicitly reason based on fairness specifications, while the second stage leverages reinforcement learning to enhance its reasoning and judgment capabilities. Our experiments, conducted across five datasets, demonstrate that BiasGuard outperforms existing tools, improving accuracy and reducing over-fairness misjudgments. We also highlight the importance of reasoning-enhanced decision-making and provide evidence for the effectiveness of our two-stage optimization pipeline.
Authors: Jongwook Han, Dongmin Choi, Woojung Song, Eun-Ju Lee, Yohan Jo
Abstract: The importance of benchmarks for assessing the values of language models has been pronounced due to the growing need of more authentic, human-aligned responses. However, existing benchmarks rely on human or machine annotations that are vulnerable to value-related biases. Furthermore, the tested scenarios often diverge from real-world contexts in which models are commonly used to generate text and express values. To address these issues, we propose the Value Portrait benchmark, a reliable framework for evaluating LLMs' value orientations with two key characteristics. First, the benchmark consists of items that capture real-life user-LLM interactions, enhancing the relevance of assessment results to real-world LLM usage. Second, each item is rated by human subjects based on its similarity to their own thoughts, and correlations between these ratings and the subjects' actual value scores are derived. This psychometrically validated approach ensures that items strongly correlated with specific values serve as reliable items for assessing those values. Through evaluating 44 LLMs with our benchmark, we find that these models prioritize Benevolence, Security, and Self-Direction values while placing less emphasis on Tradition, Power, and Achievement values. Also, our analysis reveals biases in how LLMs perceive various demographic groups, deviating from real human data.
Authors: Matan Orbach, Ohad Eytan, Benjamin Sznajder, Ariel Gera, Odellia Boni, Yoav Kantor, Gal Bloch, Omri Levy, Hadas Abraham, Nitzan Barzilay, Eyal Shnarch, Michael E. Factor, Shila Ofek-Koifman, Paula Ta-Shma, Assaf Toledo
Abstract: Finding the optimal Retrieval-Augmented Generation (RAG) configuration for a given use case can be complex and expensive. Motivated by this challenge, frameworks for RAG hyper-parameter optimization (HPO) have recently emerged, yet their effectiveness has not been rigorously benchmarked. To address this gap, we present a comprehensive study involving 5 HPO algorithms over 5 datasets from diverse domains, including a new one collected for this work on real-world product documentation. Our study explores the largest HPO search space considered to date, with three evaluation metrics as optimization targets. Analysis of the results shows that RAG HPO can be done efficiently, either greedily or with random search, and that it significantly boosts RAG performance for all datasets. For greedy HPO approaches, we show that optimizing model selection first is preferable to the prevalent practice of optimizing according to RAG pipeline order.
Authors: Ruilin Liu, Zhixiao Zhao, Jieqiong Li, Chang Liu, Dongbo Wang
Abstract: The rapid development of large language models (LLMs) has provided significant support and opportunities for the advancement of domain-specific LLMs. However, fine-tuning these large models using Intangible Cultural Heritage (ICH) data inevitably faces challenges such as bias, incorrect knowledge inheritance, and catastrophic forgetting. To address these issues, we propose a novel training method that integrates a bidirectional chains of thought and a reward mechanism. This method is built upon ICH-Qwen, a large language model specifically designed for the field of intangible cultural heritage. The proposed method enables the model to not only perform forward reasoning but also enhances the accuracy of the generated answers by utilizing reverse questioning and reverse reasoning to activate the model's latent knowledge. Additionally, a reward mechanism is introduced during training to optimize the decision-making process. This mechanism improves the quality of the model's outputs through structural and content evaluations with different weighting schemes. We conduct comparative experiments on ICH-Qwen, with results demonstrating that our method outperforms 0-shot, step-by-step reasoning, knowledge distillation, and question augmentation methods in terms of accuracy, Bleu-4, and Rouge-L scores on the question-answering task. Furthermore, the paper highlights the effectiveness of combining the bidirectional chains of thought and reward mechanism through ablation experiments. In addition, a series of generalizability experiments are conducted, with results showing that the proposed method yields improvements on various domain-specific datasets and advanced models in areas such as Finance, Wikidata, and StrategyQA. This demonstrates that the method is adaptable to multiple domains and provides a valuable approach for model training in future applications across diverse fields.
Authors: Yuhang Zhou, Jing Zhu, Shengyi Qian, Zhuokai Zhao, Xiyao Wang, Xiaoyu Liu, Ming Li, Paiheng Xu, Wei Ai, Furong Huang
Abstract: Large Language Models (LLMs) are increasingly aligned with human preferences through Reinforcement Learning from Human Feedback (RLHF). Among RLHF methods, Group Relative Policy Optimization (GRPO) has gained attention for its simplicity and strong performance, notably eliminating the need for a learned value function. However, GRPO implicitly assumes a balanced domain distribution and uniform semantic alignment across groups - assumptions that rarely hold in real-world datasets. When applied to multi-domain, imbalanced data, GRPO disproportionately optimizes for dominant domains, neglecting underrepresented ones and resulting in poor generalization and fairness. We propose Domain-Informed Self-Consistency Policy Optimization (DISCO), a principled extension to GRPO that addresses inter-group imbalance with two key innovations. Domain-aware reward scaling counteracts frequency bias by reweighting optimization based on domain prevalence. Difficulty-aware reward scaling leverages prompt-level self-consistency to identify and prioritize uncertain prompts that offer greater learning value. Together, these strategies promote more equitable and effective policy learning across domains. Extensive experiments across multiple LLMs and skewed training distributions show that DISCO improves generalization, outperforms existing GRPO variants by 5% on Qwen3 models, and sets new state-of-the-art results on multi-domain alignment benchmarks.
Authors: Tong Zheng, Lichang Chen, Simeng Han, R. Thomas McCoy, Heng Huang
Abstract: Human beings naturally utilize multiple reasoning modalities to learn and solve logical problems, i.e., different representational formats such as natural language, code, and symbolic logic. In contrast, most existing LLM-based approaches operate with a single reasoning modality during training, typically natural language. Although some methods explored modality selection or augmentation at inference time, the training process remains modality-blind, limiting synergy among modalities. To fill in this gap, we propose Mixture-of-Thought (MoT), a framework that enables LLMs to reason across three complementary modalities: natural language, code, and a newly introduced symbolic modality, truth-table, which systematically enumerates logical cases and partially mitigates key failure modes in natural language reasoning. MoT adopts a two-phase design: (1) self-evolving MoT training, which jointly learns from filtered, self-generated rationales across modalities; and (2) MoT inference, which fully leverages the synergy of three modalities to produce better predictions. Experiments on logical reasoning benchmarks including FOLIO and ProofWriter demonstrate that our MoT framework consistently and significantly outperforms strong LLM baselines with single-modality chain-of-thought approaches, achieving up to +11.7pp average accuracy gain. Further analyses show that our MoT framework benefits both training and inference stages; that it is particularly effective on harder logical reasoning problems; and that different modalities contribute complementary strengths, with truth-table reasoning helping to overcome key bottlenecks in natural language inference.
Authors: Gouki Minegishi, Hiroki Furuta, Shohei Taniguchi, Yusuke Iwasawa, Yutaka Matsuo
Abstract: Transformer-based language models exhibit In-Context Learning (ICL), where predictions are made adaptively based on context. While prior work links induction heads to ICL through a sudden jump in accuracy, this can only account for ICL when the answer is included within the context. However, an important property of practical ICL in large language models is the ability to meta-learn how to solve tasks from context, rather than just copying answers from context; how such an ability is obtained during training is largely unexplored. In this paper, we experimentally clarify how such meta-learning ability is acquired by analyzing the dynamics of the model's circuit during training. Specifically, we extend the copy task from previous research into an In-Context Meta Learning setting, where models must infer a task from examples to answer queries. Interestingly, in this setting, we find that there are multiple phases in the process of acquiring such abilities, and that a unique circuit emerges in each phase, contrasting with the single-phases change in induction heads. The emergence of such circuits can be related to several phenomena known in large language models, and our analysis lead to a deeper understanding of the source of the transformer's ICL ability.
Authors: Xinlong Chen, Yuanxing Zhang, Qiang Liu, Junfei Wu, Fuzheng Zhang, Tieniu Tan
Abstract: Large Vision-Language Models (LVLMs) have exhibited impressive capabilities across various visual tasks, yet they remain hindered by the persistent challenge of hallucinations. To address this critical issue, we propose Mixture of Decoding (MoD), a novel approach for hallucination mitigation that dynamically adapts decoding strategies by evaluating the correctness of the model's attention on image tokens. Specifically, MoD measures the consistency between outputs generated from the original image tokens and those derived from the model's attended image tokens, to distinguish the correctness aforementioned. If the outputs are consistent, indicating correct attention, MoD employs a complementary strategy to amplify critical information. Conversely, if the outputs are inconsistent, suggesting erroneous attention, MoD utilizes a contrastive strategy to suppress misleading information. Extensive experiments demonstrate that MoD significantly outperforms existing decoding methods across multiple mainstream benchmarks, effectively mitigating hallucinations in LVLMs. The code is available at https://github.com/xlchen0205/MoD.
Authors: Subrata Biswas, Mohammad Nur Hossain Khan, Bashima Islam
Abstract: Multimodal question answering (QA) often requires identifying which video, audio, or sensor tokens are relevant to the question. Yet modality disagreements are common: off-camera speech, background noise, or motion outside the field of view often mislead fusion models that weight all streams equally. We present RAVEN, a unified QA architecture whose core is QuART, a query-conditioned cross-modal gating module that assigns scalar relevance scores to each token across modalities, enabling the model to amplify informative signals and suppress distractors before fusion. RAVEN is trained through a three-stage pipeline comprising unimodal pretraining, query-aligned fusion, and disagreement-oriented fine-tuning -- each stage targeting a distinct challenge in multi-modal reasoning: representation quality, cross-modal relevance, and robustness to modality mismatch. To support training and evaluation, we release AVS-QA, a dataset of 300K synchronized Audio--Video-Sensor streams paired with automatically generated question-answer pairs. Experimental results on seven multi-modal QA benchmarks -- including egocentric and exocentric tasks -- show that RAVEN achieves up to 14.5\% and 8.0\% gains in accuracy compared to state-of-the-art multi-modal large language models, respectively. Incorporating sensor data provides an additional 16.4\% boost, and the model remains robust under modality corruption, outperforming SOTA baselines by 50.23\%. Our code and dataset are available at https://github.com/BASHLab/RAVEN.
Authors: Lei Zhang, Markus Stricker
Abstract: The discovery and optimization of materials for specific applications is hampered by the practically infinite number of possible elemental combinations and associated properties, also known as the `combinatorial explosion'. By nature of the problem, data are scarce and all possible data sources should be used. In addition to simulations and experimental results, the latent knowledge in scientific texts is not yet used to its full potential. We present an iterative framework that refines a given scientific corpus by strategic selection of the most diverse documents, training Word2Vec models, and monitoring the convergence of composition-property correlations in embedding space. Our approach is applied to predict high-performing materials for oxygen reduction (ORR), hydrogen evolution (HER), and oxygen evolution (OER) reactions for a large number of possible candidate compositions. Our method successfully predicts the highest performing compositions among a large pool of candidates, validated by experimental measurements of the electrocatalytic performance in the lab. This work demonstrates and validates the potential of iterative corpus refinement to accelerate materials discovery and optimization, offering a scalable and efficient tool for screening large compositional spaces where reliable data are scarce or non-existent.
Authors: Shuhai Zhang, Zeng You, Yaofo Chen, Zhiquan Wen, Qianyue Wang, Zhijie Qiu, Yuanqing Li, Mingkui Tan
Abstract: Transformer-based large language models (LLMs) excel in natural language processing tasks by capturing long-range dependencies through self-attention mechanisms. However, long-context modeling faces significant computational inefficiencies due to \textit{redundant} attention computations: while attention weights are often \textit{sparse}, all tokens consume \textit{equal} computational resources. In this paper, we reformulate traditional probabilistic sequence modeling as a \textit{supervised learning task}, enabling the separation of relevant and irrelevant tokens and providing a clearer understanding of redundancy. Based on this reformulation, we theoretically analyze attention sparsity, revealing that only a few tokens significantly contribute to predictions. Building on this, we formulate attention optimization as a linear coding problem and propose a \textit{group coding strategy}, theoretically showing its ability to improve robustness against random noise and enhance learning efficiency. Motivated by this, we propose \textit{Dynamic Group Attention} (DGA), which leverages the group coding to explicitly reduce redundancy by aggregating less important tokens during attention computation. Empirical results show that our DGA significantly reduces computational costs while maintaining competitive performance.Code is available at https://github.com/bolixinyu/DynamicGroupAttention.
Authors: Cihan Xiao, Ruixing Liang, Xiangyu Zhang, Mehmet Emre Tiryaki, Veronica Bae, Lavanya Shankar, Rong Yang, Ethan Poon, Emmanuel Dupoux, Sanjeev Khudanpur, Leibny Paola Garcia Perera
Abstract: The success of large language models has driven interest in developing similar speech processing capabilities. However, a key challenge is the scarcity of high-quality spontaneous speech data, as most existing datasets contain scripted dialogues. To address this, we present a novel pipeline for eliciting and recording natural dialogues and release our dataset with 100+ hours of spontaneous speech. Our approach fosters fluid, natural conversations while encouraging a diverse range of topics and interactive exchanges. Unlike traditional methods, it facilitates genuine interactions, providing a reproducible framework for future data collection. This paper introduces our dataset and methodology, laying the groundwork for addressing the shortage of spontaneous speech data. We plan to expand this dataset in future stages, offering a growing resource for the research community.
Authors: Ming Wang, Peidong Wang, Lin Wu, Xiaocui Yang, Daling Wang, Shi Feng, Yuxin Chen, Bixuan Wang, Yifei Zhang
Abstract: Constrained by the cost and ethical concerns of involving real seekers in AI-driven mental health, researchers develop LLM-based conversational agents (CAs) with tailored configurations, such as profiles, symptoms, and scenarios, to simulate seekers. While these efforts advance AI in mental health, achieving more realistic seeker simulation remains hindered by two key challenges: dynamic evolution and multi-session memory. Seekers' mental states often fluctuate during counseling, which typically spans multiple sessions. To address this, we propose AnnaAgent, an emotional and cognitive dynamic agent system equipped with tertiary memory. AnnaAgent incorporates an emotion modulator and a complaint elicitor trained on real counseling dialogues, enabling dynamic control of the simulator's configurations. Additionally, its tertiary memory mechanism effectively integrates short-term and long-term memory across sessions. Evaluation results, both automated and manual, demonstrate that AnnaAgent achieves more realistic seeker simulation in psychological counseling compared to existing baselines. The ethically reviewed and screened code can be found on https://github.com/sci-m-wang/AnnaAgent.
Authors: Chiyu Zhang, Marc-Alexandre Cote, Michael Albada, Anush Sankaran, Jack W. Stokes, Tong Wang, Amir Abdi, William Blum, Muhammad Abdul-Mageed
Abstract: Large language model (LLM) agents have shown impressive capabilities in human language comprehension and reasoning, yet their potential in cybersecurity remains underexplored. We introduce DefenderBench, a practical, open-source toolkit for evaluating language agents across offense, defense, and cybersecurity knowledge-based tasks. DefenderBench includes environments for network intrusion, malicious content detection, code vulnerability analysis, and cybersecurity knowledge assessment. It is intentionally designed to be affordable and easily accessible for researchers while providing fair and rigorous assessment. We benchmark several state-of-the-art (SoTA) and popular LLMs, including both open- and closed-weight models, using a standardized agentic framework. Our results show that Claude-3.7-sonnet performs best with a DefenderBench score of 81.65, followed by Claude-3.7-sonnet-think with 78.40, while the best open-weight model, Llama 3.3 70B, is not far behind with a DefenderBench score of 71.81. DefenderBench's modular design allows seamless integration of custom LLMs and tasks, promoting reproducibility and fair comparisons. An anonymized version of DefenderBench is available at https://github.com/microsoft/DefenderBench.
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: 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: Wenhao Li, Wenwu Li, Chuyun Shen, Junjie Sheng, Zixiao Huang, Di Wu, Yun Hua, Wei Yin, Xiangfeng Wang, Hongyuan Zha, Bo Jin
Abstract: We present TextAtari, a benchmark for evaluating language agents on very long-horizon decision-making tasks spanning up to 100,000 steps. By translating the visual state representations of classic Atari games into rich textual descriptions, TextAtari creates a challenging test bed that bridges sequential decision-making with natural language processing. The benchmark includes nearly 100 distinct tasks with varying complexity, action spaces, and planning horizons, all rendered as text through an unsupervised representation learning framework (AtariARI). We evaluate three open-source large language models (Qwen2.5-7B, Gemma-7B, and Llama3.1-8B) across three agent frameworks (zero-shot, few-shot chain-of-thought, and reflection reasoning) to assess how different forms of prior knowledge affect performance on these long-horizon challenges. Four scenarios-Basic, Obscured, Manual Augmentation, and Reference-based-investigate the impact of semantic understanding, instruction comprehension, and expert demonstrations on agent decision-making. Our results reveal significant performance gaps between language agents and human players in extensive planning tasks, highlighting challenges in sequential reasoning, state tracking, and strategic planning across tens of thousands of steps. TextAtari provides standardized evaluation protocols, baseline implementations, and a framework for advancing research at the intersection of language models and planning. Our code is available at https://github.com/Lww007/Text-Atari-Agents.
Authors: Matthieu Tehenan, Vikram Natarajan, Jonathan Michala, Milton Lin, Juri Opitz
Abstract: Sentence embeddings are central to modern NLP and AI systems, yet little is known about their internal structure. While we can compare these embeddings using measures such as cosine similarity, the contributing features are not human-interpretable, and the content of an embedding seems untraceable, as it is masked by complex neural transformations and a final pooling operation that combines individual token embeddings. To alleviate this issue, we propose a new method to mechanistically decompose sentence embeddings into interpretable components, by using dictionary learning on token-level representations. We analyze how pooling compresses these features into sentence representations, and assess the latent features that reside in a sentence embedding. This bridges token-level mechanistic interpretability with sentence-level analysis, making for more transparent and controllable representations. In our studies, we obtain several interesting insights into the inner workings of sentence embedding spaces, for instance, that many semantic and syntactic aspects are linearly encoded in the embeddings.
Authors: Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, Jingren Zhou
Abstract: In this work, we introduce the Qwen3 Embedding series, a significant advancement over its predecessor, the GTE-Qwen series, in text embedding and reranking capabilities, built upon the Qwen3 foundation models. Leveraging the Qwen3 LLMs' robust capabilities in multilingual text understanding and generation, our innovative multi-stage training pipeline combines large-scale unsupervised pre-training with supervised fine-tuning on high-quality datasets. Effective model merging strategies further ensure the robustness and adaptability of the Qwen3 Embedding series. During the training process, the Qwen3 LLMs serve not only as backbone models but also play a crucial role in synthesizing high-quality, rich, and diverse training data across multiple domains and languages, thus enhancing the training pipeline. The Qwen3 Embedding series offers a spectrum of model sizes (0.6B, 4B, 8B) for both embedding and reranking tasks, addressing diverse deployment scenarios where users can optimize for either efficiency or effectiveness. Empirical evaluations demonstrate that the Qwen3 Embedding series achieves state-of-the-art results across diverse benchmarks. Notably, it excels on the multilingual evaluation benchmark MTEB for text embedding, as well as in various retrieval tasks, including code retrieval, cross-lingual retrieval and multilingual retrieval. To facilitate reproducibility and promote community-driven research and development, the Qwen3 Embedding models are publicly available under the Apache 2.0 license.
Authors: Pengyi Li, Matvey Skripkin, Alexander Zubrey, Andrey Kuznetsov, Ivan Oseledets
Abstract: Large language models (LLMs) excel at reasoning, yet post-training remains critical for aligning their behavior with task goals. Existing reinforcement learning (RL) methods often depend on costly human annotations or external reward models. We propose Reinforcement Learning via Self-Confidence (RLSC), which uses the model's own confidence as reward signals-eliminating the need for labels, preference models, or reward engineering. Applied to Qwen2.5-Math-7B with only 16 samples per question and 10 or 20 training steps, RLSC improves accuracy by +13.4% on AIME2024, +21.2% on MATH500, +21.7% on Minerva Math, +20.8% on Olympiadbench, and +9.7% on AMC23. RLSC provides a simple, scalable post-training method for inference models, requiring only a small number of samples and unlabelled supervision.
Authors: LASA Team, Weiwen Xu, Hou Pong Chan, Long Li, Mahani Aljunied, Ruifeng Yuan, Jianyu Wang, Chenghao Xiao, Guizhen Chen, Chaoqun Liu, Zhaodonghui Li, Yu Sun, Junao Shen, Chaojun Wang, Jie Tan, Deli Zhao, Tingyang Xu, Hao Zhang, Yu Rong
Abstract: Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities in understanding common visual elements, largely due to their large-scale datasets and advanced training strategies. However, their effectiveness in medical applications remains limited due to the inherent discrepancies between data and tasks in medical scenarios and those in the general domain. Concretely, existing medical MLLMs face the following critical limitations: (1) limited coverage of medical knowledge beyond imaging, (2) heightened susceptibility to hallucinations due to suboptimal data curation processes, (3) lack of reasoning capabilities tailored for complex medical scenarios. To address these challenges, we first propose a comprehensive data curation procedure that (1) efficiently acquires rich medical knowledge data not only from medical imaging but also from extensive medical texts and general-domain data; and (2) synthesizes accurate medical captions, visual question answering (VQA), and reasoning samples. As a result, we build a multimodal dataset enriched with extensive medical knowledge. Building on the curated data, we introduce our medical-specialized MLLM: Lingshu. Lingshu undergoes multi-stage training to embed medical expertise and enhance its task-solving capabilities progressively. Besides, we preliminarily explore the potential of applying reinforcement learning with verifiable rewards paradigm to enhance Lingshu's medical reasoning ability. Additionally, we develop MedEvalKit, a unified evaluation framework that consolidates leading multimodal and textual medical benchmarks for standardized, fair, and efficient model assessment. We evaluate the performance of Lingshu on three fundamental medical tasks, multimodal QA, text-based QA, and medical report generation. The results show that Lingshu consistently outperforms the existing open-source multimodal models on most tasks ...
Authors: Prarabdh Shukla, Wei Yin Chong, Yash Patel, Brennan Schaffner, Danish Pruthi, Arjun Bhagoji
Abstract: To meet the demands of content moderation, online platforms have resorted to automated systems. Newer forms of real-time engagement($\textit{e.g.}$, users commenting on live streams) on platforms like Twitch exert additional pressures on the latency expected of such moderation systems. Despite their prevalence, relatively little is known about the effectiveness of these systems. In this paper, we conduct an audit of Twitch's automated moderation tool ($\texttt{AutoMod}$) to investigate its effectiveness in flagging hateful content. For our audit, we create streaming accounts to act as siloed test beds, and interface with the live chat using Twitch's APIs to send over $107,000$ comments collated from $4$ datasets. We measure $\texttt{AutoMod}$'s accuracy in flagging blatantly hateful content containing misogyny, racism, ableism and homophobia. Our experiments reveal that a large fraction of hateful messages, up to $94\%$ on some datasets, $\textit{bypass moderation}$. Contextual addition of slurs to these messages results in $100\%$ removal, revealing $\texttt{AutoMod}$'s reliance on slurs as a moderation signal. We also find that contrary to Twitch's community guidelines, $\texttt{AutoMod}$ blocks up to $89.5\%$ of benign examples that use sensitive words in pedagogical or empowering contexts. Overall, our audit points to large gaps in $\texttt{AutoMod}$'s capabilities and underscores the importance for such systems to understand context effectively.
Authors: Peter Ochieng
Abstract: This work introduces UDPNet, a novel architecture designed to accelerate the reverse diffusion process in speech synthesis. Unlike traditional diffusion models that rely on timestep embeddings and shared network parameters, UDPNet unrolls the reverse diffusion process directly into the network architecture, with successive layers corresponding to equally spaced steps in the diffusion schedule. Each layer progressively refines the noisy input, culminating in a high-fidelity estimation of the original data, \(x_0\). Additionally, we redefine the learning target by predicting latent variables instead of the conventional \(x_0\) or noise \(\epsilon_0\). This shift addresses the common issue of large prediction errors in early denoising stages, effectively reducing speech distortion. Extensive evaluations on single- and multi-speaker datasets demonstrate that UDPNet consistently outperforms state-of-the-art methods in both quality and efficiency, while generalizing effectively to unseen speech. These results position UDPNet as a robust solution for real-time speech synthesis applications. Sample audio is available at https://onexpeters.github.io/UDPNet.
Authors: Yuqi Zhou, Sunhao Dai, Liang Pang, Gang Wang, Zhenhua Dong, Jun Xu, Ji-Rong Wen
Abstract: Recommender systems are essential for information access, allowing users to present their content for recommendation. With the rise of large language models (LLMs), AI-generated content (AIGC), primarily in the form of text, has become a central part of the content ecosystem. As AIGC becomes increasingly prevalent, it is important to understand how it affects the performance and dynamics of recommender systems. To this end, we construct an environment that incorporates AIGC to explore its short-term impact. The results from popular sequential recommendation models reveal that AIGC are ranked higher in the recommender system, reflecting the phenomenon of source bias. To further explore the long-term impact of AIGC, we introduce a feedback loop with realistic simulators. The results show that the model's preference for AIGC increases as the user clicks on AIGC rises and the model trains on simulated click data. This leads to two issues: In the short term, bias toward AIGC encourages LLM-based content creation, increasing AIGC content, and causing unfair traffic distribution. From a long-term perspective, our experiments also show that when AIGC dominates the content ecosystem after a feedback loop, it can lead to a decline in recommendation performance. To address these issues, we propose a debiasing method based on L1-loss optimization to maintain long-term content ecosystem balance. In a real-world environment with AIGC generated by mainstream LLMs, our method ensures a balance between AIGC and human-generated content in the ecosystem. The code and dataset are available at https://github.com/Yuqi-Zhou/Rec_SourceBias.
Authors: Jiacheng Du, Zhibo Wang, Jie Zhang, Xiaoyi Pang, Jiahui Hu, Kui Ren
Abstract: Language Models (LMs) are prone to ''memorizing'' training data, including substantial sensitive user information. To mitigate privacy risks and safeguard the right to be forgotten, machine unlearning has emerged as a promising approach for enabling LMs to efficiently ''forget'' specific texts. However, despite the good intentions, is textual unlearning really as effective and reliable as expected? To address the concern, we first propose Unlearning Likelihood Ratio Attack+ (U-LiRA+), a rigorous textual unlearning auditing method, and find that unlearned texts can still be detected with very high confidence after unlearning. Further, we conduct an in-depth investigation on the privacy risks of textual unlearning mechanisms in deployment and present the Textual Unlearning Leakage Attack (TULA), along with its variants in both black- and white-box scenarios. We show that textual unlearning mechanisms could instead reveal more about the unlearned texts, exposing them to significant membership inference and data reconstruction risks. Our findings highlight that existing textual unlearning actually gives a false sense of unlearning, underscoring the need for more robust and secure unlearning mechanisms.
Authors: Changde Du, Kaicheng Fu, Bincheng Wen, Yi Sun, Jie Peng, Wei Wei, Ying Gao, Shengpei Wang, Chuncheng Zhang, Jinpeng Li, Shuang Qiu, Le Chang, Huiguang He
Abstract: Understanding how humans conceptualize and categorize natural objects offers critical insights into perception and cognition. With the advent of Large Language Models (LLMs), a key question arises: can these models develop human-like object representations from linguistic and multimodal data? In this study, we combined behavioral and neuroimaging analyses to explore the relationship between object concept representations in LLMs and human cognition. We collected 4.7 million triplet judgments from LLMs and Multimodal LLMs (MLLMs) to derive low-dimensional embeddings that capture the similarity structure of 1,854 natural objects. The resulting 66-dimensional embeddings were stable, predictive, and exhibited semantic clustering similar to human mental representations. Remarkably, the dimensions underlying these embeddings were interpretable, suggesting that LLMs and MLLMs develop human-like conceptual representations of objects. Further analysis showed strong alignment between model embeddings and neural activity patterns in brain regions such as EBA, PPA, RSC, and FFA. This provides compelling evidence that the object representations in LLMs, while not identical to human ones, share fundamental similarities that reflect key aspects of human conceptual knowledge. Our findings advance the understanding of machine intelligence and inform the development of more human-like artificial cognitive systems.
Authors: Jerome Garnier-Brun, Marc M\'ezard, Emanuele Moscato, Luca Saglietti
Abstract: Understanding the learning process and the embedded computation in transformers is becoming a central goal for the development of interpretable AI. In the present study, we introduce a hierarchical filtering procedure for data models of sequences on trees, allowing us to hand-tune the range of positional correlations in the data. Leveraging this controlled setting, we provide evidence that vanilla encoder-only transformers can approximate the exact inference algorithm when trained on root classification and masked language modeling tasks, and study how this computation is discovered and implemented. We find that correlations at larger distances, corresponding to increasing layers of the hierarchy, are sequentially included by the network during training. By comparing attention maps from models trained with varying degrees of filtering and by probing the different encoder levels, we find clear evidence of a reconstruction of correlations on successive length scales corresponding to the various levels of the hierarchy, which we relate to a plausible implementation of the exact inference algorithm within the same architecture.
Authors: Zefang Liu, Yinzhu Quan
Abstract: Temporal point processes (TPPs) are widely used to model the timing and occurrence of events in domains such as social networks, transportation systems, and e-commerce. In this paper, we introduce TPP-LLM, a novel framework that integrates large language models (LLMs) with TPPs to capture both the semantic and temporal aspects of event sequences. Unlike traditional methods that rely on categorical event type representations, TPP-LLM directly utilizes the textual descriptions of event types, enabling the model to capture rich semantic information embedded in the text. While LLMs excel at understanding event semantics, they are less adept at capturing temporal patterns. To address this, TPP-LLM incorporates temporal embeddings and employs parameter-efficient fine-tuning (PEFT) methods to effectively learn temporal dynamics without extensive retraining. This approach improves both predictive accuracy and computational efficiency. Experimental results across diverse real-world datasets demonstrate that TPP-LLM outperforms state-of-the-art baselines in sequence modeling and event prediction, highlighting the benefits of combining LLMs with TPPs.
Authors: Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, Deva Ramanan
Abstract: Vision-language models (VLMs) have made significant progress in recent visual-question-answering (VQA) benchmarks that evaluate complex visio-linguistic reasoning. However, are these models truly effective? In this work, we show that VLMs still struggle with natural images and questions that humans can easily answer, which we term natural adversarial samples. We also find it surprisingly easy to generate these VQA samples from natural image-text corpora using off-the-shelf models like CLIP and ChatGPT. We propose a semi-automated approach to collect a new benchmark, NaturalBench, for reliably evaluating VLMs with 10,000 human-verified VQA samples. Crucially, we adopt a $\textbf{vision-centric}$ design by pairing each question with two images that yield different answers, preventing blind solutions from answering without using the images. This makes NaturalBench more challenging than previous benchmarks that can be solved with commonsense priors. We evaluate 53 state-of-the-art VLMs on NaturalBench, showing that models like LLaVA-OneVision, Cambrian-1, Llama3.2-Vision, Molmo, Qwen2-VL, and even GPT-4o lag 50%-70% behind human performance (over 90%). We analyze why NaturalBench is hard from two angles: (1) Compositionality: Solving NaturalBench requires diverse visio-linguistic skills, including understanding attribute bindings, object relationships, and advanced reasoning like logic and counting. To this end, unlike prior work that uses a single tag per sample, we tag each NaturalBench sample with 1 to 8 skill tags for fine-grained evaluation. (2) Biases: NaturalBench exposes severe biases in VLMs, as models often choose the same answer regardless of the image. Lastly, we apply our benchmark curation method to diverse data sources, including long captions (over 100 words) and non-English languages like Chinese and Hindi, highlighting its potential for dynamic evaluations of VLMs.
Authors: Michael S. Ryoo, Honglu Zhou, Shrikant Kendre, Can Qin, Le Xue, Manli Shu, Jongwoo Park, Kanchana Ranasinghe, Silvio Savarese, Ran Xu, Caiming Xiong, Juan Carlos Niebles
Abstract: We present xGen-MM-Vid (BLIP-3-Video): a multimodal language model for videos, particularly designed to efficiently capture temporal information over multiple frames. BLIP-3-Video takes advantage of the 'temporal encoder' in addition to the conventional visual tokenizer, which maps a sequence of tokens over multiple frames into a compact set of visual tokens. This enables BLIP3-Video to use much fewer visual tokens than its competing models (e.g., 32 vs. 4608 tokens). We explore different types of temporal encoders, including learnable spatio-temporal pooling as well as sequential models like Token Turing Machines. We experimentally confirm that BLIP-3-Video obtains video question-answering accuracies comparable to much larger state-of-the-art models (e.g., 34B), while being much smaller (i.e., 4B) and more efficient by using fewer visual tokens. The project website is at https://www.salesforceairesearch.com/opensource/xGen-MM-Vid/index.html
URLs: https://www.salesforceairesearch.com/opensource/xGen-MM-Vid/index.html
Authors: Peter Ochieng, Dennis Kaburu
Abstract: We present a prosody-guided framework for speech-to-speech translation (S2ST) that aligns and translates speech \emph{without} transcripts by leveraging cross-linguistic pause synchrony. Analyzing a 6{,}000-hour East African news corpus spanning five languages, we show that \emph{within-phylum} language pairs exhibit 30--40\% lower pause variance and over 3$\times$ higher onset/offset correlation compared to cross-phylum pairs. These findings motivate \textbf{SPaDA}, a dynamic-programming alignment algorithm that integrates silence consistency, rate synchrony, and semantic similarity. SPaDA improves alignment $F_1$ by +3--4 points and eliminates up to 38\% of spurious matches relative to greedy VAD baselines. Using SPaDA-aligned segments, we train \textbf{SegUniDiff}, a diffusion-based S2ST model guided by \emph{external gradients} from frozen semantic and speaker encoders. SegUniDiff matches an enhanced cascade in BLEU (30.3 on CVSS-C vs.\ 28.9 for UnitY), reduces speaker error rate (EER) from 12.5\% to 5.3\%, and runs at an RTF of 1.02. To support evaluation in low-resource settings, we also release a three-tier, transcript-free BLEU suite (M1--M3) that correlates strongly with human judgments. Together, our results show that prosodic cues in multilingual speech provide a reliable scaffold for scalable, non-autoregressive S2ST.
Authors: Ying Zhou, Xinyao Wang, Yulei Niu, Yaojie Shen, Lexin Tang, Fan Chen, Ben He, Le Sun, Longyin Wen
Abstract: Recent advancements in large language models (LLMs) have significantly enhanced their knowledge and generative capabilities, leading to a surge of interest in leveraging LLMs for high-quality data synthesis. However, synthetic data generation via prompting LLMs remains challenging due to LLMs' limited understanding of target data distributions and the complexity of prompt engineering, especially for structured formatted data. To address these issues, we introduce DiffLM, a controllable data synthesis framework based on variational autoencoder (VAE), which further (1) leverages diffusion models to reserve more information of original distribution and format structure in the learned latent distribution and (2) decouples the learning of target distribution knowledge from the LLM's generative objectives via a plug-and-play latent feature injection module. As we observed significant discrepancies between the VAE's latent representations and the real data distribution, the latent diffusion module is introduced into our framework to learn a fully expressive latent distribution. Evaluations on seven real-world datasets with structured formatted data (i.e., Tabular, Code, and Tool data) demonstrate that DiffLM generates high-quality data, with performance on downstream tasks surpassing that of real data by 2%-7% in certain cases. Data and code are available at https://github.com/bytedance/DiffLM.
Authors: Zachary Coalson, Jeonghyun Woo, Yu Sun, Shiyang Chen, Lishan Yang, Prashant Nair, Bo Fang, Sanghyun Hong
Abstract: We introduce a new class of attacks on commercial-scale (human-aligned) language models that induce jailbreaking through targeted bitwise corruptions in model parameters. Our adversary can jailbreak billion-parameter language models with fewer than 25 bit-flips in all cases$-$and as few as 5 in some$-$using up to 40$\times$ less bit-flips than existing attacks on computer vision models at least 100$\times$ smaller. Unlike prompt-based jailbreaks, our attack renders these models in memory 'uncensored' at runtime, allowing them to generate harmful responses without any input modifications. Our attack algorithm efficiently identifies target bits to flip, offering up to 20$\times$ more computational efficiency than previous methods. This makes it practical for language models with billions of parameters. We show an end-to-end exploitation of our attack using software-induced fault injection, Rowhammer (RH). Our work examines 56 DRAM RH profiles from DDR4 and LPDDR4X devices with different RH vulnerabilities. We show that our attack can reliably induce jailbreaking in systems similar to those affected by prior bit-flip attacks. Moreover, our approach remains effective even against highly RH-secure systems (e.g., 46$\times$ more secure than previously tested systems). Our analyses further reveal that: (1) models with less post-training alignment require fewer bit flips to jailbreak; (2) certain model components, such as value projection layers, are substantially more vulnerable than others; and (3) our method is mechanistically different than existing jailbreaks. Our findings highlight a pressing, practical threat to the language model ecosystem and underscore the need for research to protect these models from bit-flip attacks.
Authors: Kaiqu Liang, Haimin Hu, Ryan Liu, Thomas L. Griffiths, Jaime Fern\'andez Fisac
Abstract: While Reinforcement Learning from Human Feedback (RLHF) has shown promise in aligning generative AI, we present empirical evidence that it can also cause severe, systematic misalignment. We hypothesize that this stems from evaluator feedback depending on downstream outcome predictions (foresight) that can be influenced by the AI's output, inducing Goodhart's law dynamics. We present a theoretical analysis showing that conditioning evaluator feedback on downstream observations (hindsight) inhibits this effect by decoupling the alignment signal from potentially compromised predictions--crucially, the result holds even if the observed outcomes are sampled from the AI's own world model. Building on this insight, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which presents plausible simulated outcomes to evaluators before eliciting feedback. We validate RLHS across three consultancy settings--marketplace interactions, restaurant recommendations, and online course advising--using both online (PPO) and offline (DPO) fine-tuning methods, and show that it substantially improves alignment over RLHF in experiments and human evaluations. We perform post-hoc benchmark evaluations on TruthfulQA, HaluEval, and TrustLLM, finding that even after single-task fine-tuning, RLHF misalignment persists, whereas RLHS consistently outperforms baselines and demonstrates robust alignment generalization. The project webpage and code are available at https://rl-hindsight.github.io.
Authors: Navid Rajabi, Jana Kosecka
Abstract: In this work, we propose a modular approach for the Vision-Language Navigation (VLN) task by decomposing the problem into four sub-modules that use state-of-the-art Large Language Models (LLMs) and Vision-Language Models (VLMs) in a zero-shot setting. Given navigation instruction in natural language, we first prompt LLM to extract the landmarks and the order in which they are visited. Assuming the known model of the environment, we retrieve the top-k locations of the last landmark and generate $k$ path hypotheses from the starting location to the last landmark using the shortest path algorithm on the topological map of the environment. Each path hypothesis is represented by a sequence of panoramas. We then use dynamic programming to compute the alignment score between the sequence of panoramas and the sequence of landmark names, which match scores obtained from VLM. Finally, we compute the nDTW metric between the hypothesis that yields the highest alignment score to evaluate the path fidelity. We demonstrate superior performance compared to other approaches that use joint semantic maps like VLMaps on the complex R2R-Habitat instruction dataset and quantify in detail the effect of visual grounding on navigation performance.
Authors: Yudong W. Xu, Wenhao Li, Scott Sanner, Elias B. Khalil
Abstract: We present a Transformer-based framework for Constraint Satisfaction Problems (CSPs). CSPs find use in many applications and thus accelerating their solution with machine learning is of wide interest. Most existing approaches rely on supervised learning from feasible solutions or reinforcement learning, paradigms that require either feasible solutions to these NP-Complete CSPs or large training budgets and a complex expert-designed reward signal. To address these challenges, we propose ConsFormer, a self-supervised framework that leverages a Transformer as a solution refiner. ConsFormer constructs a solution to a CSP iteratively in a process that mimics local search. Instead of using feasible solutions as labeled data, we devise differentiable approximations to the discrete constraints of a CSP to guide model training. Our model is trained to improve random assignments for a single step but is deployed iteratively at test time, circumventing the bottlenecks of supervised and reinforcement learning. Experiments on Sudoku, Graph Coloring, Nurse Rostering, and MAXCUT demonstrate that our method can tackle out-of-distribution CSPs simply through additional iterations.
Authors: Boyi Kang, Xinfa Zhu, Zihan Zhang, Zhen Ye, Mingshuai Liu, Ziqian Wang, Yike Zhu, Guobin Ma, Jun Chen, Longshuai Xiao, Chao Weng, Wei Xue, Lei Xie
Abstract: Recent advancements in language models (LMs) have demonstrated strong capabilities in semantic understanding and contextual modeling, which have flourished in generative speech enhancement (SE). However, many LM-based SE approaches primarily focus on semantic information, often neglecting the critical role of acoustic information, which leads to acoustic inconsistency after enhancement and limited generalization across diverse SE tasks. In this paper, we introduce LLaSE-G1, a LLaMA-based language model that incentivizes generalization capabilities for speech enhancement. LLaSE-G1 offers the following key contributions: First, to mitigate acoustic inconsistency, LLaSE-G1 employs continuous representations from WavLM as input and predicts speech tokens from X-Codec2, maximizing acoustic preservation. Second, to promote generalization capability, LLaSE-G1 introduces dual-channel inputs and outputs, unifying multiple SE tasks without requiring task-specific IDs. Third, LLaSE-G1 outperforms prior task-specific discriminative and generative SE models, demonstrating scaling effects at test time and emerging capabilities for unseen SE tasks. Additionally, we release our code and models to support further research in this area.
Authors: Caterina Fuligni, Daniel Dominguez Figaredo, Julia Stoyanovich
Abstract: In recent years, Large Language Models (LLMs) rapidly gained popularity across all parts of society, including education. After initial skepticism and bans, many schools have chosen to embrace this new technology by integrating it into their curricula in the form of virtual tutors and teaching assistants. However, neither the companies developing this technology nor the public institutions involved in its implementation have set up a formal system to collect feedback from the stakeholders impacted by them. In this paper, we argue that understanding the perceptions of those directly or indirectly impacted by LLMs in the classroom, including parents and school staff, is essential for ensuring responsible use of AI in this critical domain. Our contributions are two-fold. First, we propose the Contextualized Perceptions for the Adoption of LLMs in Education (Co-PALE) framework, which can be used to systematically elicit perceptions and inform whether and how LLM-based tools should be designed, developed, and deployed in the classroom. Second, we explain how our framework can be used to ground specific rubrics for eliciting perceptions of the relevant stakeholders in view of specific goals and context of implementation. Overall, Co-PALE is a practical step toward helping educational agents, policymakers, researchers, and technologists ensure the responsible and effective deployment of LLM-based systems across diverse learning contexts.
Authors: Yash Vekaria (UC Davis), Aurelio Loris Canino (Mediterranea University of Reggio Calabria), Jonathan Levitsky (UC Davis), Alex Ciechonski (University College London), Patricia Callejo (Universidad Carlos III de Madrid), Anna Maria Mandalari (University College London), Zubair Shafiq (UC Davis)
Abstract: Generative AI (GenAI) browser assistants integrate powerful capabilities of GenAI in web browsers to provide rich experiences such as question answering, content summarization, and agentic navigation. These assistants, available today as browser extensions, can not only track detailed browsing activity such as search and click data, but can also autonomously perform tasks such as filling forms, raising significant privacy concerns. It is crucial to understand the design and operation of GenAI browser extensions, including how they collect, store, process, and share user data. To this end, we study their ability to profile users and personalize their responses based on explicit or inferred demographic attributes and interests of users. We perform network traffic analysis and use a novel prompting framework to audit tracking, profiling, and personalization by the ten most popular GenAI browser assistant extensions. We find that instead of relying on local in-browser models, these assistants largely depend on server-side APIs, which can be auto-invoked without explicit user interaction. When invoked, they collect and share webpage content, often the full HTML DOM and sometimes even the user's form inputs, with their first-party servers. Some assistants also share identifiers and user prompts with third-party trackers such as Google Analytics. The collection and sharing continues even if a webpage contains sensitive information such as health or personal information such as name or SSN entered in a web form. We find that several GenAI browser assistants infer demographic attributes such as age, gender, income, and interests and use this profile--which carries across browsing contexts--to personalize responses. In summary, our work shows that GenAI browser assistants can and do collect personal and sensitive information for profiling and personalization with little to no safeguards.
Authors: Jihwan Oh, Minchan Jeong, Jongwoo Ko, Se-Young Yun
Abstract: Large Language Models $($LLMs$)$ solve complex problems using training-free methods like prompt engineering and in-context learning, yet ensuring reasoning correctness remains challenging. While self-correction methods such as self-consistency and self-refinement aim to improve reliability, they often reinforce biases due to the lack of effective feedback mechanisms. Multi-Agent Debate $($MAD$)$ has emerged as an alternative, but we identify two key limitations: bias reinforcement, where debate amplifies model biases instead of correcting them, and lack of perspective diversity, as all agents share the same model and reasoning patterns, limiting true debate effectiveness. To systematically evaluate these issues, we introduce $\textit{MetaNIM Arena}$, a benchmark designed to assess LLMs in adversarial strategic decision-making, where dynamic interactions influence optimal decisions. To overcome MAD's limitations, we propose $\textbf{DReaMAD}$ $($$\textbf{D}$iverse $\textbf{Rea}$soning via $\textbf{M}$ulti-$\textbf{A}$gent $\textbf{D}$ebate with Refined Prompt$)$, a novel framework that $(1)$ refines LLM's strategic prior knowledge to improve reasoning quality and $(2)$ promotes diverse viewpoints within a single model by systematically modifying prompts, reducing bias. Empirical results show that $\textbf{DReaMAD}$ significantly improves decision accuracy, reasoning diversity, and bias mitigation across multiple strategic tasks, establishing it as a more effective approach for LLM-based decision-making.
Authors: Hung-Yueh Chiang, Chi-Chih Chang, Natalia Frumkin, Kai-Chiang Wu, Mohamed S. Abdelfattah, Diana Marculescu
Abstract: State Space Models (SSMs) are emerging as a compelling alternative to Transformers because of their consistent memory usage and high performance. Despite this, scaling up SSMs on cloud services or limited-resource devices is challenging due to their storage requirements and computational power. To overcome this, quantizing SSMs with low bit-width data formats can reduce model size and benefit from hardware acceleration. As SSMs are prone to quantization-induced errors, recent efforts have focused on optimizing a particular model or bit-width for efficiency without sacrificing performance. However, distinct bit-width configurations are essential for different scenarios, like W4A8 for boosting large-batch decoding speed, and W4A16 for enhancing generation speed in short prompt applications for a single user. To this end, we present Quamba2, compatible with W8A8, W4A8, and W4A16 for both Mamba1 and Mamba2 backbones, addressing the growing demand for SSM deployment on various platforms. Based on the channel order preserving and activation persistence of SSMs, we propose an offline approach to quantize inputs of a linear recurrence in 8-bit by sorting and clustering for input $x$, combined with a per-state-group quantization for input-dependent parameters $B$ and $C$. To ensure compute-invariance in the SSM output, we rearrange weights offline according to the clustering sequence. The experiments show that Quamba2-8B outperforms two state-of-the-art SSM quantization methods and delivers 1.3$\times$ and 3$\times$ speed-ups in the pre-filling and generation stages, respectively, while offering 4$\times$ memory reduction with only a $1.6\%$ average accuracy drop. The evaluation on MMLU shows the generalizability and robustness of our framework. The code and quantized models will be released at: https://github.com/enyac-group/Quamba.
Authors: Oren Sultan, Eitan Stern, Dafna Shahaf
Abstract: Large language models (LLMs) struggle with formal domains that require rigorous logical deduction and symbolic reasoning, such as mathematical proof generation. We propose a neuro-symbolic approach that combines LLMs' generative strengths with structured components to overcome this challenge. As a proof-of-concept, we focus on geometry problems. Our approach is two-fold: (1) we retrieve analogous problems and use their proofs to guide the LLM, and (2) a formal verifier evaluates the generated proofs and provides feedback, helping the model fix incorrect proofs. We demonstrate that our method significantly improves proof accuracy for OpenAI's o1 model (58%-70% improvement); both analogous problems and the verifier's feedback contribute to these gains. More broadly, shifting to LLMs that generate provably correct conclusions could dramatically improve their reliability, accuracy and consistency, unlocking complex tasks and critical real-world applications that require trustworthiness.
Authors: Chengzhuo Tong, Ziyu Guo, Renrui Zhang, Wenyu Shan, Xinyu Wei, Zhenghao Xing, Hongsheng Li, Pheng-Ann Heng
Abstract: Recent advancements underscore the significant role of Reinforcement Learning (RL) in enhancing the Chain-of-Thought (CoT) reasoning capabilities of large language models (LLMs). Two prominent RL algorithms, Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO), are central to these developments, showcasing different pros and cons. Autoregressive image generation, also interpretable as a sequential CoT reasoning process, presents unique challenges distinct from LLM-based CoT reasoning. These encompass ensuring text-image consistency, improving image aesthetic quality, and designing sophisticated reward models, rather than relying on simpler rule-based rewards. While recent efforts have extended RL to this domain, these explorations typically lack an in-depth analysis of the domain-specific challenges and the characteristics of different RL strategies. To bridge this gap, we provide the first comprehensive investigation of the GRPO and DPO algorithms in autoregressive image generation, evaluating their in-domain performance and out-of-domain generalization, while scrutinizing the impact of different reward models on their respective capabilities. Our findings reveal that GRPO and DPO exhibit distinct advantages, and crucially, that reward models possessing stronger intrinsic generalization capabilities potentially enhance the generalization potential of the applied RL algorithms. Furthermore, we systematically explore three prevalent scaling strategies to enhance both their in-domain and out-of-domain proficiency, deriving unique insights into efficiently scaling performance for each paradigm. We hope our study paves a new path for inspiring future work on developing more effective RL algorithms to achieve robust CoT reasoning in the realm of autoregressive image generation. Code is released at https://github.com/ZiyuGuo99/Image-Generation-CoT
Authors: Wasi Uddin Ahmad, Somshubra Majumdar, Boris Ginsburg
Abstract: Post-processing is crucial for the automatic evaluation of LLMs in fill-in-the-middle (FIM) code generation due to the frequent presence of extraneous code in raw outputs. This extraneous generation suggests a lack of awareness regarding output boundaries, requiring truncation for effective evaluation. The determination of an optimal truncation strategy, however, often proves intricate, particularly when the scope includes several programming languages. This study investigates the necessity of post-processing instruction-tuned LLM outputs. Our findings reveal that supervised fine-tuning significantly enhances FIM code generation, enabling LLMs to generate code that seamlessly integrates with the surrounding context. Evaluating our fine-tuned \texttt{Qwen2.5-Coder} (base and instruct) models on HumanEval Infilling and SAFIM benchmarks demonstrates improved performances without post-processing, especially when the \emph{middle} consist of complete lines. However, post-processing of the LLM outputs remains necessary when the \emph{middle} is a random span of code.
Authors: Kidist Amde Mekonnen, Yosef Worku Alemneh, Maarten de Rijke
Abstract: Neural retrieval methods using transformer-based pre-trained language models have advanced multilingual and cross-lingual retrieval. However, their effectiveness for low-resource, morphologically rich languages such as Amharic remains underexplored due to data scarcity and suboptimal tokenization. We address this gap by introducing Amharic-specific dense retrieval models based on pre-trained Amharic BERT and RoBERTa backbones. Our proposed RoBERTa-Base-Amharic-Embed model (110M parameters) achieves a 17.6% relative improvement in MRR@10 and a 9.86% gain in Recall@10 over the strongest multilingual baseline, Arctic Embed 2.0 (568M parameters). More compact variants, such as RoBERTa-Medium-Amharic-Embed (42M), remain competitive while being over 13x smaller. Additionally, we train a ColBERT-based late interaction retrieval model that achieves the highest MRR@10 score (0.843) among all evaluated models. We benchmark our proposed models against both sparse and dense retrieval baselines to systematically assess retrieval effectiveness in Amharic. Our analysis highlights key challenges in low-resource settings and underscores the importance of language-specific adaptation. To foster future research in low-resource IR, we publicly release our dataset, codebase, and trained models at https://github.com/kidist-amde/amharic-ir-benchmarks.
Authors: Guangfu Hao, Haojie Wen, Liangxuna Guo, Yang Chen, Yanchao Bi, Shan Yu
Abstract: Flexible tool selection reflects a complex cognitive ability that distinguishes humans from other species, yet computational models that capture this ability remain underdeveloped. We developed a framework using low-dimensional attribute representations to bridge visual tool perception and linguistic task understanding. We constructed a comprehensive dataset (ToolNet) containing 115 common tools labeled with 13 carefully designed attributes spanning physical, functional, and psychological properties, paired with natural language scenarios describing tool usage. Visual encoders (ResNet or ViT) extract attributes from tool images while fine-tuned language models (GPT-2, LLaMA, DeepSeek) derive required attributes from task descriptions. Our approach achieves 74% accuracy in tool selection tasks-significantly outperforming direct tool matching (20%) and smaller multimodal models (21%-58%), while approaching performance of much larger models like GPT-4o (73%) with substantially fewer parameters. Ablation studies revealed that manipulation-related attributes (graspability, hand-relatedness, elongation) consistently prove most critical across modalities. This work provides a parameter-efficient, interpretable solution that mimics human-like tool cognition, advancing both cognitive science understanding and practical applications in tool selection tasks.
Authors: Lin Sun, Weihong Lin, Jinzhu Wu, Yongfu Zhu, Xiaoqi Jian, Guangxiang Zhao, Change Jia, Linglin Zhang, Sai-er Hu, Yuhan Wu, Xiangzheng Zhang
Abstract: Reasoning models represented by the Deepseek-R1-Distill series have been widely adopted by the open-source community due to their strong performance in mathematics, science, programming, and other domains. However, our study reveals that their benchmark evaluation results are subject to significant fluctuations caused by various factors. Subtle differences in evaluation conditions can lead to substantial variations in results. Similar phenomena are observed in other open-source inference models fine-tuned based on the Deepseek-R1-Distill series, as well as in the QwQ-32B model, making their claimed performance improvements difficult to reproduce reliably. Therefore, we advocate for the establishment of a more rigorous paradigm for model performance evaluation and present our empirical assessments of the Deepseek-R1-Distill series models.
Authors: Keinichi Fujita, Shota Horiguchi, Yusuke Ijima
Abstract: Para-/non-linguistic information in speech is pivotal in shaping the listeners' impression. Although zero-shot text-to-speech (TTS) has achieved high speaker fidelity, modulating subtle para-/non-linguistic information to control perceived voice characteristics, i.e., impressions, remains challenging. We have therefore developed a voice impression control method in zero-shot TTS that utilizes a low-dimensional vector to represent the intensities of various voice impression pairs (e.g., dark-bright). The results of both objective and subjective evaluations have demonstrated our method's effectiveness in impression control. Furthermore, generating this vector via a large language model enables target-impression generation from a natural language description of the desired impression, thus eliminating the need for manual optimization. Audio examples are available on our demo page (https://ntt-hilab-gensp.github.io/is2025voiceimpression/).
URLs: https://ntt-hilab-gensp.github.io/is2025voiceimpression/).
Authors: Daniel Thilo Schroeder, Meeyoung Cha, Andrea Baronchelli, Nick Bostrom, Nicholas A. Christakis, David Garcia, Amit Goldenberg, Yara Kyrychenko, Kevin Leyton-Brown, Nina Lutz, Gary Marcus, Filippo Menczer, Gordon Pennycook, David G. Rand, Frank Schweitzer, Christopher Summerfield, Audrey Tang, Jay Van Bavel, Sander van der Linden, Dawn Song, Jonas R. Kunst
Abstract: Advances in AI portend a new era of sophisticated disinformation operations. While individual AI systems already create convincing -- and at times misleading -- information, an imminent development is the emergence of malicious AI swarms. These systems can coordinate covertly, infiltrate communities, evade traditional detectors, and run continuous A/B tests, with round-the-clock persistence. The result can include fabricated grassroots consensus, fragmented shared reality, mass harassment, voter micro-suppression or mobilization, contamination of AI training data, and erosion of institutional trust. With democratic processes worldwide increasingly vulnerable, we urge a three-pronged response: (1) platform-side defenses -- always-on swarm-detection dashboards, pre-election high-fidelity swarm-simulation stress-tests, transparency audits, and optional client-side "AI shields" for users; (2) model-side safeguards -- standardized persuasion-risk tests, provenance-authenticating passkeys, and watermarking; and (3) system-level oversight -- a UN-backed AI Influence Observatory.
Authors: Akash Gupta, Amos Storkey, Mirella Lapata
Abstract: Large Multimodal Models (LMMs) often rely on in-context learning (ICL) to perform new tasks with minimal supervision. However, ICL performance, especially in smaller LMMs, is inconsistent and does not always improve monotonically with increasing examples. We hypothesize that this occurs due to the LMM being overwhelmed by additional information present in the image embeddings, which is not required for the downstream task. To address this, we propose a meta-learning approach that provides an alternative for inducing few-shot capabilities in LMMs, using a fixed set of soft prompts that are distilled from task-relevant image features and can be adapted at test time using a few examples. To facilitate this distillation, we introduce an attention-mapper module that can be easily integrated with the popular LLaVA v1.5 architecture and is jointly learned with soft prompts, enabling task adaptation in LMMs under low-data regimes with just a few gradient steps. Evaluation on the VL-ICL Bench shows that our method consistently outperforms ICL and related prompt-tuning approaches, even under image perturbations, improving task induction and reasoning across visual question answering tasks.
Authors: Xiaoyuan Zhu, Yaowen Ye, Tianyi Qiu, Hanlin Zhu, Sijun Tan, Ajraf Mannan, Jonathan Michala, Raluca Ada Popa, Willie Neiswanger
Abstract: As API access becomes a primary interface to large language models (LLMs), users often interact with black-box systems that offer little transparency into the deployed model. To reduce costs or maliciously alter model behaviors, API providers may discreetly serve quantized or fine-tuned variants, which can degrade performance and compromise safety. Detecting such substitutions is difficult, as users lack access to model weights and, in most cases, even output logits. To tackle this problem, we propose a rank-based uniformity test that can verify the behavioral equality of a black-box LLM to a locally deployed authentic model. Our method is accurate, query-efficient, and avoids detectable query patterns, making it robust to adversarial providers that reroute or mix responses upon the detection of testing attempts. We evaluate the approach across diverse threat scenarios, including quantization, harmful fine-tuning, jailbreak prompts, and full model substitution, showing that it consistently achieves superior statistical power over prior methods under constrained query budgets.
Authors: Peiran Li, Xinkai Zou, Zhuohang Wu, Ruifeng Li, Shuo Xing, Hanwen Zheng, Zhikai Hu, Yuping Wang, Haoxi Li, Qin Yuan, Yingmo Zhang, Zhengzhong Tu
Abstract: Recent advances in large language models (LLMs) and vision-language models (VLMs) have enabled powerful autonomous agents capable of complex reasoning and multi-modal tool use. Despite their growing capabilities, today's agent frameworks remain fragile, lacking principled mechanisms for secure information flow, reliability, and multi-agent coordination. In this work, we introduce SAFEFLOW, a new protocol-level framework for building trustworthy LLM/VLM-based agents. SAFEFLOW enforces fine-grained information flow control (IFC), precisely tracking provenance, integrity, and confidentiality of all the data exchanged between agents, tools, users, and environments. By constraining LLM reasoning to respect these security labels, SAFEFLOW prevents untrusted or adversarial inputs from contaminating high-integrity decisions. To ensure robustness in concurrent multi-agent settings, SAFEFLOW introduces transactional execution, conflict resolution, and secure scheduling over shared state, preserving global consistency across agents. We further introduce mechanisms, including write-ahead logging, rollback, and secure caches, that further enhance resilience against runtime errors and policy violations. To validate the performances, we built SAFEFLOWBENCH, a comprehensive benchmark suite designed to evaluate agent reliability under adversarial, noisy, and concurrent operational conditions. Extensive experiments demonstrate that agents built with SAFEFLOW maintain impressive task performance and security guarantees even in hostile environments, substantially outperforming state-of-the-art. Together, SAFEFLOW and SAFEFLOWBENCH lay the groundwork for principled, robust, and secure agent ecosystems, advancing the frontier of reliable autonomy.