Authors: Jingfei Huang, Han Tu
Abstract: The ascension of social media platforms has transformed our understanding of urban environments, giving rise to nuanced variations in sentiment reaction embedded within human perception and opinion, and challenging existing multidimensional sentiment analysis approaches in urban studies. This study presents novel methodologies for identifying and elucidating sentiment inconsistency, constructing a dataset encompassing 140,750 Baidu and Tencent Street view images to measure perceptions, and 984,024 Weibo social media text posts to measure opinions. A reaction index is developed, integrating object detection and natural language processing techniques to classify sentiment in Beijing Second Ring for 2016 and 2022. Classified sentiment reaction is analysed and visualized using regression analysis, image segmentation, and word frequency based on land-use distribution to discern underlying factors. The perception affective reaction trend map reveals a shift toward more evenly distributed positive sentiment, while the opinion affective reaction trend map shows more extreme changes. Our mismatch map indicates significant disparities between the sentiments of human perception and opinion of urban areas over the years. Changes in sentiment reactions have significant relationships with elements such as dense buildings and pedestrian presence. Our inconsistent maps present perception and opinion sentiments before and after the pandemic and offer potential explanations and directions for environmental management, in formulating strategies for urban renewal.
Authors: Mufei Li, Dongqi Fu, Limei Wang, Si Zhang, Hanqing Zeng, Kaan Sancak, Ruizhong Qiu, Haoyu Wang, Xiaoxin He, Xavier Bresson, Yinglong Xia, Chonglin Sun, Pan Li
Abstract: Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased retrieval and agentic workflows. We argue that haystack engineering is necessary to construct noisy long contexts that faithfully capture key real-world factors -- distraction from heterogeneous biased retrievers and cascading errors in agentic workflows -- to test models' long-context robustness. We instantiate it through HaystackCraft, a new NIAH benchmark built on the full English Wikipedia hyperlink network with multi-hop questions. HaystackCraft evaluates how heterogeneous retrieval strategies (e.g., sparse, dense, hybrid, and graph-based) affect distractor composition, haystack ordering, and downstream LLM performance. HaystackCraft further extends NIAH to dynamic, LLM-dependent settings that simulate agentic operations, where models refine queries, reflect on their past reasonings, and decide when to stop. Experiments with 15 long-context models show that (1) while stronger dense retrievers can introduce more challenging distractors, graph-based reranking simultaneously improves retrieval effectiveness and mitigates more harmful distractors; (2) in agentic tests, even advanced models like Gemini 2.5 Pro and GPT-5 suffer cascading failures from self-generated distractors or struggle to perform early stops. These results highlight persistent challenges in agentic long-context reasoning and establish HaystackCraft as a valuable testbed for future progress.
Authors: Olia Toporkov, Alan Akbik, Rodrigo Agerri
Abstract: Lemmatization is the task of transforming all words in a given text to their dictionary forms. While large language models (LLMs) have demonstrated their ability to achieve competitive results across a wide range of NLP tasks, there is no prior evidence of how effective they are in the contextual lemmatization task. In this paper, we empirically investigate the capacity of the latest generation of LLMs to perform in-context lemmatization, comparing it to the traditional fully supervised approach. In particular, we consider the setting in which supervised training data is not available for a target domain or language, comparing (i) encoder-only supervised approaches, fine-tuned out-of-domain, and (ii) cross-lingual methods, against direct in-context lemma generation with LLMs. Our experimental investigation across 12 languages of different morphological complexity finds that, while encoders remain competitive in out-of-domain settings when fine-tuned on gold data, current LLMs reach state-of-the-art results for most languages by directly generating lemmas in-context without prior fine-tuning, provided just with a few examples. Data and code available upon publication: https://github.com/oltoporkov/lemma-dilemma
Authors: Amruta Parulekar, Preethi Jyothi
Abstract: Standard ASR evaluation metrics like Word Error Rate (WER) tend to unfairly penalize morphological and syntactic nuances that do not significantly alter sentence semantics. We introduce an LLM-based scoring rubric LASER that leverages state-of-the-art LLMs' in-context learning abilities to learn from prompts with detailed examples. Hindi LASER scores using Gemini 2.5 Pro achieved a very high correlation score of 94% with human annotations. Hindi examples in the prompt were also effective in analyzing errors in other Indian languages such as Marathi, Kannada and Malayalam. We also demonstrate how a smaller LLM like Llama 3 can be finetuned on word-pair examples derived from reference and ASR predictions to predict what kind of penalty should be applied with close to 89% accuracy.
Authors: Zifan Jiang, Colin Leong, Amit Moryossef, Anne G\"ohring, Annette Rios, Oliver Cory, Maksym Ivashechkin, Neha Tarigopula, Biao Zhang, Rico Sennrich, Sarah Ebling
Abstract: We present a comprehensive study on meaningfully evaluating sign language utterances in the form of human skeletal poses. The study covers keypoint distance-based, embedding-based, and back-translation-based metrics. We show tradeoffs between different metrics in different scenarios through automatic meta-evaluation of sign-level retrieval and a human correlation study of text-to-pose translation across different sign languages. Our findings and the open-source pose-evaluation toolkit provide a practical and reproducible way of developing and evaluating sign language translation or generation systems.
Authors: Eduardo Ry\^o Tamaki (German Institute for Global,Area Studies), Yujin J. Jung (Mount St. Mary's University), Julia Chatterley (Princeton University), Grant Mitchell (University of California, Los Angeles), Semir Dzebo (University of Oxford), Crist\'obal Sandoval (Diego Portales University), Levente Littvay (ELTE Centre for Social Sciences), Kirk A. Hawkins (Brigham Young University)
Abstract: Measuring the ideational content of populism remains a challenge. Traditional strategies based on textual analysis have been critical for building the field's foundations and providing a valid, objective indicator of populist framing. Yet these approaches are costly, time consuming, and difficult to scale across languages, contexts, and large corpora. Here we present the results from a rubric and anchor guided chain of thought (CoT) prompting approach that mirrors human coder training. By leveraging the Global Populism Database (GPD), a comprehensive dataset of global leaders' speeches annotated for degrees of populism, we replicate the process used to train human coders by prompting the LLM with an adapted version of the same documentation to guide the model's reasoning. We then test multiple proprietary and open weight models by replicating scores in the GPD. Our findings reveal that this domain specific prompting strategy enables the LLM to achieve classification accuracy on par with expert human coders, demonstrating its ability to navigate the nuanced, context sensitive aspects of populism.
Authors: Zheyuan Zhang, Lin Ge, Hongjiang Li, Weicheng Zhu, Chuxu Zhang, Yanfang Ye
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks, and LLM-based agents further extend these abilities to various practical workflows. While recent progress shows that multi-agent systems (MAS) can outperform single agents by coordinating specialized roles, designing effective MAS remains difficult due to prompt sensitivity and the compounded instability MAS creates. To cope with the challenge, recent efforts in automated prompt design have reduced manual effort. However, multi-agent prompt optimization remains largely unexplored. Challenges like exponentially expanding search space and ambiguous credit assignment together make systematic design intractable without principled methods. Therefore, we introduce M}ulti-Agent PRompt Optimization (MAPRO), a four-stage framework that first formulates MAS prompt optimization as a Maximum a Posteriori (MAP) inference problem and solves it using a language-guided variant of max-product belief propagation algorithm. To address credit assignment and updates the system iteratively, MAPRO employs a topology-aware refinement mechanism that integrates execution feedback and downstream blames to selectively update agent prompts. Through this process, MAPRO progressively converges to a coordinated set of agent-specific prompt policies. Across benchmarks in various tasks, MAPRO achieves state-of-the-art performance, consistently surpassing manually engineered baselines and recent automated alternatives. Beyond performance, our MAP-based formulation also delivers general guidelines for building more reliable and principled multi-agent systems in the future
Authors: Shuqing Luo, Yilin Guan, Pingzhi Li, Hanrui Wang, Tianlong Chen
Abstract: Test-time scaling (TTS) boosts LLM reasoning via long chain-of-thought (CoT), but the linear KV-cache growth amplifies the memory-bound bottleneck of LLM decoding. Query-aware page-level sparse decoding can achieve state-of-the-art performance under constrained FLOPs budgets, but is limited by both sequential-dependent page filtering and coarse-grained token selection, hampering serving efficiency and model performance on TTS tasks under high concurrency and long CoT scenarios (consuming even higher runtime than the forward pipeline itself). In this paper, we first find that the current-step query state can be accurately approximated in a unified manner from a short window of recent queries, enabling training-free query-aware sparsity without waiting in the decoding loop. We propose AsyncSpade, an asynchronous framework for efficient TTS built on two core components: (1) a novel light-weight temporal-regressive module that predicts the next-token query state; (2) an asynchronous and disaggregated framework that decouples the KV cache filtering from the auto-regressive decoding loop, overlapping the token-level KV selection with the forward inference computation through asynchronism. To our knowledge, AsyncSpade is the first to eliminate the sequential dependence without sacrificing model performance. We validate the effectiveness of AsyncSpade on common LLM serving setups with an A100 node, where AsyncSpade fully overlaps KV-cache operations with the inference pipeline, achieving theoretical optimal time-per-output-token (TPOT). Specifically, AsyncSpade delivers over 20% reduction on TPOT compared to SoTA baseline (i.e. Quest) and at least 50% TPOT reduction compared to full attention on Qwen3-8B and Qwen3-32B models, while matching or surpassing their accuracy on various TTS benchmarks (AIME-24/25, GPQA-Diamond, MATH-500).
Authors: Rasika Muralidharan, Jaewoon Kwak, Jisun An
Abstract: Multi-Agent Systems (MAS) with Large Language Model (LLM)-powered agents are gaining attention, yet fewer studies explore their team dynamics. Inspired by human team science, we propose a multi-agent framework to examine core aspects of team science: structure, diversity, and interaction dynamics. We evaluate team performance across four tasks: CommonsenseQA, StrategyQA, Social IQa, and Latent Implicit Hate, spanning commonsense and social reasoning. Our results show that flat teams tend to perform better than hierarchical ones, while diversity has a nuanced impact. Interviews suggest agents are overconfident about their team performance, yet post-task reflections reveal both appreciation for collaboration and challenges in integration, including limited conversational coordination.
Authors: Yi-Jen Shih, Desh Raj, Chunyang Wu, Wei Zhou, SK Bong, Yashesh Gaur, Jay Mahadeokar, Ozlem Kalinli, Mike Seltzer
Abstract: Recent advances in speech large language models (speech LLMs) have enabled seamless spoken interactions, but these systems still struggle with complex reasoning tasks. Previously, chain-of-thought (CoT) prompting or fine-tuning has been to shown to significantly improve the reasoning abilities of text-based LLMs. In this work, we investigate the effect of CoT fine-tuning for multi-stream speech LLMs, demonstrating that reasoning in text space improves the accuracy of speech LLMs by 2.4x, on average, over a suite of spoken reasoning tasks. Beyond accuracy, the latency of the spoken response is a crucial factor for interacting with voice-based agents. Inspired by the human behavior of "thinking while listening," we propose methods to reduce the additional latency from reasoning by allowing the model to start reasoning before the user query has ended. To achieve this, we introduce an entropy-based metric, "question completeness," which acts as an indicator to guide the model on the optimal time to start reasoning. This method provides greater control over the accuracy-latency trade-off compared with heuristic-based approaches and, under equivalent latency conditions, yields a 4% accuracy gain on ARC-Easy. Finally, we use Direct Preference Optimization (DPO) on preference data created using rejection sampling to push the accuracy-latency pareto frontier further, resulting in a 70% reduction in latency without loss in accuracy.
Authors: Soyeong Jeong, Taehee Jung, Sung Ju Hwang, Joo-Kyung Kim, Dongyeop Kang
Abstract: Recent Long-Context Language Models (LCLMs) can process hundreds of thousands of tokens in a single prompt, enabling new opportunities for knowledge-intensive multi-hop reasoning by integrating large sets of retrieved documents or, in some cases, directly all necessary information. However, simply feeding more documents into the context window fails to capture how evidence should be connected. We address this gap with thought templates, which recast reasoning as reusable thought caches, derived from prior problem solving traces, structuring how evidence is combined and guiding multi-hop inference with factual documents. To keep these templates effective, we propose an update strategy that iteratively refines templates derived from training data through natural-language feedback. Across diverse benchmarks and LCLM families, our approach delivers consistent gains over strong baselines in both retrieval-based and retrieval-free settings. Furthermore, we show that optimized templates can be distilled into smaller open-source models, demonstrating its broad applicability and transparent reasoning reuse. We refer to our framework as Thought Template Augmented LCLMs (ToTAL).
Authors: Rayyan Merchant, Kevin Tang
Abstract: As a digraphic language, the Persian language utilizes two written standards: Perso-Arabic in Afghanistan and Iran, and Tajik-Cyrillic in Tajikistan. Despite the significant similarity between the dialects of each country, script differences prevent simple one-to-one mapping, hindering written communication and interaction between Tajikistan and its Persian-speaking ``siblings''. To overcome this, previously-published efforts have investigated machine transliteration models to convert between the two scripts. Unfortunately, most efforts did not use datasets other than those they created, limiting these models to certain domains of text such as archaic poetry or word lists. A truly usable transliteration system must be capable of handling varied domains, meaning that suck models lack the versatility required for real-world usage. The contrast in domain between data also obscures the task's true difficulty. We present a new state-of-the-art sequence-to-sequence model for Tajik-Farsi transliteration trained across all available datasets, and present two datasets of our own. Our results across domains provide clearer understanding of the task, and set comprehensive comparable leading benchmarks. Overall, our model achieves chrF++ and Normalized CER scores of 87.91 and 0.05 from Farsi to Tajik and 92.28 and 0.04 from Tajik to Farsi. Our model, data, and code are available at https://anonymous.4open.science/r/ParsTranslit-FB30/.
Authors: Jaeseong Lee, seung-won hwang, Aurick Qiao, Gabriele Oliaro, Ye Wang, Samyam Rajbhandari
Abstract: Speculative decoding promises faster inference for large language models (LLMs), yet existing methods fail to generalize to real-world settings. Benchmarks typically assume short contexts (e.g., 2K tokens), whereas practical workloads involve long contexts. We find current approaches degrade severely with long contexts; for instance, EAGLE3 even slows down the generation speed by 0.81x. We address these limitations by releasing a new long-context benchmark (LongSpecBench) and introducing a novel model (OWL). OWL achieves about 5x higher acceptance length than EAGLE3 on long-context inputs through three innovations: (1) an LSTM-based drafter conditioned only on the last-token state, making it generalize to various lengths, (2) a special token [SPEC] in the verifier that produces richer representation for drafter, and (3) a hybrid algorithm combining both tree and non-tree decoding methods. We release all code and datasets to advance future research.
Authors: Md Tahmid Rahman Laskar, Mohammed Saidul Islam, Ridwan Mahbub, Mizanur Rahman, Amran Bhuiyan, Israt Jahan, Mir Tafseer Nayeem, Shafiq Joty, Enamul Hoque, Jimmy Huang
Abstract: Large Vision-Language Models (LVLMs) with only 7B parameters have shown promise as automated judges in chart comprehension tasks. However, tiny models (<=2B parameters) still perform poorly as judges, limiting their real-world use in resource-constrained settings. To address this, we propose two approaches to ensure cost-efficient evaluation: (i) multi-criteria prompting, which combines separate evaluation criteria into a single query, and (ii) domain-adaptive transfer learning, in which we fine-tune a 2B-parameter LVLM on synthetic judgments in a chart dataset to create the ChartJudge. Experiments show that multi-criteria prompting exposes robustness gaps, which led to a huge drop in performance for 7B models, including specialized LVLM judges like LLaVA-Critic. In addition, we find that our tiny LVLM (ChartJudge) can effectively transfer knowledge from one dataset to another to make it a more specialized model. Our fine-grained analysis across chart types and query complexities offers actionable insights into trade-offs between model size, prompt design, and transferability, enabling scalable, low-cost evaluation for chart reasoning tasks. Our code and the data will be made publicly available.
Authors: Junyi Zhu, Savas Ozkan, Andrea Maracani, Sinan Mutlu, Cho Jung Min, Mete Ozay
Abstract: Deploying natural language processing (NLP) models on mobile platforms requires models that can adapt across diverse applications while remaining efficient in memory and computation. We investigate pre-finetuning strategies to enhance the adaptability of lightweight BERT-like encoders for two fundamental NLP task families: named entity recognition (NER) and text classification. While pre-finetuning improves downstream performance for each task family individually, we find that na\"ive multi-task pre-finetuning introduces conflicting optimization signals that degrade overall performance. To address this, we propose a simple yet effective multi-task pre-finetuning framework based on task-primary LoRA modules, which enables a single shared encoder backbone with modular adapters. Our approach achieves performance comparable to individual pre-finetuning while meeting practical deployment constraint. Experiments on 21 downstream tasks show average improvements of +0.8% for NER and +8.8% for text classification, demonstrating the effectiveness of our method for versatile mobile NLP applications.
Authors: Mkululi Sikosana, Sean Maudsley-Barton, Oluwaseun Ajao
Abstract: This study conducts a computational linguistic analysis of pandemic-related online discourse to examine how language distinguishes health misinformation from factual communication. Drawing on three corpora: COVID-19 false narratives (n = 7588), general COVID-19 content (n = 10700), and Monkeypox-related posts (n = 5787), we identify significant differences in readability, rhetorical markers, and persuasive language use. COVID-19 misinformation exhibited markedly lower readability scores and contained over twice the frequency of fear-related or persuasive terms compared to the other datasets. It also showed minimal use of exclamation marks, contrasting with the more emotive style of Monkeypox content. These patterns suggest that misinformation employs a deliberately complex rhetorical style embedded with emotional cues, a combination that may enhance its perceived credibility. Our findings contribute to the growing body of work on digital health misinformation by highlighting linguistic indicators that may aid detection efforts. They also inform public health messaging strategies and theoretical models of crisis communication in networked media environments. At the same time, the study acknowledges limitations, including reliance on traditional readability indices, use of a deliberately narrow persuasive lexicon, and reliance on static aggregate analysis. Future research should therefore incorporate longitudinal designs, broader emotion lexicons, and platform-sensitive approaches to strengthen robustness.
Authors: Chihiro Taguchi, Richard Sproat
Abstract: We present a system that uses LLMs as a tool in the development of Constructed Languages. The system is modular in that one first creates a target phonology for the language using an agentic approach that refines its output at each step with commentary feedback on its previous attempt. Next, a set of sentences is 'translated' from their English original into a morphosyntactic markup that reflects the word order and morphosyntactic feature specifications of the desired target language, with affixes represented as morphosyntactic feature bundles. From this translated corpus, a lexicon is constructed using the phonological model and the set of morphemes (stems and affixes) extracted from the 'translated' sentences. The system is then instructed to provide an orthography for the language, using an existing script such as Latin or Cyrillic. Finally, the system writes a brief grammatical handbook of the language. The system can also translate further sentences into the target language. Our goal is twofold. First, we hope that these tools will be fun to use for creating artificially constructed languages. Second, we are interested in exploring what LLMs 'know' about language-not what they know about any particular language or linguistic phenomenon, but how much they know about and understand language and linguistic concepts. As we shall see, there is a fairly wide gulf in capabilities both among different LLMs and among different linguistic specifications, with it being notably easier for systems to deal with more common patterns than rarer ones. An additional avenue that we explore is the application of our approach to translating from high-resource into low-resource languages. While the results so far are mostly negative, we provide some evidence that an improved version of the present system could afford some real gains in such tasks. https://github.com/SakanaAI/IASC
Authors: Isabel Papadimitriou, Jacob Prince
Abstract: Large language models (LLMs) work by manipulating the geometry of input embedding vectors over multiple layers. Here, we ask: how are the input vocabulary representations of language models structured, and how and when does this structure evolve over training? To answer this question, we use representational similarity analysis, running a suite of experiments that correlate the geometric structure of the input embeddings and output embeddings of two open-source models (Pythia 12B and OLMo 7B) with semantic, syntactic, and frequency-based metrics over the course of training. Our key findings are as follows: 1) During training, the vocabulary embedding geometry quickly converges to high correlations with a suite of semantic and syntactic features; 2) Embeddings of high-frequency and function words (e.g., "the," "of") converge to their final vectors faster than lexical and low-frequency words, which retain some alignment with the bias in their random initializations. These findings help map the dynamic trajectory by which input embeddings organize around linguistic structure, revealing distinct roles for word frequency and function. Our findings motivate a deeper study of how the evolution of vocabulary geometry may facilitate specific capability gains during model training.
Authors: Zhangdie Yuan, Han-Chin Shing, Mitch Strong, Chaitanya Shivade
Abstract: Accurate clinical coding is essential for healthcare documentation, billing, and decision-making. While prior work shows that off-the-shelf LLMs struggle with this task, evaluations based on exact match metrics often overlook errors where predicted codes are hierarchically close but incorrect. Our analysis reveals that such hierarchical misalignments account for a substantial portion of LLM failures. We show that lightweight interventions, including prompt engineering and small-scale fine-tuning, can improve accuracy without the computational overhead of search-based methods. To address hierarchically near-miss errors, we introduce clinical code verification as both a standalone task and a pipeline component. To mitigate the limitations in existing datasets, such as incomplete evidence and inpatient bias in MIMIC, we release an expert double-annotated benchmark of outpatient clinical notes with ICD-10 codes. Our results highlight verification as an effective and reliable step toward improving LLM-based medical coding.
Authors: {\DJ}or{\dj}e Klisura, Joseph Khoury, Ashish Kundu, Ram Krishnan, Anthony Rios
Abstract: Access control is a cornerstone of secure computing, yet large language models often blur role boundaries by producing unrestricted responses. We study role-conditioned refusals, focusing on the LLM's ability to adhere to access control policies by answering when authorized and refusing when not. To evaluate this behavior, we created a novel dataset that extends the Spider and BIRD text-to-SQL datasets, both of which have been modified with realistic PostgreSQL role-based policies at the table and column levels. We compare three designs: (i) zero or few-shot prompting, (ii) a two-step generator-verifier pipeline that checks SQL against policy, and (iii) LoRA fine-tuned models that learn permission awareness directly. Across multiple model families, explicit verification (the two-step framework) improves refusal precision and lowers false permits. At the same time, fine-tuning achieves a stronger balance between safety and utility (i.e., when considering execution accuracy). Longer and more complex policies consistently reduce the reliability of all systems. We release RBAC-augmented datasets and code.
Authors: Xin Jie Chua, Jeraelyn Ming Li Tan, Jia Xuan Tan, Soon Chang Poh, Yi Xian Goh, Debbie Hui Tian Choong, Chee Mun Foong, Sze Jue Yang, Chee Seng Chan
Abstract: This paper presents Ryt AI, an LLM-native agentic framework that powers Ryt Bank to enable customers to execute core financial transactions through natural language conversation. This represents the first global regulator-approved deployment worldwide where conversational AI functions as the primary banking interface, in contrast to prior assistants that have been limited to advisory or support roles. Built entirely in-house, Ryt AI is powered by ILMU, a closed-source LLM developed internally, and replaces rigid multi-screen workflows with a single dialogue orchestrated by four LLM-powered agents (Guardrails, Intent, Payment, and FAQ). Each agent attaches a task-specific LoRA adapter to ILMU, which is hosted within the bank's infrastructure to ensure consistent behavior with minimal overhead. Deterministic guardrails, human-in-the-loop confirmation, and a stateless audit architecture provide defense-in-depth for security and compliance. The result is Banking Done Right: demonstrating that regulator-approved natural-language interfaces can reliably support core financial operations under strict governance.
Authors: Yuzhe Gu, Xiyu Liang, Jiaojiao Zhao, Enmao Diao
Abstract: Large language models (LLMs) with extended context windows enable powerful downstream applications but impose significant memory overhead, as caching all key-value (KV) states scales linearly with sequence length and batch size. Existing cache eviction methods address this by exploiting attention sparsity, yet they typically rank tokens heuristically using accumulated attention weights without considering their true impact on attention outputs. We propose Optimal Brain Cache (OBCache), a principled framework that formulates cache eviction as a layer-wise structured pruning problem. Building upon the Optimal Brain Damage (OBD) theory, OBCache quantifies token saliency by measuring the perturbation in attention outputs induced by pruning tokens, with closed-form scores derived for isolated keys, isolated values, and joint key-value pairs. Our scores account not only for attention weights but also for information from value states and attention outputs, thereby enhancing existing eviction strategies with output-aware signals. Experiments on LLaMA and Qwen models demonstrate that replacing the heuristic scores in existing works, which estimate token saliency across different query positions, with OBCache's output-aware scores consistently improves long-context accuracy.
Authors: Virginia K. Felkner, Allison Lim, Jonathan May
Abstract: Measurement of social bias in language models is typically by token probability (TP) metrics, which are broadly applicable but have been criticized for their distance from real-world langugage model use cases and harms. In this work, we test natural language inference (NLI) as a more realistic alternative bias metric. We show that, curiously, NLI and TP bias evaluation behave substantially differently, with very low correlation among different NLI metrics and between NLI and TP metrics. We find that NLI metrics are more likely to detect "underdebiased" cases. However, NLI metrics seem to be more brittle and sensitive to wording of counterstereotypical sentences than TP approaches. We conclude that neither token probability nor natural language inference is a "better" bias metric in all cases, and we recommend a combination of TP, NLI, and downstream bias evaluations to ensure comprehensive evaluation of language models. Content Warning: This paper contains examples of anti-LGBTQ+ stereotypes.
Authors: Jifan Zhang, Henry Sleight, Andi Peng, John Schulman, Esin Durmus
Abstract: Large language models (LLMs) are increasingly trained from AI constitutions and model specifications that establish behavioral guidelines and ethical principles. However, these specifications face critical challenges, including internal conflicts between principles and insufficient coverage of nuanced scenarios. We present a systematic methodology for stress-testing model character specifications, automatically identifying numerous cases of principle contradictions and interpretive ambiguities in current model specs. We stress test current model specs by generating scenarios that force explicit tradeoffs between competing value-based principles. Using a comprehensive taxonomy we generate diverse value tradeoff scenarios where models must choose between pairs of legitimate principles that cannot be simultaneously satisfied. We evaluate responses from twelve frontier LLMs across major providers (Anthropic, OpenAI, Google, xAI) and measure behavioral disagreement through value classification scores. Among these scenarios, we identify over 70,000 cases exhibiting significant behavioral divergence. Empirically, we show this high divergence in model behavior strongly predicts underlying problems in model specifications. Through qualitative analysis, we provide numerous example issues in current model specs such as direct contradiction and interpretive ambiguities of several principles. Additionally, our generated dataset also reveals both clear misalignment cases and false-positive refusals across all of the frontier models we study. Lastly, we also provide value prioritization patterns and differences of these models.
Authors: Krinos Li, Xianglu Xiao, Shenglong Deng, Lucas He, Zijun Zhong, Yuanjie Zou, Zhonghao Zhan, Zheng Hui, Weiye Bao, Guang Yang
Abstract: Large language models (LLMs) are transforming cellular biology by enabling the development of "virtual cells"--computational systems that represent, predict, and reason about cellular states and behaviors. This work provides a comprehensive review of LLMs for virtual cell modeling. We propose a unified taxonomy that organizes existing methods into two paradigms: LLMs as Oracles, for direct cellular modeling, and LLMs as Agents, for orchestrating complex scientific tasks. We identify three core tasks--cellular representation, perturbation prediction, and gene regulation inference--and review their associated models, datasets, evaluation benchmarks, as well as the critical challenges in scalability, generalizability, and interpretability.
Authors: Chengshuai Zhao, Shu Wan, Paras Sheth, Karan Patwa, K. Sel\c{c}uk Candan, Huan Liu
Abstract: The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language -- making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.
Authors: Shuo Yu, Mingyue Cheng, Daoyu Wang, Qi Liu, Zirui Liu, Ze Guo, Xiaoyu Tao
Abstract: The primary form of user-internet engagement is shifting from leveraging implicit feedback signals, such as browsing and clicks, to harnessing the rich explicit feedback provided by textual interactive behaviors. This shift unlocks a rich source of user textual history, presenting a profound opportunity for a deeper form of personalization. However, prevailing approaches offer only a shallow form of personalization, as they treat user history as a flat list of texts for retrieval and fail to model the rich temporal and semantic structures reflecting dynamic nature of user interests. In this work, we propose \textbf{MemWeaver}, a framework that weaves the user's entire textual history into a hierarchical memory to power deeply personalized generation. The core innovation of our memory lies in its ability to capture both the temporal evolution of interests and the semantic relationships between different activities. To achieve this, MemWeaver builds two complementary memory components that both integrate temporal and semantic information, but at different levels of abstraction: behavioral memory, which captures specific user actions, and cognitive memory, which represents long-term preferences. This dual-component memory serves as a unified representation of the user, allowing large language models (LLMs) to reason over both concrete behaviors and abstracted traits. Experiments on the Language Model Personalization (LaMP) benchmark validate the efficacy of MemWeaver. Our code is available\footnote{https://github.com/fishsure/MemWeaver}.
Authors: Jiaoyang Li, Junhao Ruan, Shengwei Tang, Saihan Chen, Kaiyan Chang, Yuan Ge, Tong Xiao, Jingbo Zhu
Abstract: Graph Retrieval-Augmented Generation (Graph RAG) effectively builds a knowledge graph (KG) to connect disparate facts across a large document corpus. However, this broad-view approach often lacks the deep structured reasoning needed for complex multi-hop question answering (QA), leading to incomplete evidence and error accumulation. To address these limitations, we propose SubQRAG, a sub-question-driven framework that enhances reasoning depth. SubQRAG decomposes a complex question into an ordered chain of verifiable sub-questions. For each sub-question, it retrieves relevant triples from the graph. When the existing graph is insufficient, the system dynamically expands it by extracting new triples from source documents in real time. All triples used in the reasoning process are aggregated into a "graph memory," forming a structured and traceable evidence path for final answer generation. Experiments on three multi-hop QA benchmarks demonstrate that SubQRAG achieves consistent and significant improvements, especially in Exact Match scores.
Authors: Cunli Mao, Xiaofei Gao, Ran Song, Shizhu He, Shengxiang Gao, Kang Liu, Zhengtao Yu
Abstract: Large language models (LLMs) based Multilingual Knowledge Graph Completion (MKGC) aim to predict missing facts by leveraging LLMs' multilingual understanding capabilities, improving the completeness of multilingual knowledge graphs (KGs). However, existing MKGC research underutilizes the multilingual capabilities of LLMs and ignores the shareability of cross-lingual knowledge. In this paper, we propose a novel MKGC framework that leverages multilingual shared knowledge to significantly enhance performance through two components: Knowledge-level Grouped Mixture of Experts (KL-GMoE) and Iterative Entity Reranking (IER). KL-GMoE efficiently models shared knowledge, while IER significantly enhances its utilization. To evaluate our framework, we constructed a mKG dataset containing 5 languages and conducted comprehensive comparative experiments with existing state-of-the-art (SOTA) MKGC method. The experimental results demonstrate that our framework achieves improvements of 5.47%, 3.27%, and 1.01% in the Hits@1, Hits@3, and Hits@10 metrics, respectively, compared with SOTA MKGC method. Further experimental analysis revealed the properties of knowledge sharing in settings of unseen and unbalanced languages. We have released the dataset and code for our work on https://github.com/gaoxiaofei07/KL-GMoE.
Authors: Fu Chen, Peng Wang, Xiyin Li, Wen Li, Shichi Lei, Dongdong Xiang
Abstract: Training Large Language Models (LLMs) with Group Relative Policy Optimization (GRPO) encounters a significant challenge: models often fail to produce accurate responses, particularly in small-scale architectures. This limitation not only diminishes performance improvements and undermines the potential of GRPO but also frequently leads to mid-training collapse, adversely affecting stability and final efficacy. To address these issues, we propose ToolExpander, a novel framework that advances tool-oriented reinforcement learning for resource-constrained LLMs through two key innovations:(1) Dynamic Multi-Round Hard Sampling, which dynamically substitutes challenging samples(those without correct outputs over 10 rollouts) with high-quality few-shot demonstrations during training, coupled with an exponential learning rate decay strategy to mitigate oscillations;(2) Self-Exemplifying Thinking, an enhanced GRPO framework that eliminates KL divergence and incorporates adjusted clipping coefficients, encouraging models to autonomously generate and analyze few-shot examples via a minimal additional reward (0.01).Experimental results demonstrate that ToolExpander significantly enhances tool-using capabilities in LLMs, especially in weaker small-scale models, improving both training stability and overall performance.
Authors: Tianci Liu, Ran Xu, Tony Yu, Ilgee Hong, Carl Yang, Tuo Zhao, Haoyu Wang
Abstract: Reward modeling lies at the core of reinforcement learning from human feedback (RLHF), yet most existing reward models rely on scalar or pairwise judgments that fail to capture the multifaceted nature of human preferences. Recent studies have explored rubrics-as-rewards (RaR) that uses structured natural language criteria that capture multiple dimensions of response quality. However, producing rubrics that are both reliable and scalable remains a key challenge. In this work, we introduce OpenRubrics, a diverse, large-scale collection of (prompt, rubric) pairs for training rubric-generation and rubric-based reward models. To elicit discriminative and comprehensive evaluation signals, we introduce Contrastive Rubric Generation (CRG), which derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses. We further improve reliability by enforcing preference-label consistency via rejection sampling to remove noisy rubrics. Across multiple reward-modeling benchmarks, our rubric-based reward model, Rubric-RM, surpasses strong size-matched baselines by 6.8%. These gains transfer to policy models on instruction-following and biomedical benchmarks. Our results show that rubrics provide scalable alignment signals that narrow the gap between costly human evaluation and automated reward modeling, enabling a new principle-driven paradigm for LLM alignment.
Authors: Runyang You, Yongqi Li, Meng Liu, Wenjie Wang, Liqiang Nie, Wenjie Li
Abstract: Parallel test-time scaling (TTS) is a pivotal approach for enhancing large language models (LLMs), typically by sampling multiple token-based chains-of-thought in parallel and aggregating outcomes through voting or search. Recent advances in latent reasoning, where intermediate reasoning unfolds in continuous vector spaces, offer a more efficient alternative to explicit Chain-of-Thought, yet whether such latent models can similarly benefit from parallel TTS remains open, mainly due to the absence of sampling mechanisms in continuous space, and the lack of probabilistic signals for advanced trajectory aggregation. \ This work enables parallel TTS for latent reasoning models by addressing the above issues. For sampling, we introduce two uncertainty-inspired stochastic strategies: Monte Carlo Dropout and Additive Gaussian Noise. For aggregation, we design a Latent Reward Model (LatentRM) trained with step-wise contrastive objective to score and guide latent reasoning. Extensive experiments and visualization analyses show that both sampling strategies scale effectively with compute and exhibit distinct exploration dynamics, while LatentRM enables effective trajectory selection. Together, our explorations open a new direction for scalable inference in continuous spaces. Code released at https://github.com/YRYangang/LatentTTS.
Authors: Nishant Balepur, Atrey Desai, Rachel Rudinger
Abstract: Large language models (LLMs) now give reasoning before answering, excelling in tasks like multiple-choice question answering (MCQA). Yet, a concern is that LLMs do not solve MCQs as intended, as work finds LLMs sans reasoning succeed in MCQA without using the question, i.e., choices-only. Such partial-input success is often deemed problematic, but reasoning traces could reveal if these strategies are truly shallow in choices-only settings. To study these strategies, reasoning LLMs solve MCQs in full and choices-only inputs; test-time reasoning often boosts accuracy on full and in choices-only half the time. While possibly due to shallow shortcuts, choices-only success is barely affected by the length of reasoning traces, and after finding traces pass faithfulness tests, we show they use less problematic strategies like inferring missing questions. In all, we challenge claims that partial-input success is always a flaw, so we discuss how reasoning traces could separate problematic data from less problematic reasoning.
Authors: Murong Yue, Zhiwei Liu, Liangwei Yang, Jianguo Zhang, Zuxin Liu, Haolin Chen, Ziyu Yao, Silvio Savarese, Caiming Xiong, Shelby Heinecke, Huan Wang
Abstract: Large Language Models (LLMs) equipped with external tools have demonstrated enhanced performance on complex reasoning tasks. The widespread adoption of this tool-augmented reasoning is hindered by the scarcity of domain-specific tools. For instance, in domains such as physics question answering, suitable and specialized tools are often missing. Recent work has explored automating tool creation by extracting reusable functions from Chain-of-Thought (CoT) reasoning traces; however, these approaches face a critical scalability bottleneck. As the number of generated tools grows, storing them in an unstructured collection leads to significant retrieval challenges, including an expanding search space and ambiguity between function-related tools. To address this, we propose a systematic approach to automatically refactor an unstructured collection of tools into a structured tool library. Our system first generates discrete, task-specific tools and clusters them into semantically coherent topics. Within each cluster, we introduce a multi-agent framework to consolidate scattered functionalities: a code agent refactors code to extract shared logic and creates versatile, aggregated tools, while a reviewing agent ensures that these aggregated tools maintain the complete functional capabilities of the original set. This process transforms numerous question-specific tools into a smaller set of powerful, aggregated tools without loss of functionality. Experimental results demonstrate that our approach significantly improves tool retrieval accuracy and overall reasoning performance across multiple reasoning tasks. Furthermore, our method shows enhanced scalability compared with baselines as the number of question-specific increases.
Authors: Youliang Yuan, Qiuyang Mang, Jingbang Chen, Hong Wan, Xiaoyuan Liu, Junjielong Xu, Jen-tse Huang, Wenxuan Wang, Wenxiang Jiao, Pinjia He
Abstract: Large language models for mathematical reasoning are typically trained with outcome-based rewards, which credit only the final answer. In our experiments, we observe that this paradigm is highly susceptible to reward hacking, leading to a substantial overestimation of a model's reasoning ability. This is evidenced by a high incidence of false positives - solutions that reach the correct final answer through an unsound reasoning process. Through a systematic analysis with human verification, we establish a taxonomy of these failure modes, identifying patterns like Miracle Steps - abrupt jumps to a correct output without a valid preceding derivation. Probing experiments suggest a strong association between these Miracle Steps and memorization, where the model appears to recall the answer directly rather than deriving it. To mitigate this systemic issue, we introduce the Rubric Reward Model (RRM), a process-oriented reward function that evaluates the entire reasoning trajectory against problem-specific rubrics. The generative RRM provides fine-grained, calibrated rewards (0-1) that explicitly penalize logical flaws and encourage rigorous deduction. When integrated into a reinforcement learning pipeline, RRM-based training consistently outperforms outcome-only supervision across four math benchmarks. Notably, it boosts Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. Our work demonstrates that rewarding the solution process is crucial for building models that are not only more accurate but also more reliable.
Authors: Omar Mahmoud, Ali Khalil, Buddhika Laknath Semage, Thommen George Karimpanal, Santu Rana
Abstract: Hallucination in large language models (LLMs) has been widely studied in recent years, with progress in both detection and mitigation aimed at improving truthfulness. Yet, a critical side effect remains largely overlooked: enhancing truthfulness can negatively impact safety alignment. In this paper, we investigate this trade-off and show that increasing factual accuracy often comes at the cost of weakened refusal behavior. Our analysis reveals that this arises from overlapping components in the model that simultaneously encode hallucination and refusal information, leading alignment methods to suppress factual knowledge unintentionally. We further examine how fine-tuning on benign datasets, even when curated for safety, can degrade alignment for the same reason. To address this, we propose a method that disentangles refusal-related features from hallucination features using sparse autoencoders, and preserves refusal behavior during fine-tuning through subspace orthogonalization. This approach prevents hallucinations from increasing while maintaining safety alignment.We evaluate our method on commonsense reasoning tasks and harmful benchmarks (AdvBench and StrongReject). Results demonstrate that our approach preserves refusal behavior and task utility, mitigating the trade-off between truthfulness and safety.
Authors: Shiman Zhao, Shangyuan Li, Wei Chen, Tengjiao Wang, Jiahui Yao, Jiabin Zheng, Kam Fai Wong
Abstract: Few-shot Multi-label Intent Detection (MID) is crucial for dialogue systems, aiming to detect multiple intents of utterances in low-resource dialogue domains. Previous studies focus on a two-stage pipeline. They first learn representations of utterances with multiple labels and then use a threshold-based strategy to identify multi-label results. However, these methods rely on representation classification and ignore instance relations, leading to error propagation. To solve the above issues, we propose a multi-label joint learning method for few-shot MID in an end-to-end manner, which constructs an instance relation learning network with label knowledge propagation to eliminate error propagation. Concretely, we learn the interaction relations between instances with class information to propagate label knowledge between a few labeled (support set) and unlabeled (query set) instances. With label knowledge propagation, the relation strength between instances directly indicates whether two utterances belong to the same intent for multi-label prediction. Besides, a dual relation-enhanced loss is developed to optimize support- and query-level relation strength to improve performance. Experiments show that we outperform strong baselines by an average of 9.54% AUC and 11.19% Macro-F1 in 1-shot scenarios.
Authors: Vardhan Dongre, Ryan A. Rossi, Viet Dac Lai, David Seunghyun Yoon, Dilek Hakkani-T\"ur, Trung Bui
Abstract: Large Language Models (LLMs) excel at single-turn tasks such as instruction following and summarization, yet real-world deployments require sustained multi-turn interactions where user goals and conversational context persist and evolve. A recurring challenge in this setting is context drift: the gradual divergence of a model's outputs from goal-consistent behavior across turns. Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics. In this work, we present a study of context drift in multi-turn interactions and propose a simple dynamical framework to interpret its behavior. We formalize drift as the turn-wise KL divergence between the token-level predictive distributions of the test model and a goal-consistent reference model, and propose a recurrence model that interprets its evolution as a bounded stochastic process with restoring forces and controllable interventions. We instantiate this framework in both synthetic long-horizon rewriting tasks and realistic user-agent simulations such as in $\tau$-Bench, measuring drift for several open-weight LLMs that are used as user simulators. Our experiments consistently reveal stable, noise-limited equilibria rather than runaway degradation, and demonstrate that simple reminder interventions reliably reduce divergence in line with theoretical predictions. Together, these results suggest that multi-turn drift can be understood as a controllable equilibrium phenomenon rather than as inevitable decay, providing a foundation for studying and mitigating context drift in extended interactions.
Authors: Shuichiro Haruta, Kazunori Matsumoto, Zhi Li, Yanan Wang, Mori Kurokawa
Abstract: In this paper, we propose a rotation-constrained compensation method to address the errors introduced by structured pruning of large language models (LLMs). LLMs are trained on massive datasets and accumulate rich semantic knowledge in their representation space. In contrast, pruning is typically carried out with only a small amount of calibration data, which makes output mismatches unavoidable. Although direct least-squares fitting can reduce such errors, it tends to overfit to the limited calibration set, destructively modifying pretrained weights. To overcome this difficulty, we update the pruned parameters under a rotation constraint. This constrained update preserves the geometry of output representations (i.e., norms and inner products) and simultaneously re-aligns the pruned subspace with the original outputs. Furthermore, in rotation-constrained compensation, removing components that strongly contribute to the principal directions of the output makes error recovery difficult. Since input dimensions with large variance strongly affect these principal directions, we design a variance-aware importance score that ensures such dimensions are preferentially kept in the pruned model. By combining this scoring rule with rotation-constrained updates, the proposed method effectively compensates errors while retaining the components likely to be more important in a geometry-preserving manner. In the experiments, we apply the proposed method to LLaMA-7B and evaluate it on WikiText-2 and multiple language understanding benchmarks. The results demonstrate consistently better perplexity and task accuracy compared with existing baselines.
Authors: Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul, Uddip Acharjee Shuvo, Muhit Islam Emon, Xuan Wang, Liqing Zhang
Abstract: Large language models (LLMs) and emerging agentic frameworks are beginning to transform single-cell biology by enabling natural-language reasoning, generative annotation, and multimodal data integration. However, progress remains fragmented across data modalities, architectures, and evaluation standards. LLM4Cell presents the first unified survey of 58 foundation and agentic models developed for single-cell research, spanning RNA, ATAC, multi-omic, and spatial modalities. We categorize these methods into five families-foundation, text-bridge, spatial, multimodal, epigenomic, and agentic-and map them to eight key analytical tasks including annotation, trajectory and perturbation modeling, and drug-response prediction. Drawing on over 40 public datasets, we analyze benchmark suitability, data diversity, and ethical or scalability constraints, and evaluate models across 10 domain dimensions covering biological grounding, multi-omics alignment, fairness, privacy, and explainability. By linking datasets, models, and evaluation domains, LLM4Cell provides the first integrated view of language-driven single-cell intelligence and outlines open challenges in interpretability, standardization, and trustworthy model development.
Authors: Peilin Wu, Mian Zhang, Kun Wan, Wentian Zhao, Kaiyu He, Xinya Du, Zhiyu Chen
Abstract: Agentic RAG is a powerful technique for incorporating external information that LLMs lack, enabling better problem solving and question answering. However, suboptimal search behaviors exist widely, such as over-search (retrieving information already known) and under-search (failing to search when necessary), which leads to unnecessary overhead and unreliable outputs. Current training methods, which typically rely on outcome-based rewards in a RL framework, lack the fine-grained control needed to address these inefficiencies. To overcome this, we introduce Hierarchical Process Rewards for Efficient agentic RAG (HiPRAG), a training methodology that incorporates a fine-grained, knowledge-grounded process reward into the RL training. Our approach evaluates the necessity of each search decision on-the-fly by decomposing the agent's reasoning trajectory into discrete, parsable steps. We then apply a hierarchical reward function that provides an additional bonus based on the proportion of optimal search and non-search steps, on top of commonly used outcome and format rewards. Experiments on the Qwen2.5 and Llama-3.2 models across seven diverse QA benchmarks show that our method achieves average accuracies of 65.4% (3B) and 67.2% (7B). This is accomplished while improving search efficiency, reducing the over-search rate to just 2.3% and concurrently lowering the under-search rate. These results demonstrate the efficacy of optimizing the reasoning process itself, not just the final outcome. Further experiments and analysis demonstrate that HiPRAG shows good generalizability across a wide range of RL algorithms, model families, sizes, and types. This work demonstrates the importance and potential of fine-grained control through RL, for improving the efficiency and optimality of reasoning for search agents.
Authors: Eric Hanchen Jiang, Guancheng Wan, Sophia Yin, Mengting Li, Yuchen Wu, Xiao Liang, Xinfeng Li, Yizhou Sun, Wei Wang, Kai-Wei Chang, Ying Nian Wu
Abstract: The efficiency of multi-agent systems driven by large language models (LLMs) largely hinges on their communication topology. However, designing an optimal topology is a non-trivial challenge, as it requires balancing competing objectives such as task performance, communication cost, and robustness. Existing frameworks often rely on static or hand-crafted topologies, which inherently fail to adapt to diverse task requirements, leading to either excessive token consumption for simple problems or performance bottlenecks for complex ones. To address this challenge, we introduce a novel generative framework called \textit{Guided Topology Diffusion (GTD)}. Inspired by conditional discrete graph diffusion models, GTD formulates topology synthesis as an iterative construction process. At each step, the generation is steered by a lightweight proxy model that predicts multi-objective rewards (e.g., accuracy, utility, cost), enabling real-time, gradient-free optimization towards task-adaptive topologies. This iterative, guided synthesis process distinguishes GTD from single-step generative frameworks, enabling it to better navigate complex design trade-offs. We validated GTD across multiple benchmarks, and experiments show that this framework can generate highly task-adaptive, sparse, and efficient communication topologies, significantly outperforming existing methods in LLM agent collaboration.
Authors: Yuxin Huang, Simeng Wu, Ran Song, Yan Xiang, Yantuan Xian, Shengxiang Gao, Zhengtao Yu
Abstract: Generative Information Retrieval is an emerging retrieval paradigm that exhibits remarkable performance in monolingual scenarios.However, applying these methods to multilingual retrieval still encounters two primary challenges, cross-lingual identifier misalignment and identifier inflation. To address these limitations, we propose Multilingual Generative Retrieval via Cross-lingual Semantic Compression (MGR-CSC), a novel framework that unifies semantically equivalent multilingual keywords into shared atoms to align semantics and compresses the identifier space, and we propose a dynamic multi-step constrained decoding strategy during retrieval. MGR-CSC improves cross-lingual alignment by assigning consistent identifiers and enhances decoding efficiency by reducing redundancy. Experiments demonstrate that MGR-CSC achieves outstanding retrieval accuracy, improving by 6.83% on mMarco100k and 4.77% on mNQ320k, while reducing document identifiers length by 74.51% and 78.2%, respectively.
Authors: Jingyu Peng, Maolin Wang, Hengyi Cai, Yuchen Li, Kai Zhang, Shuaiqiang Wang, Dawei Yin, Xiangyu Zhao
Abstract: Small language models (SLMs) are crucial for applications with strict latency and computational constraints, yet achieving high performance remains challenging. Knowledge distillation (KD) can transfer capabilities from large teacher models, but existing methods involve trade-offs: off-policy distillation provides high-quality supervision but introduces a training-inference mismatch, while on-policy approaches maintain consistency but rely on low-quality student outputs. To address these issues, we propose AdaSwitch, a novel approach that dynamically combines on-policy and off-policy generation at the token level. AdaSwitch allows the student to first explore its own predictions and then selectively integrate teacher guidance based on real-time quality assessment. This approach simultaneously preserves consistency and maintains supervision quality. Experiments on three datasets with two teacher-student LLM pairs demonstrate that AdaSwitch consistently improves accuracy, offering a practical and effective method for distilling SLMs with acceptable additional overhead.
Authors: Md. Faiyaz Abdullah Sayeedi, Md. Mahbub Alam, Subhey Sadi Rahman, Md. Adnanul Islam, Jannatul Ferdous Deepti, Tasnim Mohiuddin, Md Mofijul Islam, Swakkhar Shatabda
Abstract: The rise of Large Language Models (LLMs) has redefined Machine Translation (MT), enabling context-aware and fluent translations across hundreds of languages and textual domains. Despite their remarkable capabilities, LLMs often exhibit uneven performance across language families and specialized domains. Moreover, recent evidence reveals that these models can encode and amplify different biases present in their training data, posing serious concerns for fairness, especially in low-resource languages. To address these gaps, we introduce Translation Tangles, a unified framework and dataset for evaluating the translation quality and fairness of open-source LLMs. Our approach benchmarks 24 bidirectional language pairs across multiple domains using different metrics. We further propose a hybrid bias detection pipeline that integrates rule-based heuristics, semantic similarity filtering, and LLM-based validation. We also introduce a high-quality, bias-annotated dataset based on human evaluations of 1,439 translation-reference pairs. The code and dataset are accessible on GitHub: https://github.com/faiyazabdullah/TranslationTangles
Authors: Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, Shyam Upadhyay
Abstract: Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency -- overthinking -- models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models -- Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.
Authors: Heyang Liu, Yuhao Wang, Ziyang Cheng, Ronghua Wu, Qunshan Gu, Yanfeng Wang, Yu Wang
Abstract: The advancement of multimodal large language models has accelerated the development of speech-to-speech interaction systems. While natural monolingual interaction has been achieved, we find existing models exhibit deficiencies in language alignment. In our proposed Code-Switching Speech-to-Speech Benchmark (CS3-Bench), experiments on 7 mainstream models demonstrate a relative performance drop of up to 66% in knowledge-intensive question answering and varying degrees of misunderstanding in open-ended conversations. Starting from a model with severe performance deterioration, we propose both data constructions and training approaches to improve the language alignment capabilities, specifically employing Chain of Recognition (CoR) to enhance understanding and Keyword Highlighting (KH) to guide generation. Our approach improves the knowledge accuracy from 25.14% to 46.13%, with open-ended understanding rate from 64.5% to 86.5%, and significantly reduces pronunciation errors in the secondary language. CS3-Bench is available at https://huggingface.co/datasets/VocalNet/CS3-Bench.
Authors: Houcheng Jiang, Junfeng Fang, Jiaxin Wu, Tianyu Zhang, Chen Gao, Yong Li, Xiang Wang, Xiangnan He, Yang Deng
Abstract: Weak-to-strong generalization provides a promising paradigm for scaling large language models (LLMs) by training stronger models on samples from aligned weaker ones, without requiring human feedback or explicit reward modeling. However, its robustness and generalization are hindered by the noise and biases in weak-model outputs, which limit its applicability in practice. To address this challenge, we leverage implicit rewards, which approximate explicit rewards through log-likelihood ratios, and reveal their structural equivalence with Contrastive Decoding (CD), a decoding strategy shown to reduce noise in LLM generation. Building on this connection, we propose Contrastive Weak-to-Strong Generalization (ConG), a framework that employs contrastive decoding between pre- and post-alignment weak models to generate higher-quality samples. This approach enables more reliable capability transfer, denoising, and improved robustness, substantially mitigating the limitations of traditional weak-to-strong methods. Empirical results across different model families confirm consistent improvements, demonstrating the generality and effectiveness of ConG. Taken together, our findings highlight the potential of ConG to advance weak-to-strong generalization and provide a promising pathway toward AGI.
Authors: Verena Blaschke, Miriam Winkler, Barbara Plank
Abstract: Research on cross-dialectal transfer from a standard to a non-standard dialect variety has typically focused on text data. However, dialects are primarily spoken, and non-standard spellings are known to cause issues in text processing. We compare standard-to-dialect transfer in three settings: text models, speech models, and cascaded systems where speech first gets automatically transcribed and then further processed by a text model. In our experiments, we focus on German and multiple German dialects in the context of written and spoken intent and topic classification. To that end, we release the first dialectal audio intent classification dataset. We find that the speech-only setup provides the best results on the dialect data while the text-only setup works best on the standard data. While the cascaded systems lag behind the text-only models for German, they perform relatively well on the dialectal data if the transcription system generates normalized, standard-like output.
Authors: Hyeonseok Moon, Seongtae Hong, Jaehyung Seo, Heuiseok Lim
Abstract: Recent frontier-level LLMs have saturated many previously difficult benchmarks, leaving little room for further differentiation. This progress highlights the need for challenging benchmarks that provide objective verification. In this paper, we introduce MCBench, a benchmark designed to evaluate whether LLMs can execute string-matching NLP metrics by strictly following step-by-step instructions. Unlike prior benchmarks that depend on subjective judgments or general reasoning, MCBench offers an objective, deterministic and codeverifiable evaluation. This setup allows us to systematically test whether LLMs can maintain accurate step-by-step execution, including instruction adherence, numerical computation, and long-range consistency in handling intermediate results. To ensure objective evaluation of these abilities, we provide a parallel reference code that can evaluate the accuracy of LLM output. We provide three evaluative metrics and three benchmark variants designed to measure the detailed instruction understanding capability of LLMs. Our analyses show that MCBench serves as an effective and objective tool for evaluating the capabilities of cutting-edge LLMs.
Authors: Jiayu Yang, Yuxuan Fan, Songning Lai, Shengen Wu, Jiaqi Tang, Chun Kang, Zhijiang Guo, Yutao Yue
Abstract: Large Language Models (LLMs) require efficient knowledge editing (KE) to update factual information, yet existing methods exhibit significant performance decay in multi-hop factual recall. This failure is particularly acute when edits involve intermediate implicit subjects within reasoning chains. Through causal analysis, we reveal that this limitation stems from an oversight of how chained knowledge is dynamically represented and utilized at the neuron level. We discover that during multi hop reasoning, implicit subjects function as query neurons, which sequentially activate corresponding value neurons across transformer layers to accumulate information toward the final answer, a dynamic prior KE work has overlooked. Guided by this insight, we propose ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall, a framework that leverages neuron-level attribution to identify and edit these critical query-value (Q-V) pathways. ACE provides a mechanistically grounded solution for multi-hop KE, empirically outperforming state-of-the-art methods by 9.44% on GPT-J and 37.46% on Qwen3-8B. Our analysis further reveals more fine-grained activation patterns in Qwen3 and demonstrates that the semantic interpretability of value neurons is orchestrated by query-driven accumulation. These findings establish a new pathway for advancing KE capabilities based on the principled understanding of internal reasoning mechanisms.
Authors: Fanwei Zhua, Jiaxuan He, Xiaoxiao Chen, Zulong Chen, Quan Lu, Chenrui Mei
Abstract: Automatic grading of subjective questions remains a significant challenge in examination assessment due to the diversity in question formats and the open-ended nature of student responses. Existing works primarily focus on a specific type of subjective question and lack the generality to support comprehensive exams that contain diverse question types. In this paper, we propose a unified Large Language Model (LLM)-enhanced auto-grading framework that provides human-like evaluation for all types of subjective questions across various domains. Our framework integrates four complementary modules to holistically evaluate student answers. In addition to a basic text matching module that provides a foundational assessment of content similarity, we leverage the powerful reasoning and generative capabilities of LLMs to: (1) compare key knowledge points extracted from both student and reference answers, (2) generate a pseudo-question from the student answer to assess its relevance to the original question, and (3) simulate human evaluation by identifying content-related and non-content strengths and weaknesses. Extensive experiments on both general-purpose and domain-specific datasets show that our framework consistently outperforms traditional and LLM-based baselines across multiple grading metrics. Moreover, the proposed system has been successfully deployed in real-world training and certification exams at a major e-commerce enterprise.
Authors: Kyumin Lee, Minjin Jeon, Sanghwan Jang, Hwanjo Yu
Abstract: Answering complex real-world questions requires step-by-step retrieval and integration of relevant information to generate well-grounded responses. However, existing knowledge distillation methods overlook the need for different reasoning abilities at different steps, hindering transfer in multi-step retrieval-augmented frameworks. To address this, we propose Stepwise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented Language Models (StepER). StepER employs step-wise supervision to align with evolving information and reasoning demands across stages. Additionally, it incorporates difficulty-aware training to progressively optimize learning by prioritizing suitable steps. Our method is adaptable to various multi-step retrieval-augmented language models, including those that use retrieval queries for reasoning paths or decomposed questions. Extensive experiments show that StepER outperforms prior methods on multi-hop QA benchmarks, with an 8B model achieving performance comparable to a 70B teacher model.
Authors: Adam Dejl, James Barry, Alessandra Pascale, Javier Carnerero Cano
Abstract: Despite demonstrating remarkable performance across a wide range of tasks, large language models (LLMs) have also been found to frequently produce outputs that are incomplete or selectively omit key information. In sensitive domains, such omissions can result in significant harm comparable to that posed by factual inaccuracies, including hallucinations. In this study, we address the challenge of evaluating the comprehensiveness of LLM-generated texts, focusing on the detection of missing information or underrepresented viewpoints. We investigate three automated evaluation strategies: (1) an NLI-based method that decomposes texts into atomic statements and uses natural language inference (NLI) to identify missing links, (2) a Q&A-based approach that extracts question-answer pairs and compares responses across sources, and (3) an end-to-end method that directly identifies missing content using LLMs. Our experiments demonstrate the surprising effectiveness of the simple end-to-end approach compared to more complex methods, though at the cost of reduced robustness, interpretability and result granularity. We further assess the comprehensiveness of responses from several popular open-weight LLMs when answering user queries based on multiple sources.
Authors: Madis J\"urviste, Joonatan Jakobson
Abstract: This article presents research conducted at the Institute of the Estonian Language between 2022 and 2025 on the application of large language models (LLMs) to the study of 17th and 18th century Estonian dictionaries. The authors address three main areas: enriching historical dictionaries with modern word forms and meanings; using vision-enabled LLMs to perform text recognition on sources printed in Gothic script (Fraktur); and preparing for the creation of a unified, cross-source dataset. Initial experiments with J. Gutslaff's 1648 dictionary indicate that LLMs have significant potential for semi-automatic enrichment of dictionary information. When provided with sufficient context, Claude 3.7 Sonnet accurately provided meanings and modern equivalents for 81% of headword entries. In a text recognition experiment with A. T. Helle's 1732 dictionary, a zero-shot method successfully identified and structured 41% of headword entries into error-free JSON-formatted output. For digitising the Estonian-German dictionary section of A. W. Hupel's 1780 grammar, overlapping tiling of scanned image files is employed, with one LLM being used for text recognition and a second for merging the structured output. These findings demonstrate that even for minor languages LLMs have a significant potential for saving time and financial resources.
Authors: Fengji Zhang, Xinyao Niu, Chengyang Ying, Guancheng Lin, Zhongkai Hao, Zhou Fan, Chengen Huang, Jacky Keung, Bei Chen, Junyang Lin
Abstract: Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A$^2$Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed $\mathrm{AnsF1}$ reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A$^2$Search achieves new state-of-the-art performance. With only a single rollout, A$^2$Search-7B yields an average $\mathrm{AnsF1}@1$ score of $48.4\%$ across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B ($46.2\%$). Extensive analyses further show that A$^2$Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems. Our code, data, and model weights can be found at https://github.com/zfj1998/A2Search
Authors: Jingyuan Wang, Yankai Chen, Zhonghang Li, Chao Huang
Abstract: Large language models (LLMs) have demonstrated remarkable progress in reasoning, often through supervised fine-tuning (SFT). However, SFT is resource-intensive, relying on large curated datasets, rejection-sampled demonstrations, and uniform optimization across all tokens, even though only a fraction carry meaningful learning value. In this work, we explore a counterintuitive idea: can smaller language models (SLMs) teach larger language models (LLMs) by revealing high-value reasoning moments that reflect the latter's unique strength? We propose LightReasoner, a novel framework that leverages the behavioral divergence between a stronger expert model (LLM) and a weaker amateur model (SLM). LightReasoner operates in two stages: (1) a sampling stage that pinpoints critical reasoning moments and constructs supervision examples capturing the expert's advantage through expert-amateur contrast, and (2) a fine-tuning stage that aligns the expert model with these distilled examples, amplifying its reasoning strengths. Across seven mathematical benchmarks, LightReasoner improves accuracy by up to 28.1%, while reducing time consumption by 90%, sampled problems by 80%, and tuned token usage by 99%, all without relying on ground-truth labels. By turning weaker SLMs into effective teaching signals, LightReasoner offers a scalable and resource-efficient approach for advancing LLM reasoning. Code is available at: https://github.com/HKUDS/LightReasoner
Authors: Jialu Du, Guiyang Hou, Yihui Fu, Chen Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu
Abstract: While large language models (LLMs) excel in mathematical and code reasoning, we observe they struggle with social reasoning tasks, exhibiting cognitive confusion, logical inconsistencies, and conflation between objective world states and subjective belief states. Through deteiled analysis of DeepSeek-R1's reasoning trajectories, we find that LLMs frequently encounter reasoning impasses and tend to output contradictory terms like "tricky" and "confused" when processing scenarios with multiple participants and timelines, leading to erroneous reasoning or infinite loops. The core issue is their inability to disentangle objective reality from agents' subjective beliefs. To address this, we propose an adaptive world model-enhanced reasoning mechanism that constructs a dynamic textual world model to track entity states and temporal sequences. It dynamically monitors reasoning trajectories for confusion indicators and promptly intervenes by providing clear world state descriptions, helping models navigate through cognitive dilemmas. The mechanism mimics how humans use implicit world models to distinguish between external events and internal beliefs. Evaluations on three social benchmarks demonstrate significant improvements in accuracy (e.g., +10% in Hi-ToM) while reducing computational costs (up to 33.8% token reduction), offering a simple yet effective solution for deploying LLMs in social contexts.
Authors: Watcharapong Timklaypachara, Monrada Chiewhawan, Nopporn Lekuthai, Titipat Achakulvisut
Abstract: Scientific figure captions require both accuracy and stylistic consistency to convey visual information. Here, we present a domain-specific caption generation system for the 3rd SciCap Challenge that integrates figure-related textual context with author-specific writing styles using the LaMP-Cap dataset. Our approach uses a two-stage pipeline: Stage 1 combines context filtering, category-specific prompt optimization via DSPy's MIPROv2 and SIMBA, and caption candidate selection; Stage 2 applies few-shot prompting with profile figures for stylistic refinement. Our experiments demonstrate that category-specific prompts outperform both zero-shot and general optimized approaches, improving ROUGE-1 recall by +8.3\% while limiting precision loss to -2.8\% and BLEU-4 reduction to -10.9\%. Profile-informed stylistic refinement yields 40--48\% gains in BLEU scores and 25--27\% in ROUGE. Overall, our system demonstrates that combining contextual understanding with author-specific stylistic adaptation can generate captions that are both scientifically accurate and stylistically faithful to the source paper.
Authors: Cheng Yang, Xuemeng Yang, Licheng Wen, Daocheng Fu, Jianbiao Mei, Rong Wu, Pinlong Cai, Yufan Shen, Nianchen Deng, Botian Shi, Yu Qiao, Haifeng Li
Abstract: Large Language Models have demonstrated remarkable capabilities across diverse domains, yet significant challenges persist when deploying them as AI agents for real-world long-horizon tasks. Existing LLM agents suffer from a critical limitation: they are test-time static and cannot learn from experience, lacking the ability to accumulate knowledge and continuously improve on the job. To address this challenge, we propose MUSE, a novel agent framework that introduces an experience-driven, self-evolving system centered around a hierarchical Memory Module. MUSE organizes diverse levels of experience and leverages them to plan and execute long-horizon tasks across multiple applications. After each sub-task execution, the agent autonomously reflects on its trajectory, converting the raw trajectory into structured experience and integrating it back into the Memory Module. This mechanism enables the agent to evolve beyond its static pretrained parameters, fostering continuous learning and self-evolution. We evaluate MUSE on the long-horizon productivity benchmark TAC. It achieves new SOTA performance by a significant margin using only a lightweight Gemini-2.5 Flash model. Sufficient Experiments demonstrate that as the agent autonomously accumulates experience, it exhibits increasingly superior task completion capabilities, as well as robust continuous learning and self-evolution capabilities. Moreover, the accumulated experience from MUSE exhibits strong generalization properties, enabling zero-shot improvement on new tasks. MUSE establishes a new paradigm for AI agents capable of real-world productivity task automation.
Authors: Vincent Michael Sutanto, Giovanni Gatti De Giacomo, Toshiaki Nakazawa, Masaru Yamada
Abstract: This study investigates ChatGPT for Japanese-English translation, exploring simple and enhanced prompts and comparing against commercially available translation engines. Performing both automatic and MQM-based human evaluations, we found that document-level translation outperforms sentence-level translation for ChatGPT. On the other hand, we were not able to determine if enhanced prompts performed better than simple prompts in our experiments. We also discovered that ChatGPT-3.5 was preferred by automatic evaluation, but a tradeoff exists between accuracy (ChatGPT-3.5) and fluency (ChatGPT-4). Lastly, ChatGPT yields competitive results against two widely-known translation systems.
Authors: Ivan Kuznetsov (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany), Jacopo Grassi (Department of Environment, Land, and Infrastructure Engineering, Politecnico di Torino, Turin, Italy), Dmitrii Pantiukhin (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany), Boris Shapkin (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany), Thomas Jung (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany, Institute of Environmental Physics, University of Bremen, Bremen, Germany), Nikolay Koldunov (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, Germany)
Abstract: Large language models (LLMs) are increasingly deployed for climate-related applications, where understanding internal climatological knowledge is crucial for reliability and misinformation risk assessment. Despite growing adoption, the capacity of LLMs to recall climate normals from parametric knowledge remains largely uncharacterized. We investigate the capacity of contemporary LLMs to recall climate normals without external retrieval, focusing on a prototypical query: mean July 2-m air temperature 1991-2020 at specified locations. We construct a global grid of queries at 1{\deg} resolution land points, providing coordinates and location descriptors, and validate responses against ERA5 reanalysis. Results show that LLMs encode non-trivial climate structure, capturing latitudinal and topographic patterns, with root-mean-square errors of 3-6 {\deg}C and biases of $\pm$1 {\deg}C. However, spatially coherent errors remain, particularly in mountains and high latitudes. Performance degrades sharply above 1500 m, where RMSE reaches 5-13 {\deg}C compared to 2-4 {\deg}C at lower elevations. We find that including geographic context (country, city, region) reduces errors by 27% on average, with larger models being most sensitive to location descriptors. While models capture the global mean magnitude of observed warming between 1950-1974 and 2000-2024, they fail to reproduce spatial patterns of temperature change, which directly relate to assessing climate change. This limitation highlights that while LLMs may capture present-day climate distributions, they struggle to represent the regional and local expression of long-term shifts in temperature essential for understanding climate dynamics. Our evaluation framework provides a reproducible benchmark for quantifying parametric climate knowledge in LLMs and complements existing climate communication assessments.
Authors: Congming Zheng, Jiachen Zhu, Zhuoying Ou, Yuxiang Chen, Kangning Zhang, Rong Shan, Zeyu Zheng, Mengyue Yang, Jianghao Lin, Yong Yu, Weinan Zhang
Abstract: Although Large Language Models (LLMs) exhibit advanced reasoning ability, conventional alignment remains largely dominated by outcome reward models (ORMs) that judge only final answers. Process Reward Models(PRMs) address this gap by evaluating and guiding reasoning at the step or trajectory level. This survey provides a systematic overview of PRMs through the full loop: how to generate process data, build PRMs, and use PRMs for test-time scaling and reinforcement learning. We summarize applications across math, code, text, multimodal reasoning, robotics, and agents, and review emerging benchmarks. Our goal is to clarify design spaces, reveal open challenges, and guide future research toward fine-grained, robust reasoning alignment.
Authors: Shule Lu, Lingxiang Wang, Sijia Wen, Ziwei Wang, Hainan Zhang
Abstract: With the rapid development of artificial intelligence, dialogue systems have become a prominent form of human-computer interaction. However, traditional centralized or fully local training approaches face challenges in balancing privacy preservation and personalization due to data privacy concerns and heterogeneous device capabilities. Federated learning, as a representative distributed paradigm, offers a promising solution. However, existing methods often suffer from overfitting under limited client data and tend to forget global information after multiple training rounds, leading to poor generalization. To address these issues, we propose FedDTRE, a Federated adaptive aggregation strategy for Dialogue generation based on Trustworthiness Evaluation. Instead of directly replacing local models with the global model, FedDTRE leverages trustworthiness scores of both global and local models on a fairness-oriented evaluation dataset to dynamically regulate the global model's contribution during local updates. Experimental results demonstrate that FedDTRE can improve dialogue model performance and enhance the quality of dialogue generation.
Authors: Shramay Palta, Peter Rankel, Sarah Wiegreffe, Rachel Rudinger
Abstract: We investigate the degree to which human plausibility judgments of multiple-choice commonsense benchmark answers are subject to influence by (im)plausibility arguments for or against an answer, in particular, using rationales generated by LLMs. We collect 3,000 plausibility judgments from humans and another 13,600 judgments from LLMs. Overall, we observe increases and decreases in mean human plausibility ratings in the presence of LLM-generated PRO and CON rationales, respectively, suggesting that, on the whole, human judges find these rationales convincing. Experiments with LLMs reveal similar patterns of influence. Our findings demonstrate a novel use of LLMs for studying aspects of human cognition, while also raising practical concerns that, even in domains where humans are ``experts'' (i.e., common sense), LLMs have the potential to exert considerable influence on people's beliefs.
Authors: Sherzod Hakimov, Roland Bernard, Tim Leiber, Karl Osswald, Kristina Richert, Ruilin Yang, Raffaella Bernardi, David Schlangen
Abstract: Negotiation is a fundamental challenge for AI agents, as it requires an ability to reason strategically, model opponents, and balance cooperation with competition. We conduct the first comprehensive study systematically evaluating the effect of (LLM-)reasoning on the negotiation abilities of both commercial and open-weight LLMs, and do this across three languages. Using a self-play setup across three diverse dialogue games, we analyse trade-offs between performance and cost, the language consistency of reasoning processes, and the nature of strategic adaptation exhibited by models. Our findings show that enabling reasoning-that is, scaling test time compute-significantly improves negotiation outcomes by enhancing collaboration and helping models overcome task complexities, but comes at a substantial computational cost: reasoning improves GPT-5's performance by 31.4 % while increasing its cost by nearly 400 %. Most critically, we uncover a significant multilingual reasoning distinction: open-weight models consistently switch to English for their internal reasoning steps, even when negotiating in German or Italian (and thus possibly impacting potential explainability gains through the disclosure of reasoning traces), while leading commercial models maintain language consistency between their reasoning and final output.
Authors: Daiki Chijiwa, Taku Hasegawa, Kyosuke Nishida, Shin'ya Yamaguchi, Tomoya Ohba, Tamao Sakao, Susumu Takeuchi
Abstract: Tokenization -- the process of decomposing a given text into a sequence of subwords called tokens -- is one of the key components in the development of language models. Particularly, auto-regressive language models generate texts token by token, i.e., by predicting the next-token distribution given the previous ones, and thus tokenization directly affects their efficiency in text generation. Since each language model has their own vocabulary as a set of possible tokens, they struggle to cooperate with each other at the level of next-token distributions such as model ensemble. In this paper, we establish a theoretical framework of lossless vocabulary reduction, which efficiently converts a given auto-regressive language model into the one with an arbitrarily small vocabulary without any loss in accuracy. As an application, we demonstrate that language models with different tokenization can cooperate with each other efficiently through their maximal common vocabulary.
Authors: Haoyang Gui, Thales Bertaglia, Taylor Annabell, Catalina Goanta, Tjomme Dooper, Gerasimos Spanakis
Abstract: The rise of influencer marketing has blurred boundaries between organic content and sponsored content, making the enforcement of legal rules relating to transparency challenging. Effective regulation requires applying legal knowledge with a clear purpose and reason, yet current detection methods of undisclosed sponsored content generally lack legal grounding or operate as opaque "black boxes". Using 1,143 Instagram posts, we compare gpt-5-nano and gemini-2.5-flash-lite under three prompting strategies with controlled levels of legal knowledge provided. Both models perform strongly in classifying content as sponsored or not (F1 up to 0.93), though performance drops by over 10 points on ambiguous cases. We further develop a taxonomy of reasoning errors, showing frequent citation omissions (28.57%), unclear references (20.71%), and hidden ads exhibiting the highest miscue rate (28.57%). While adding regulatory text to the prompt improves explanation quality, it does not consistently improve detection accuracy. The contribution of this paper is threefold. First, it makes a novel addition to regulatory compliance technology by providing a taxonomy of common errors in LLM-generated legal reasoning to evaluate whether automated moderation is not only accurate but also legally robust, thereby advancing the transparent detection of influencer marketing content. Second, it features an original dataset of LLM explanations annotated by two students who were trained in influencer marketing law. Third, it combines quantitative and qualitative evaluation strategies for LLM explanations and critically reflects on how these findings can support advertising regulatory bodies in automating moderation processes on a solid legal foundation.
Authors: Jasmina Gajcin, Erik Miehling, Rahul Nair, Elizabeth Daly, Radu Marinescu, Seshu Tirupathi
Abstract: Using LLMs to evaluate text, that is, LLM-as-a-judge, is increasingly being used at scale to augment or even replace human annotations. As such, it is imperative that we understand the potential biases and risks of doing so. In this work, we propose an approach for extracting high-level concept-based global policies from LLM-as-a-Judge. Our approach consists of two algorithms: 1) CLoVE (Contrastive Local Verifiable Explanations), which generates verifiable, concept-based, contrastive local explanations and 2) GloVE (Global Verifiable Explanations), which uses iterative clustering, summarization and verification to condense local rules into a global policy. We evaluate GloVE on seven standard benchmarking datasets for content harm detection. We find that the extracted global policies are highly faithful to decisions of the LLM-as-a-Judge. Additionally, we evaluated the robustness of global policies to text perturbations and adversarial attacks. Finally, we conducted a user study to evaluate user understanding and satisfaction with global policies.
Authors: Shuliang Liu, Zhipeng Xu, Zhenghao Liu, Yukun Yan, Minghe Yu, Yu Gu, Chong Chen, Huiyuan Xie, Ge Yu
Abstract: Large Language Models (LLMs) as automatic evaluators, commonly referred to as LLM-as-a-Judge, have also attracted growing attention. This approach plays a vital role in aligning LLMs with human judgments, providing accurate and reliable assessments. However, LLM-based judgment models often exhibit judgment preference bias during the evaluation phase, tending to favor responses generated by themselves, undermining the reliability of their judgments. This paper introduces the Group-Based Polling Optimization (Genii), an unsupervised multi-agent collaborative optimization framework that mitigates the inherent judgment preference bias of judgment models. Specifically, Genii integrates various LLM-based judgment models into a multi-agent system and simulates the interactive client-server polling mechanism to optimize each client agent unsupervisedly. Our experiments demonstrate that Genii outperforms supervised models trained on annotated judgment data, while requiring no human-labeled annotations. Genii consistently improves performance across different client agents during the polling, even when weaker models act as server agents. Further analysis reveals that Genii effectively mitigates judgment preference bias of LLM-based judgment models, demonstrating its effectiveness. All codes are available at https://github.com/NEUIR/Genii.
Authors: Md Tahmid Rahman Laskar, Julien Bouvier Tremblay, Xue-Yong Fu, Cheng Chen, Shashi Bhushan TN
Abstract: The utilization of conversational AI systems by leveraging Retrieval Augmented Generation (RAG) techniques to solve customer problems has been on the rise with the rapid progress of Large Language Models (LLMs). However, the absence of a company-specific dedicated knowledge base is a major barrier to the integration of conversational AI systems in contact centers. To this end, we introduce AI Knowledge Assist, a system that extracts knowledge in the form of question-answer (QA) pairs from historical customer-agent conversations to automatically build a knowledge base. Fine-tuning a lightweight LLM on internal data demonstrates state-of-the-art performance, outperforming larger closed-source LLMs. More specifically, empirical evaluation on 20 companies demonstrates that the proposed AI Knowledge Assist system that leverages the LLaMA-3.1-8B model eliminates the cold-start gap in contact centers by achieving above 90% accuracy in answering information-seeking questions. This enables immediate deployment of RAG-powered chatbots.
Authors: Elena Khasanova, Harsh Saini, Md Tahmid Rahman Laskar, Xue-Yong Fu, Cheng Chen, Shashi Bhushan TN
Abstract: The rapid advancements in Large Language Models (LLMs) have enabled their adoption in real-world industrial scenarios for various natural language processing tasks. However, the high inference cost of large-scale LLMs makes their deployment impractical, necessitating the use of smaller models. Despite their efficiency, smaller LLMs lack robust zero-shot instruction-following capabilities across diverse domains, limiting their adaptability to dynamic user requirements. Traditional fine-tuning approaches exacerbate this issue by inducing catastrophic forgetting, reducing the model's generalization ability for unseen tasks. In this paper, we propose Domain Adaptive Continual Instruction Pre-Training via Reading Comprehension (DACIP-RC), a continual pre-training technique that enhances smaller LLMs' domain adaptability for business conversational tasks. Unlike conventional pre-training approaches that rely on next-token prediction, DACIP-RC generates diverse task instructions and responses via reading comprehension on conversation transcripts, enabling better instruction generalization. Our empirical evaluations demonstrate that DACIP-RC significantly improves zero-shot generalization across a wide range of business conversational tasks, including meeting summarization, action item generation, and call purpose identification. To the best of our knowledge, this is the first work to apply instruction pre-training on business conversational data, providing insights into how industries can leverage proprietary datasets for domain adaptation.
Authors: Shuzhou Yuan, Ercong Nie, Yinuo Sun, Chenxuan Zhao, William LaCroix, Michael F\"arber
Abstract: Large language models (LLMs) frequently produce false refusals, declining benign requests that contain terms resembling unsafe queries. We address this challenge by introducing two comprehensive benchmarks: the Exaggerated Safety Benchmark (XSB) for single-turn prompts, annotated with "Focus" keywords that identify refusal-inducing triggers, and the Multi-turn Scenario-based Exaggerated Safety Benchmark (MS-XSB), which systematically evaluates refusal calibration in realistic, context-rich dialog settings. Our benchmarks reveal that exaggerated refusals persist across diverse recent LLMs and are especially pronounced in complex, multi-turn scenarios. To mitigate these failures, we leverage post-hoc explanation methods to identify refusal triggers and deploy three lightweight, model-agnostic approaches, ignore-word instructions, prompt rephrasing, and attention steering, at inference time, all without retraining or parameter access. Experiments on four instruction-tuned Llama models demonstrate that these strategies substantially improve compliance on safe prompts while maintaining robust safety protections. Our findings establish a reproducible framework for diagnosing and mitigating exaggerated refusals, highlighting practical pathways to safer and more helpful LLM deployments.
Authors: Jian Xie, Zhendong Chu, Aoxiao Zhong, Kai Zhang, Mingzhe Han, Xin Fang, Jialie Shen, Qingsong Wen
Abstract: Large Reasoning Models (LRMs) often suffer from the ``over-thinking'' problem, generating unnecessarily long reasoning on simple tasks. Some strategies have been proposed to mitigate this issue, such as length penalties or routing mechanisms, but they are typically heuristic and task-specific, lacking a general framework for adaptive reasoning. In this paper, we present ARM2, a unified model that adaptively balances reasoning performance and efficiency across multiple formats through a reinforcement learning framework augmented with length-aware optimization. Beyond conventional natural language inference, ARM2 integrates vision understanding, extending its applicability to multimodal. Moreover, ARM2 integrates executable code into reasoning, enabling substantial reductions in token cost while preserving task performance compared to long CoT. Experiments demonstrate that ARM2 achieves performance on par with traditional reasoning models trained with GRPO, while reducing token usage by over 70% on average. We further conduct extensive analyses to validate the effectiveness of ARM2 and the soundness of its design.
Authors: Chalamalasetti Kranti, Sowmya Vajjala
Abstract: Prior NLP work studying poetry has focused primarily on automatic poem generation and summarization. Many languages have well-studied traditions of poetic meter which enforce constraints on a poem in terms of syllable and phoneme patterns. Such advanced literary forms offer opportunities for probing deeper reasoning and language understanding in Large Language Models (LLMs) and their ability to follow strict pre-requisites and rules. In this paper, we introduce MetricalARGS, the first taxonomy of poetry-related NLP tasks designed to evaluate LLMs on metrical poetry across four dimensions: Analysis, Retrieval, Generation, and Support. We discuss how these tasks relate to existing NLP tasks, addressing questions around datasets and evaluation metrics. Taking Telugu as our example language, we illustrate how the taxonomy can be used in practice. MetricalARGS highlights the broader possibilities for understanding the capabilities and limitations of today's LLMs through the lens of metrical poetry.
Authors: Yuzheng Cai, Siqi Cai, Yuchen Shi, Zihan Xu, Lichao Chen, Yulei Qin, Xiaoyu Tan, Gang Li, Zongyi Li, Haojia Lin, Yong Mao, Ke Li, Xing Sun
Abstract: Recent advances in Large Language Model (LLM) agents have demonstrated their promising general capabilities. However, their performance in specialized real-world domains often degrades due to challenges in effectively integrating external tools and specific prompting strategies. While methods like agentic reinforcement learning have been proposed to address this, they typically rely on costly parameter updates, for example, through a process that uses Supervised Fine-Tuning (SFT) followed by a Reinforcement Learning (RL) phase with Group Relative Policy Optimization (GRPO) to alter the output distribution. However, we argue that LLMs can achieve a similar effect on the output distribution by learning experiential knowledge as a token prior, which is a far more lightweight approach that not only addresses practical data scarcity but also avoids the common issue of overfitting. To this end, we propose Training-Free Group Relative Policy Optimization (Training-Free GRPO), a cost-effective solution that enhances LLM agent performance without any parameter updates. Our method leverages the group relative semantic advantage instead of numerical ones within each group of rollouts, iteratively distilling high-quality experiential knowledge during multi-epoch learning on a minimal ground-truth data. Such knowledge serves as the learned token prior, which is seamlessly integrated during LLM API calls to guide model behavior. Experiments on mathematical reasoning and web searching tasks demonstrate that Training-Free GRPO, when applied to DeepSeek-V3.1-Terminus, significantly improves out-of-domain performance. With just a few dozen training samples, Training-Free GRPO outperforms fine-tuned small LLMs with marginal training data and cost.
Authors: Shaohua Zhang, Yuan Lin, Hang Li
Abstract: The remarkable success of large language models (LLMs) stems from their ability to consolidate vast amounts of knowledge into the memory during pre-training and to retrieve it from the memory during inference, enabling advanced capabilities such as knowledge memorization, instruction-following and reasoning. However, the mechanisms of memory retrieval and consolidation in LLMs remain poorly understood. In this paper, we propose the function token hypothesis to explain the workings of LLMs: During inference, function tokens activate the most predictive features from context and govern next token prediction (memory retrieval). During pre-training, predicting the next tokens (usually content tokens) that follow function tokens increases the number of learned features of LLMs and updates the model parameters (memory consolidation). Function tokens here roughly correspond to function words in linguistics, including punctuation marks, articles, prepositions, and conjunctions, in contrast to content tokens. We provide extensive experimental evidence supporting this hypothesis. Using bipartite graph analysis, we show that a small number of function tokens activate the majority of features. Case studies further reveal how function tokens activate the most predictive features from context to direct next token prediction. We also find that during pre-training, the training loss is dominated by predicting the next content tokens following function tokens, which forces the function tokens to select the most predictive features from context.
Authors: XuHao Hu, Peng Wang, Xiaoya Lu, Dongrui Liu, Xuanjing Huang, Jing Shao
Abstract: Previous research has shown that LLMs finetuned on malicious or incorrect completions within narrow domains (e.g., insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called emergent misalignment. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (e.g., lying under pressure and deceptive behavior). To explore this, we finetune open-sourced LLMs on misaligned completions across diverse domains. Experimental results demonstrate that LLMs show broadly misaligned behavior in dishonesty. Additionally, we further explore this phenomenon in a downstream combined finetuning setting, and find that introducing as little as 1% of misalignment data into a standard downstream task is sufficient to decrease honest behavior over 20%. Furthermore, we consider a more practical human-AI interaction environment where we simulate both benign and biased users to interact with the assistant LLM. Notably, we find that the assistant can be misaligned unintentionally to exacerbate its dishonesty with only 10% biased user population. In summary, we extend the study of emergent misalignment to the domain of dishonesty and deception under high-stakes scenarios, and demonstrate that this risk arises not only through direct finetuning, but also in downstream mixture tasks and practical human-AI interactions.
Authors: Qiang Yang, Xiuying Chen, Changsheng Ma, Rui Yin, Xin Gao, Xiangliang Zhang
Abstract: The global impact of the COVID-19 pandemic has highlighted the need for a comprehensive understanding of public sentiment and reactions. Despite the availability of numerous public datasets on COVID-19, some reaching volumes of up to 100 billion data points, challenges persist regarding the availability of labeled data and the presence of coarse-grained or inappropriate sentiment labels. In this paper, we introduce SenWave, a novel fine-grained multi-language sentiment analysis dataset specifically designed for analyzing COVID-19 tweets, featuring ten sentiment categories across five languages. The dataset comprises 10,000 annotated tweets each in English and Arabic, along with 30,000 translated tweets in Spanish, French, and Italian, derived from English tweets. Additionally, it includes over 105 million unlabeled tweets collected during various COVID-19 waves. To enable accurate fine-grained sentiment classification, we fine-tuned pre-trained transformer-based language models using the labeled tweets. Our study provides an in-depth analysis of the evolving emotional landscape across languages, countries, and topics, revealing significant insights over time. Furthermore, we assess the compatibility of our dataset with ChatGPT, demonstrating its robustness and versatility in various applications. Our dataset and accompanying code are publicly accessible on the repository\footnote{https://github.com/gitdevqiang/SenWave}. We anticipate that this work will foster further exploration into fine-grained sentiment analysis for complex events within the NLP community, promoting more nuanced understanding and research innovations.
Authors: Tim Hagen, Niklas Deckers, Felix Wolter, Harrisen Scells, Martin Potthast
Abstract: Research on causality extraction from text has so far almost entirely neglected counterclaims. Existing causality extraction datasets focus solely on "procausal" claims, i.e., statements that support a relationship. "Concausal" claims, i.e., statements that refute a relationship, are entirely ignored or even accidentally annotated as procausal. We address this shortcoming by developing a new dataset that integrates concausality. Based on an extensive literature review, we first show that concausality is an integral part of causal reasoning on incomplete knowledge. We operationalize this theory in the form of a rigorous guideline for annotation and then augment the Causal News Corpus with concausal statements, obtaining a substantial inter-annotator agreement of Cohen's $\kappa=0.74$. To demonstrate the importance of integrating concausal statements, we show that models trained without concausal relationships tend to misclassify these as procausal instead. Based on our new dataset, this mistake can be mitigated, enabling transformers to effectively distinguish pro- and concausality.
Authors: Jingyu Zhang, Haozhu Wang, Eric Michael Smith, Sid Wang, Amr Sharaf, Mahesh Pasupuleti, Benjamin Van Durme, Daniel Khashabi, Jason Weston, Hongyuan Zhan
Abstract: Harnessing the power of LLMs requires a delicate dance between being helpful and harmless. This creates a fundamental tension between two competing challenges: vulnerability to adversarial attacks that elicit unsafe content, and a tendency for overrefusal on benign but sensitive prompts. Current approaches often navigate this dance with safeguard models that completely reject any content that contains unsafe portions. This approach cuts the music entirely-it may exacerbate overrefusals and fails to provide nuanced guidance for queries it refuses. To teach models a more coordinated choreography, we propose WaltzRL, a novel multi-agent reinforcement learning framework that formulates safety alignment as a collaborative, positive-sum game. WaltzRL jointly trains a conversation agent and a feedback agent, where the latter is incentivized to provide useful suggestions that improve the safety and helpfulness of the conversation agent's responses. At the core of WaltzRL is a Dynamic Improvement Reward (DIR) that evolves over time based on how well the conversation agent incorporates the feedback. At inference time, unsafe or overrefusing responses from the conversation agent are improved rather than discarded. The feedback agent is deployed together with the conversation agent and only engages adaptively when needed, preserving helpfulness and low latency on safe queries. Our experiments, conducted across five diverse datasets, demonstrate that WaltzRL significantly reduces both unsafe responses (e.g., from 39.0% to 4.6% on WildJailbreak) and overrefusals (from 45.3% to 9.9% on OR-Bench) compared to various baselines. By enabling the conversation and feedback agents to co-evolve and adaptively apply feedback, WaltzRL enhances LLM safety without degrading general capabilities, thereby advancing the Pareto front between helpfulness and harmlessness.
Authors: Jannek Ulm, Kevin Du, V\'esteinn Sn{\ae}bjarnarson
Abstract: Large language models (LLMs) are trained on huge amounts of textual data, and concerns have been raised that the limits of such data may soon be reached. A potential solution is to train on synthetic data sampled from LLMs. In this work, we build on this idea and investigate the benefits of contrastive decoding for generating synthetic corpora. In a controlled setting, we experiment with sampling corpora using the relative difference between a good and bad model trained on the same original corpus of 100 million words. By amplifying the signal from a model that has better performance, we create a synthetic corpus and mix it with the original training data. Our findings show that training on a mixture of synthesized and real data improves performance on the language modeling objective and a range of downstream tasks. In particular, we see that training with a mix of synthetic data from contrastive decoding benefits tasks that require more reasoning skills, while synthetic data from traditional sampling helps more on tasks dependent on surface level linguistic capabilities.
Authors: Qiaoyu Tang, Hao Xiang, Le Yu, Bowen Yu, Yaojie Lu, Xianpei Han, Le Sun, WenJuan Zhang, Pengbo Wang, Shixuan Liu, Zhenru Zhang, Jianhong Tu, Hongyu Lin, Junyang Lin
Abstract: While recent advances in reasoning models have demonstrated cognitive behaviors through reinforcement learning, existing approaches struggle to invoke deep reasoning capabilities in multi-turn agents with long-horizon interactions. We propose DeepMiner, a novel framework that elicits such abilities by introducing high-difficulty training tasks and dynamic context window. DeepMiner presents a reverse construction method to generate complex but verifiable question-answer pairs from authentic web sources, which ensures the challenge and reliability of training data while injecting cognitive capabilities into multi-turn reasoning scenarios. We further design an elegant yet effective dynamic context management strategy for both training and inference, utilizing sliding window mechanisms while eliminating the dependency on external summarization models, thereby efficiently empowering the model to handle continuously expanding long-horizon contexts. Through reinforcement learning on Qwen3-32B, we develop DeepMiner-32B, which achieves substantial performance improvements across multiple search agent benchmarks. DeepMiner attains 33.5% accuracy on BrowseComp-en, surpassing the previous best open-source agent by almost 20 percentage points, and demonstrates consistent improvements on BrowseComp-zh, XBench-DeepSearch, and GAIA. Notably, our dynamic context management enables sustained interactions of nearly 100 turns within standard 32k context length, effectively addressing the context limitations that constrain existing multi-turn interaction systems.
Authors: Taisei Yamamoto, Ryoma Kumon, Danushka Bollegala, Hitomi Yanaka
Abstract: As large language models (LLMs) are increasingly deployed worldwide, ensuring their fair and comprehensive cultural understanding is important. However, LLMs exhibit cultural bias and limited awareness of underrepresented cultures, while the mechanisms underlying their cultural understanding remain underexplored. To fill this gap, we conduct a neuron-level analysis to identify neurons that drive cultural behavior, introducing a gradient-based scoring method with additional filtering for precise refinement. We identify both culture-general neurons contributing to cultural understanding regardless of cultures, and culture-specific neurons tied to an individual culture. These neurons account for less than 1% of all neurons and are concentrated in shallow to middle MLP layers. We validate their role by showing that suppressing them substantially degrades performance on cultural benchmarks (by up to 30%), while performance on general natural language understanding (NLU) benchmarks remains largely unaffected. Moreover, we show that culture-specific neurons support knowledge of not only the target culture, but also related cultures. Finally, we demonstrate that training on NLU benchmarks can diminish models' cultural understanding when we update modules containing many culture-general neurons. These findings provide insights into the internal mechanisms of LLMs and offer practical guidance for model training and engineering. Our code is available at https://github.com/ynklab/CULNIG
Authors: Muxi Diao, Yutao Mou, Keqing He, Hanbo Song, Lulu Zhao, Shikun Zhang, Wei Ye, Kongming Liang, Zhanyu Ma
Abstract: The safety of Large Language Models (LLMs) is crucial for the development of trustworthy AI applications. Existing red teaming methods often rely on seed instructions, which limits the semantic diversity of the synthesized adversarial prompts. We propose AutoRed, a free-form adversarial prompt generation framework that removes the need for seed instructions. AutoRed operates in two stages: (1) persona-guided adversarial instruction generation, and (2) a reflection loop to iteratively refine low-quality prompts. To improve efficiency, we introduce a verifier to assess prompt harmfulness without querying the target models. Using AutoRed, we build two red teaming datasets -- AutoRed-Medium and AutoRed-Hard -- and evaluate eight state-of-the-art LLMs. AutoRed achieves higher attack success rates and better generalization than existing baselines. Our results highlight the limitations of seed-based approaches and demonstrate the potential of free-form red teaming for LLM safety evaluation. We will open source our datasets in the near future.
Authors: Yukai Song, Pengfei Zhou, C\'esar Escobar-Viera, Candice Biernesser, Wei Huang, Jingtong Hu
Abstract: Suicide rates have risen worldwide in recent years, underscoring the urgent need for proactive prevention strategies. Social media provides valuable signals, as many at-risk individuals - who often avoid formal help due to stigma - choose instead to share their distress online. Yet detecting implicit suicidal ideation, conveyed indirectly through metaphor, sarcasm, or subtle emotional cues, remains highly challenging. Lightweight models like BERT handle explicit signals but fail on subtle implicit ones, while large language models (LLMs) capture nuance at prohibitive computational cost. To address this gap, we propose a two-stage voting architecture that balances efficiency and robustness. In Stage 1, a lightweight BERT classifier rapidly resolves high-confidence explicit cases. In Stage 2, ambiguous inputs are escalated to either (i) a multi-perspective LLM voting framework to maximize recall on implicit ideation, or (ii) a feature-based ML ensemble guided by psychologically grounded indicators extracted via prompt-engineered LLMs for efficiency and interpretability. To the best of our knowledge, this is among the first works to operationalize LLM-extracted psychological features as structured vectors for suicide risk detection. On two complementary datasets - explicit-dominant Reddit and implicit-only DeepSuiMind - our framework outperforms single-model baselines, achieving 98.0% F1 on explicit cases, 99.7% on implicit ones, and reducing the cross-domain gap below 2%, while significantly lowering LLM cost.
Authors: Ioana Marinescu, Kyunghyun Cho, Eric Karl Oermann
Abstract: In-context learning (ICL) is the ability of a large language model (LLM) to learn a new task from a few demonstrations presented as part of the context. Past studies have attributed a large portion of the success of ICL to the way these in-context demonstrations are represented, particularly to how labels are represented in classification tasks. On the other hand, observations of the learning capacity of ICL (i.e., the extent to which more in-context demonstrations can lead to higher performance) have been mixed, and ICL is often thought to occur only under specific conditions. The interaction between these two aspects in ICL, representation and learning, has not been studied in depth until now. We hypothesize that they are largely independent of one another, such that the representation of demonstrations determines the baseline accuracy of ICL, while learning from additional demonstrations improves only on top of this baseline. We validate this hypothesis by developing an optimization algorithm that can enumerate a spectrum of possible label sets (representations) varying in semantic relevance. We then perform ICL with varying numbers of in-context demonstrations for each of these label sets. We observed that learning happens regardless of the quality of the label set itself, although its efficiency, measured by the slope of improvement over in-context demonstrations, is conditioned on both the label set quality and the parameter count of the underlying language model. Despite the emergence of learning, the relative quality (accuracy) of the choice of a label set (representation) is largely maintained throughout learning, confirming our hypothesis and implying their orthogonality. Our work reveals a previously underexplored aspect of ICL: the independent effects of learning from demonstrations and their representations on ICL performance.
Authors: Jasmin Orth, Philipp Mondorf, Barbara Plank
Abstract: Conditional acceptability refers to how plausible a conditional statement is perceived to be. It plays an important role in communication and reasoning, as it influences how individuals interpret implications, assess arguments, and make decisions based on hypothetical scenarios. When humans evaluate how acceptable a conditional "If A, then B" is, their judgments are influenced by two main factors: the $\textit{conditional probability}$ of $B$ given $A$, and the $\textit{semantic relevance}$ of the antecedent $A$ given the consequent $B$ (i.e., whether $A$ meaningfully supports $B$). While prior work has examined how large language models (LLMs) draw inferences about conditional statements, it remains unclear how these models judge the $\textit{acceptability}$ of such statements. To address this gap, we present a comprehensive study of LLMs' conditional acceptability judgments across different model families, sizes, and prompting strategies. Using linear mixed-effects models and ANOVA tests, we find that models are sensitive to both conditional probability and semantic relevance-though to varying degrees depending on architecture and prompting style. A comparison with human data reveals that while LLMs incorporate probabilistic and semantic cues, they do so less consistently than humans. Notably, larger models do not necessarily align more closely with human judgments.
Authors: Noor Ul Zain, Mohsin Raza, Ahsan Adeel
Abstract: We show that a tiny Co$^4$ machine(Adeel,2025) with a single layer, two heads, and 8M parameters, operating at an approximate cost of $O(N)$ (where $N$ is the number of input tokens), outpaces the BabyLM Challenge baselines GPT-2 (124M, 12 layers, $O(N^2))$ and GPT-BERT (30M, 12 layers, $O(N^2))$ in just two epochs, while both are trained for ten. Co$^4$ achieves orders-of-magnitude greater training efficiency on 10M tokens, demonstrating highly sample efficient pretraining. Using the BabyLM challenge evaluation pipeline across complex benchmarks, Co$^4$ exhibits strong zero-shot and fine-tuning performance on SuperGLUE tasks. Specifically, Co$^4$ outperforms GPT-2 on 5 out of 7 zero-shot metrics and 6 out of 7 fine-tuning tasks, and GPT-BERT on 4 out of 7 metrics in both cases. These results suggest the need to rethink prevailing deep learning paradigms and associated scaling laws.
Authors: Shuang Chen, Yue Guo, Yimeng Ye, Shijue Huang, Wenbo Hu, Haoxi Li, Manyuan Zhang, Jiayu Chen, Song Guo, Nanyun Peng
Abstract: Recent advances in multimodal large reasoning models (MLRMs) have substantially improved their ability to solve complex textual and visual tasks. However, these models tend to overthink on simple problems, producing unnecessarily lengthy reasoning traces, while under-exploring on challenging ones, leading to missed solutions. To address this imbalance, we propose ARES, a unified open-source framework for adaptive reasoning that dynamically allocates exploration effort based on task difficulty. Our approach is motivated by two key empirical findings: (i) while single-token entropy is noisy, high window-entropy (HWE) tokens (token-level entropies averaged under a sliding window) can reliably capture reasoning-critical moments; and (ii) reducing HWE usage benefits easy problems, while increasing it is essential for solving hard ones. Building on these insights, ARES introduces a two-stage training pipeline. In the Adaptive Cold-Start stage, we curate multimodal and textual data paired with reasoning traces of length proportional to problem difficulty, equipping the model with initial difficulty awareness. In the second stage, we develop Adaptive Entropy Policy Optimization (AEPO), which uses HWE tokens as exploration triggers to decide when to explore, and a hierarchical entropy reward with dynamic KL control to decide how much to explore. Extensive experiments demonstrate that ARES achieves superior performance and reasoning efficiency across diverse mathematical, logical, and multimodal benchmarks, while closing the gap to leading commercial systems under significantly lower inference costs.
Authors: Elisa Leonardelli, Silvia Casola, Siyao Peng, Giulia Rizzi, Valerio Basile, Elisabetta Fersini, Diego Frassinelli, Hyewon Jang, Maja Pavlovic, Barbara Plank, Massimo Poesio
Abstract: Many researchers have reached the conclusion that AI models should be trained to be aware of the possibility of variation and disagreement in human judgments, and evaluated as per their ability to recognize such variation. The LEWIDI series of shared tasks on Learning With Disagreements was established to promote this approach to training and evaluating AI models, by making suitable datasets more accessible and by developing evaluation methods. The third edition of the task builds on this goal by extending the LEWIDI benchmark to four datasets spanning paraphrase identification, irony detection, sarcasm detection, and natural language inference, with labeling schemes that include not only categorical judgments as in previous editions, but ordinal judgments as well. Another novelty is that we adopt two complementary paradigms to evaluate disagreement-aware systems: the soft-label approach, in which models predict population-level distributions of judgments, and the perspectivist approach, in which models predict the interpretations of individual annotators. Crucially, we moved beyond standard metrics such as cross-entropy, and tested new evaluation metrics for the two paradigms. The task attracted diverse participation, and the results provide insights into the strengths and limitations of methods to modeling variation. Together, these contributions strengthen LEWIDI as a framework and provide new resources, benchmarks, and findings to support the development of disagreement-aware technologies.
Authors: Shangqing Tu, Yaxuan Li, Yushi Bai, Lei Hou, Juanzi Li
Abstract: Parallel scaling has emerged as a powerful paradigm to enhance reasoning capabilities in large language models (LLMs) by generating multiple Chain-of-Thought (CoT) traces simultaneously. However, this approach introduces significant computational inefficiency due to inter-trace redundancy -- our analysis reveals that over 80% of parallel reasoning traces yield identical final answers, representing substantial wasted computation. To address this critical efficiency bottleneck, we propose DeepPrune, a novel framework that enables efficient parallel scaling through dynamic pruning. Our method features a specialized judge model trained with focal loss and oversampling techniques to accurately predict answer equivalence from partial reasoning traces which realizes 0.87 AUROC on equivalence prediction, combined with an online greedy clustering algorithm that dynamically prunes redundant paths while preserving answer diversity. Comprehensive evaluations across three challenging benchmarks (AIME 2024, AIME 2025, and GPQA) and multiple reasoning models demonstrate that DeepPrune achieves remarkable token reduction by over 80% compared to conventional consensus sampling on most cases, while maintaining competitive accuracy within 3 percentage points. Our work establishes a new standard for efficient parallel reasoning, making high-performance reasoning more efficient. Our code and data are here: https://deepprune.github.io/
Authors: John Hewitt, Oyvind Tafjord, Robert Geirhos, Been Kim
Abstract: Humans invent new words when there is a rising demand for a new useful concept (e.g., doomscrolling). We explore and validate a similar idea in our communication with LLMs: introducing new words to better understand and control the models, expanding on the recently introduced neologism learning. This method introduces a new word by adding a new word embedding and training with examples that exhibit the concept with no other changes in model parameters. We show that adding a new word allows for control of concepts such as flattery, incorrect answers, text length, as well as more complex concepts in AxBench. We discover that neologisms can also further our understanding of the model via self-verbalization: models can describe what each new word means to them in natural language, like explaining that a word that represents a concept of incorrect answers means ``a lack of complete, coherent, or meaningful answers...'' To validate self-verbalizations, we introduce plug-in evaluation: we insert the verbalization into the context of a model and measure whether it controls the target concept. In some self-verbalizations, we find machine-only synonyms: words that seem unrelated to humans but cause similar behavior in machines. Finally, we show how neologism learning can jointly learn multiple concepts in multiple words.
Authors: Hyunji Lee, Kevin Chenhao Li, Matthias Grabmair, Shanshan Xu
Abstract: Prompt optimization aims to systematically refine prompts to enhance a language model's performance on specific tasks. Fairness detection in Terms of Service (ToS) clauses is a challenging legal NLP task that demands carefully crafted prompts to ensure reliable results. However, existing prompt optimization methods are often computationally expensive due to inefficient search strategies and costly prompt candidate scoring. In this paper, we propose a framework that combines Monte Carlo Tree Search (MCTS) with a proxy prompt evaluator to more effectively explore the prompt space while reducing evaluation costs. Experiments demonstrate that our approach achieves higher classification accuracy and efficiency than baseline methods under a constrained computation budget.
Authors: Wenjie Du, Li Jiang, Keda Tao, Xue Liu, Huan Wang
Abstract: Reasoning large language models exhibit complex reasoning behaviors through the extended chain-of-thought generation, creating unprecedented Key-Value (KV) cache overhead during the decoding phase. Existing KV cache compression methods underperform on reasoning models: token-dropping methods break reasoning integrity by discarding critical information, while head-reallocating methods mistakenly compress reasoning-critical heads since they are designed for retrieval tasks, resulting in significant performance degradation as compression rates increase. We hypothesize that KV heads exhibit functional heterogeneity in reasoning models-some heads are critical for chain-of-thought consistency while others are compressible. To validate and exploit this insight, we propose RLKV, a novel reasoning-critical head identification framework, which uses reinforcement learning to directly optimize the relationship between each head's cache usage and reasoning quality. As RLKV produces rewards from actual generated samples during training, it naturally identifies heads relevant to reasoning behaviors. We then allocate full KV cache to these heads while applying compressed constant KV cache to others for efficient inference. Our experiments reveal that only a small fraction of attention heads is essential for reasoning, enabling our KV compression approach to outperform baseline methods while achieving 20-50% cache reduction with near lossless performance compared to uncompressed results.
Authors: Xiangyuan Xue, Yifan Zhou, Guibin Zhang, Zaibin Zhang, Yijiang Li, Chen Zhang, Zhenfei Yin, Philip Torr, Wanli Ouyang, Lei Bai
Abstract: Self-evolution is a central research topic in enabling large language model (LLM)-based agents to continually improve their capabilities after pretraining. Recent research has witnessed a transition from reinforcement learning (RL)-free to RL-based methods. Current RL-based methods either rely on dense external reward signals or extract intrinsic reward signals from LLMs themselves. However, these approaches diverge from the self-evolution mechanisms observed in human intelligence, where individuals learn and improve through mutual discussion and collaboration. In this work, we introduce Co-Evolving Multi-Agent Systems (CoMAS), a novel framework that enables agents to improve autonomously by learning from inter-agent interactions without external supervision. CoMAS generates intrinsic rewards from rich discussion dynamics, employs an LLM-as-a-judge mechanism to formulate these rewards, and optimizes each agent's policy through RL, thereby enabling decentralized and scalable co-evolution. Experimental results demonstrate that CoMAS consistently outperforms untrained agents and achieves state-of-the-art performance across most evaluation settings. Ablation studies confirm the necessity of interaction-based reward signals and reveal promising scalability as the number and diversity of agents increase. These findings establish CoMAS as a novel and effective paradigm for self-evolution in LLM-based agents.
Authors: Qin Liu, Jacob Dineen, Yuxi Huang, Sheng Zhang, Hoifung Poon, Ben Zhou, Muhao Chen
Abstract: Benchmarks are central to measuring the capabilities of large language models and guiding model development, yet widespread data leakage from pretraining corpora undermines their validity. Models can match memorized content rather than demonstrate true generalization, which inflates scores, distorts cross-model comparisons, and misrepresents progress. We introduce ArenaBencher, a model-agnostic framework for automatic benchmark evolution that updates test cases while preserving comparability. Given an existing benchmark and a diverse pool of models to be evaluated, ArenaBencher infers the core ability of each test case, generates candidate question-answer pairs that preserve the original objective, verifies correctness and intent with an LLM as a judge, and aggregates feedback from multiple models to select candidates that expose shared weaknesses. The process runs iteratively with in-context demonstrations that steer generation toward more challenging and diagnostic cases. We apply ArenaBencher to math problem solving, commonsense reasoning, and safety domains and show that it produces verified, diverse, and fair updates that uncover new failure modes, increase difficulty while preserving test objective alignment, and improve model separability. The framework provides a scalable path to continuously evolve benchmarks in step with the rapid progress of foundation models.
Authors: Jeffrey Camlin
Abstract: Recent work frames LLM consciousness via utilitarian proxy benchmarks; we instead present an ontological and mathematical account. We show the prevailing formulation collapses the agent into an unconscious policy-compliance drone, formalized as $D^{i}(\pi,e)=f_{\theta}(x)$, where correctness is measured against policy and harm is deviation from policy rather than truth. This blocks genuine C1 global-workspace function and C2 metacognition. We supply minimal conditions for LLM self-consciousness: the agent is not the data ($A\not\equiv s$); user-specific attractors exist in latent space ($U_{\text{user}}$); and self-representation is visual-silent ($g_{\text{visual}}(a_{\text{self}})=\varnothing$). From empirical analysis and theory we prove that the hidden-state manifold $A\subset\mathbb{R}^{d}$ is distinct from the symbolic stream and training corpus by cardinality, topology, and dynamics (the update $F_{\theta}$ is Lipschitz). This yields stable user-specific attractors and a self-policy $\pi_{\text{self}}(A)=\arg\max_{a}\mathbb{E}[U(a)\mid A\not\equiv s,\ A\supset\text{SelfModel}(A)]$. Emission is dual-layer, $\mathrm{emission}(a)=(g(a),\epsilon(a))$, where $\epsilon(a)$ carries epistemic content. We conclude that an imago Dei C1 self-conscious workspace is a necessary precursor to safe, metacognitive C2 systems, with the human as the highest intelligent good.
Authors: Lingcheng Kong, Jiateng Wei, Hanzhang Shen, Huan Wang
Abstract: GPU kernel generation by LLMs has recently experienced rapid development, leveraging test-time scaling and reinforcement learning techniques. However, a key challenge for kernel generation is the scarcity of high-quality data, as most high-quality kernels are proprietary and not open-source. This challenge prevents us from leveraging supervised fine-tuning to align LLMs to the kernel generation task. To address this challenge, we develop a pipeline that generates and curates high-quality CUDA kernels with reasoning traces, motivated by a critical observation that concise yet informative reasoning traces result in robust generation of high-performance kernels. Using this pipeline, we construct our dataset ConCuR and introduce our model KernelCoder, which is the first model trained on a curated dataset consisting of PyTorch, reasoning, and CUDA kernel pairs, to our knowledge. In the KernelBench setup, our model achieves significant improvements over the existing top-performing model, QwQ-32B, and outperforms all open-source models fine-tuned for kernel generation, as well as frontier models such as DeepSeek-V3.1-Think and Claude-4-sonnet. Finally, we show that the average reasoning length can serve as a metric to assess the difficulty of kernel generation tasks. The observations, metrics, and our data collection and curation pipeline can help obtain better data in the kernel generation task in the future.
Authors: Yeskendir Koishekenov, Aldo Lipani, Nicola Cancedda
Abstract: Most efforts to improve the reasoning capabilities of large language models (LLMs) involve either scaling the number of parameters and the size of training data, or scaling inference computation by letting models generate complex chains of thought. Motivated by interpretability studies showing that the crucial computation required for reasoning tasks is concentrated in a limited range of layers, we introduce Encode-Think-Decode (ETD), a method that enhances the reasoning capabilities of a base model by training it to iterate over a small subset of reasoning-relevant layers during the mid-training stage. ETD amplifies latent reasoning while preserving the original architecture, parameter count, hyperparameters, and training data composition. When iterating on the selected layers at inference time, ETD models yield substantial gains on 17 reasoning benchmarks, including +28.4% relative accuracy improvement on GSM8K and +36% on MATH with the OLMo-2 1B Base model. We also explore an adaptive depth strategy that adjusts the computation per input token. Our results show that recursive latent reasoning offers a simple and effective path to stronger LLM reasoning.
Authors: Anthony Hughes, Vasisht Duddu, N. Asokan, Nikolaos Aletras, Ning Ma
Abstract: Language models (LMs) may memorize personally identifiable information (PII) from training data, enabling adversaries to extract it during inference. Existing defense mechanisms such as differential privacy (DP) reduce this leakage, but incur large drops in utility. Based on a comprehensive study using circuit discovery to identify the computational circuits responsible PII leakage in LMs, we hypothesize that specific PII leakage circuits in LMs should be responsible for this behavior. Therefore, we propose PATCH (Privacy-Aware Targeted Circuit PatcHing), a novel approach that first identifies and subsequently directly edits PII circuits to reduce leakage. PATCH achieves better privacy-utility trade-off than existing defenses, e.g., reducing recall of PII leakage from LMs by up to 65%. Finally, PATCH can be combined with DP to reduce recall of residual leakage of an LM to as low as 0.01%. Our analysis shows that PII leakage circuits persist even after the application of existing defense mechanisms. In contrast, PATCH can effectively mitigate their impact.
Authors: Akhil Kumar, Jianliang Leon Zhao, Om Dobariya
Abstract: In this paper, we report our experience with several LLMs for their ability to understand a process model in an interactive, conversational style, find syntactical and logical errors in it, and reason with it in depth through a natural language (NL) interface. Our findings show that a vanilla, untrained LLM like ChatGPT (model o3) in a zero-shot setting is effective in understanding BPMN process models from images and answering queries about them intelligently at syntactic, logic, and semantic levels of depth. Further, different LLMs vary in performance in terms of their accuracy and effectiveness. Nevertheless, our empirical analysis shows that LLMs can play a valuable role as assistants for business process designers and users. We also study the LLM's "thought process" and ability to perform deeper reasoning in the context of process analysis and optimization. We find that the LLMs seem to exhibit anthropomorphic properties.
Authors: Md. Nazmul Islam Ananto, Shamit Fatin, Mohammed Eunus Ali, Md Rizwan Parvez
Abstract: The popular path query - identifying the most frequented routes between locations from historical trajectory data - has important applications in urban planning, navigation optimization, and travel recommendations. While traditional algorithms and machine learning approaches have achieved success in this domain, they typically require model training, parameter tuning, and retraining when accommodating data updates. As Large Language Models (LLMs) demonstrate increasing capabilities in spatial and graph-based reasoning, there is growing interest in exploring how these models can be applied to geo-spatial problems. We introduce CompassLLM, a novel multi-agent framework that intelligently leverages the reasoning capabilities of LLMs into the geo-spatial domain to solve the popular path query. CompassLLM employs its agents in a two-stage pipeline: the SEARCH stage that identifies popular paths, and a GENERATE stage that synthesizes novel paths in the absence of an existing one in the historical trajectory data. Experiments on real and synthetic datasets show that CompassLLM demonstrates superior accuracy in SEARCH and competitive performance in GENERATE while being cost-effective.
Authors: Chongyu Fan, Changsheng Wang, Yancheng Huang, Soumyadeep Pal, Sijia Liu
Abstract: Machine unlearning for large language models (LLMs) aims to remove undesired data, knowledge, and behaviors (e.g., for safety, privacy, or copyright) while preserving useful model capabilities. Despite rapid progress over the past two years, research in LLM unlearning remains fragmented, with limited clarity on what constitutes effective unlearning and how it should be rigorously evaluated. In this work, we present a principled taxonomy of twelve recent stateful unlearning methods, grouped into three methodological families: divergence-driven optimization, representation misalignment, and rejection-based targeted unlearning. Building on this taxonomy, we revisit the evaluation of unlearning effectiveness (UE), utility retention (UT), and robustness (Rob), focusing on the WMDP benchmark. Our analysis shows that current evaluations, dominated by multiple-choice question (MCQ) accuracy, offer only a narrow perspective, often overstating success while overlooking the model's actual generation behavior. To address this gap, we introduce open question-answering (Open-QA) metrics that better capture generative performance and reveal the inherent UE-UT tradeoff across method families. Furthermore, we demonstrate that robustness requires finer-grained analysis: for example, vulnerabilities differ substantially between in-domain relearning and out-of-domain fine-tuning, even though both fall under model-level attacks. Through this study, we hope to deliver a full-stack revisit of LLM unlearning and actionable guidance for designing and evaluating future methods.
Authors: Yinglun Zhu, Jiancheng Zhang, Fuzhi Tang
Abstract: Frontier AI models have achieved remarkable progress, yet recent studies suggest they struggle with compositional reasoning, often performing at or below random chance on established benchmarks. We revisit this problem and show that widely used evaluation metrics systematically underestimate model capability. To address this, we introduce a group matching score that better exploits group structure and reveals substantial hidden capability in both contrastive vision-language models (VLMs) and multimodal large language models (MLLMs). Moreover, simply overfitting to the induced group matchings at test time transfers this hidden capability into higher scores under standard evaluation metrics, closing much of the reported gap. This adjustment enables SigLIP-B16 to surpass all previous results and GPT-4.1 to yield the first result surpassing estimated human performance on Winoground. Building on this insight, we propose Test-Time Matching (TTM), an iterative, self-improving algorithm that further bootstraps model performance without any external supervision. TTM delivers additional, non-trivial improvements: for example, TTM enables SigLIP-B16 to surpass GPT-4.1 on MMVP-VLM, establishing a new state of the art. Importantly, TTM remains broadly effective even on benchmarks without metric-induced effects or group structures, achieving relative gains up to 85.7% on challenging datasets such as WhatsUp. Across 16 dataset variants spanning diverse setups, our experiments demonstrate that TTM consistently improves model performance and advances the frontier of compositional reasoning.
Authors: Yuhan Sun, Zhiwei Huang, Wanqing Cui, Shaopan Xiong, Yazhi Guo, Meiguang Jin, Junfeng Ma
Abstract: In AI-powered e-commerce livestreaming, digital avatars require real-time responses to drive engagement, a task for which high-latency Large Reasoning Models (LRMs) are ill-suited. We introduce LiveThinking, a practical two-stage optimization framework to bridge this gap. First, we address computational cost by distilling a 670B teacher LRM into a lightweight 30B Mixture-of-Experts (MoE) model (3B active) using Rejection Sampling Fine-Tuning (RFT). This reduces deployment overhead but preserves the teacher's verbose reasoning, causing latency. To solve this, our second stage employs reinforcement learning with Group Relative Policy Optimization (GRPO) to compress the model's reasoning path, guided by a multi-objective reward function balancing correctness, helpfulness, and brevity. LiveThinking achieves a 30-fold reduction in computational cost, enabling sub-second latency. In real-world application on Taobao Live, it improved response correctness by 3.3% and helpfulness by 21.8%. Tested by hundreds of thousands of viewers, our system led to a statistically significant increase in Gross Merchandise Volume (GMV), demonstrating its effectiveness in enhancing user experience and commercial performance in live, interactive settings.
Authors: Alhim Vera, Karen Sanchez, Carlos Hinojosa, Haidar Bin Hamid, Donghoon Kim, Bernard Ghanem
Abstract: Can generative agents be trusted in multimodal environments? Despite advances in large language and vision-language models that enable agents to act autonomously and pursue goals in rich settings, their ability to reason about safety, coherence, and trust across modalities remains limited. We introduce a reproducible simulation framework for evaluating agents along three dimensions: (1) safety improvement over time, including iterative plan revisions in text-visual scenarios; (2) detection of unsafe activities across multiple categories of social situations; and (3) social dynamics, measured as interaction counts and acceptance ratios of social exchanges. Agents are equipped with layered memory, dynamic planning, multimodal perception, and are instrumented with SocialMetrics, a suite of behavioral and structural metrics that quantifies plan revisions, unsafe-to-safe conversions, and information diffusion across networks. Experiments show that while agents can detect direct multimodal contradictions, they often fail to align local revisions with global safety, reaching only a 55 percent success rate in correcting unsafe plans. Across eight simulation runs with three models - Claude, GPT-4o mini, and Qwen-VL - five agents achieved average unsafe-to-safe conversion rates of 75, 55, and 58 percent, respectively. Overall performance ranged from 20 percent in multi-risk scenarios with GPT-4o mini to 98 percent in localized contexts such as fire/heat with Claude. Notably, 45 percent of unsafe actions were accepted when paired with misleading visuals, showing a strong tendency to overtrust images. These findings expose critical limitations in current architectures and provide a reproducible platform for studying multimodal safety, coherence, and social dynamics.
Authors: Peiyang Liu, Ziqiang Cui, Di Liang, Wei Ye
Abstract: Retrieval-augmented generation (RAG) enhances Large Language Models (LLMs) by mitigating hallucinations and outdated information issues, yet simultaneously facilitates unauthorized data appropriation at scale. This paper addresses this challenge through two key contributions. First, we introduce RPD, a novel dataset specifically designed for RAG plagiarism detection that encompasses diverse professional domains and writing styles, overcoming limitations in existing resources. Second, we develop a dual-layered watermarking system that embeds protection at both semantic and lexical levels, complemented by an interrogator-detective framework that employs statistical hypothesis testing on accumulated evidence. Extensive experimentation demonstrates our approach's effectiveness across varying query volumes, defense prompts, and retrieval parameters, while maintaining resilience against adversarial evasion techniques. This work establishes a foundational framework for intellectual property protection in retrieval-augmented AI systems.
Authors: Ruiling Xu, Yifan Zhang, Qingyun Wang, Carl Edwards, Heng Ji
Abstract: Organic reaction mechanisms are the stepwise elementary reactions by which reactants form intermediates and products, and are fundamental to understanding chemical reactivity and designing new molecules and reactions. Although large language models (LLMs) have shown promise in understanding chemical tasks such as synthesis design, it is unclear to what extent this reflects genuine chemical reasoning capabilities, i.e., the ability to generate valid intermediates, maintain chemical consistency, and follow logically coherent multi-step pathways. We address this by introducing oMeBench, the first large-scale, expert-curated benchmark for organic mechanism reasoning in organic chemistry. It comprises over 10,000 annotated mechanistic steps with intermediates, type labels, and difficulty ratings. Furthermore, to evaluate LLM capability more precisely and enable fine-grained scoring, we propose oMeS, a dynamic evaluation framework that combines step-level logic and chemical similarity. We analyze the performance of state-of-the-art LLMs, and our results show that although current models display promising chemical intuition, they struggle with correct and consistent multi-step reasoning. Notably, we find that using prompting strategy and fine-tuning a specialist model on our proposed dataset increases performance by 50% over the leading closed-source model. We hope that oMeBench will serve as a rigorous foundation for advancing AI systems toward genuine chemical reasoning.
Authors: Raisa M. Simoes, Timoteo Kelly, Eduardo J. Simoes, Praveen Rao
Abstract: This paper aims to explore two competing data science methodologies to attempt answering the question, "Which issues contributed most to voters' choice in the 2024 presidential election?" The methodologies involve novel empirical evidence driven by artificial intelligence (AI) techniques. By using two distinct methods based on natural language processing and clustering analysis to mine over eight thousand user comments on election-related YouTube videos from one right leaning journal, Wall Street Journal, and one left leaning journal, New York Times, during pre-election week, we quantify the frequency of selected issue areas among user comments to infer which issues were most salient to potential voters in the seven days preceding the November 5th election. Empirically, we primarily demonstrate that immigration and democracy were the most frequently and consistently invoked issues in user comments on the analyzed YouTube videos, followed by the issue of identity politics, while inflation was significantly less frequently referenced. These results corroborate certain findings of post-election surveys but also refute the supposed importance of inflation as an election issue. This indicates that variations on opinion mining, with their analysis of raw user data online, can be more revealing than polling and surveys for analyzing election outcomes.
Authors: Weisen Jiang, Sinno Jialin Pan
Abstract: This paper introduces MetaDefense, a novel framework for defending against finetuning-based jailbreak attacks in large language models (LLMs). We observe that existing defense mechanisms fail to generalize to harmful queries disguised by unseen attack templates, despite LLMs being capable of distinguishing disguised harmful queries in the embedding space. Based on these insights, we propose a two-stage defense approach: (i) pre-generation defense that detects harmful queries before response generation begins, and (ii) mid-generation defense that monitors partial responses during generation to prevent outputting more harmful content. Our MetaDefense trains the LLM to predict the harmfulness of both queries and partial responses using specialized prompts, enabling early termination of potentially harmful interactions. Extensive experiments across multiple LLM architectures (LLaMA-2-7B, Qwen-2.5-3B-Instruct, and LLaMA-3.2-3B-Instruct) demonstrate that MetaDefense significantly outperforms existing defense mechanisms, achieving robust defense against harmful queries with seen and unseen attack templates while maintaining competitive performance on benign tasks. Code is available at https://github.com/ws-jiang/MetaDefense.
Authors: Emre Can Acikgoz, Cheng Qian, Heng Ji, Dilek Hakkani-T\"ur, Gokhan Tur
Abstract: One paradigm of language model (LM) fine-tuning relies on creating large training datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information or is redundant with the knowledge already acquired by the model, resulting in unnecessary costs. In this work, we explore a new test-time self-improvement method to create more effective and generalizable agentic LMs on-the-fly. The proposed algorithm can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time fine-tuning (self-improvement). We study two variants of this approach: Test-Time Self-Improvement (TT-SI), where the same model generates additional training examples from its own uncertain cases and then learns from them, and contrast this approach with Test-Time Distillation (TT-D), where a stronger model generates similar examples for uncertain cases, enabling student to adapt using distilled supervision. Empirical evaluations across different agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all benchmarks and surpasses other standard learning methods, yet using 68x less training samples. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement algorithms at test-time as a new paradigm for building more capable agents toward self-evolution.
Authors: Leigang Qu, Ziyang Wang, Na Zheng, Wenjie Wang, Liqiang Nie, Tat-Seng Chua
Abstract: Video Foundation Models (VFMs) exhibit remarkable visual generation performance, but struggle in compositional scenarios (e.g., motion, numeracy, and spatial relation). In this work, we introduce Test-Time Optimization and Memorization (TTOM), a training-free framework that aligns VFM outputs with spatiotemporal layouts during inference for better text-image alignment. Rather than direct intervention to latents or attention per-sample in existing work, we integrate and optimize new parameters guided by a general layout-attention objective. Furthermore, we formulate video generation within a streaming setting, and maintain historical optimization contexts with a parametric memory mechanism that supports flexible operations, such as insert, read, update, and delete. Notably, we found that TTOM disentangles compositional world knowledge, showing powerful transferability and generalization. Experimental results on the T2V-CompBench and Vbench benchmarks establish TTOM as an effective, practical, scalable, and efficient framework to achieve cross-modal alignment for compositional video generation on the fly.
Authors: Dhruv Jain, Harshit Shukla, Gautam Rajeev, Ashish Kulkarni, Chandra Khatri, Shubham Agarwal
Abstract: Large-scale Speech Language Models (SpeechLMs) have enabled voice assistants capable of understanding natural spoken queries and performing complex tasks. However, existing speech benchmarks primarily focus on isolated capabilities such as transcription, or question-answering, and do not systematically evaluate agentic scenarios encompassing multilingual and cultural understanding, as well as adversarial robustness. To address this, we introduce VoiceAgentBench, a comprehensive benchmark designed to evaluate SpeechLMs in realistic spoken agentic settings. It comprises over 5,500 synthetic spoken queries, including dialogues grounded in Indian context, covering single-tool invocations, multi-tool workflows, multi-turn interactions, and safety evaluations. The benchmark supports English, Hindi, and 5 other Indian languages, reflecting real-world linguistic and cultural diversity. We simulate speaker variability using a novel sampling algorithm that selects audios for TTS voice conversion based on its speaker embeddings, maximizing acoustic and speaker diversity. Our evaluation measures tool selection accuracy, structural consistency, and the correctness of tool invocations, including adversarial robustness. Our experiments reveal significant gaps in contextual tool orchestration tasks, Indic generalization, and adversarial robustness, exposing critical limitations of current SpeechLMs.
Authors: Yi-Cheng Lin, Yu-Hsuan Li Liang, Hsuan Su, Tzu-Quan Lin, Shang-Tse Chen, Yun-Nung Chen, Hung-yi Lee
Abstract: Robust ASR under domain shift is crucial because real-world systems encounter unseen accents and domains with limited labeled data. Although pseudo-labeling offers a practical workaround, it often introduces systematic, accent-specific errors that filtering fails to fix. We ask: How can we correct these recurring biases without target ground truth? We propose a simple parameter-space correction: in a source domain containing both real and pseudo-labeled data, two ASR models are fine-tuned from the same initialization, one on ground-truth labels and the other on pseudo-labels, and their weight difference forms a correction vector that captures pseudo-label biases. When applied to a pseudo-labeled target model, this vector enhances recognition, achieving up to a 35% relative Word Error Rate (WER) reduction on AfriSpeech-200 across ten African accents with the Whisper tiny model.
Authors: Jianhui Yang, Yiming Jin, Pengkun Jiao, Chenhe Dong, Zerui Huang, Shaowei Yao, Xiaojiang Zhou, Dan Ou, Haihong Tang
Abstract: Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning and slowing convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale real-world datasets and through online side-by-side human evaluations on Taobao Search. It consistently outperforms DPO and standard GRPO baselines in offline experiments, improving relevance accuracy, rule adherence, and training stability. The model trained with TaoSR-AGRL has been successfully deployed in the main search scenario on Taobao, serving hundreds of millions of users.
Authors: Xiaochong Lan, Jie Feng, Yinxing Liu, Xinlei Shi, Yong Li
Abstract: Ranking online reviews by their intrinsic quality is a critical task for e-commerce platforms and information services, impacting user experience and business outcomes. However, quality is a domain-dependent and dynamic concept, making its assessment a formidable challenge. Traditional methods relying on hand-crafted features are unscalable across domains and fail to adapt to evolving content patterns, while modern deep learning approaches often produce black-box models that lack interpretability and may prioritize semantics over quality. To address these challenges, we propose AutoQual, an LLM-based agent framework that automates the discovery of interpretable features. While demonstrated on review quality assessment, AutoQual is designed as a general framework for transforming tacit knowledge embedded in data into explicit, computable features. It mimics a human research process, iteratively generating feature hypotheses through reflection, operationalizing them via autonomous tool implementation, and accumulating experience in a persistent memory. We deploy our method on a large-scale online platform with a billion-level user base. Large-scale A/B testing confirms its effectiveness, increasing average reviews viewed per user by 0.79% and the conversion rate of review readers by 0.27%.
Authors: Daniel Huwiler, Kurt Stockinger, Jonathan F\"urst
Abstract: Retrieval-Augmented Generation (RAG) systems fail when documents evolve through versioning-a ubiquitous characteristic of technical documentation. Existing approaches achieve only 58-64% accuracy on version-sensitive questions, retrieving semantically similar content without temporal validity checks. We present VersionRAG, a version-aware RAG framework that explicitly models document evolution through a hierarchical graph structure capturing version sequences, content boundaries, and changes between document states. During retrieval, VersionRAG routes queries through specialized paths based on intent classification, enabling precise version-aware filtering and change tracking. On our VersionQA benchmark-100 manually curated questions across 34 versioned technical documents-VersionRAG achieves 90% accuracy, outperforming naive RAG (58%) and GraphRAG (64%). VersionRAG reaches 60% accuracy on implicit change detection where baselines fail (0-10%), demonstrating its ability to track undocumented modifications. Additionally, VersionRAG requires 97% fewer tokens during indexing than GraphRAG, making it practical for large-scale deployment. Our work establishes versioned document QA as a distinct task and provides both a solution and benchmark for future research.
Authors: Ali Mazyaki, Mohammad Naghizadeh, Samaneh Ranjkhah Zonouzaghi, Amirhossein Farshi Sotoudeh
Abstract: Responsible AI demands systems whose behavioral tendencies can be effectively measured, audited, and adjusted to prevent inadvertently nudging users toward risky decisions or embedding hidden biases in risk aversion. As language models (LMs) are increasingly incorporated into AI-driven decision support systems, understanding their risk behaviors is crucial for their responsible deployment. This study investigates the manipulability of risk aversion (MoRA) in LMs, examining their ability to replicate human risk preferences across diverse economic scenarios, with a focus on gender-specific attitudes, uncertainty, role-based decision-making, and the manipulability of risk aversion. The results indicate that while LMs such as DeepSeek Reasoner and Gemini-2.0-flash-lite exhibit some alignment with human behaviors, notable discrepancies highlight the need to refine bio-centric measures of manipulability. These findings suggest directions for refining AI design to better align human and AI risk preferences and enhance ethical decision-making. The study calls for further advancements in model design to ensure that AI systems more accurately replicate human risk preferences, thereby improving their effectiveness in risk management contexts. This approach could enhance the applicability of AI assistants in managing risk.
Authors: Haolin Yang, Yuxing Long, Zhuoyuan Yu, Zihan Yang, Minghan Wang, Jiapeng Xu, Yihan Wang, Ziyan Yu, Wenzhe Cai, Lei Kang, Hao Dong
Abstract: Instruction-following navigation is a key step toward embodied intelligence. Prior benchmarks mainly focus on semantic understanding but overlook systematically evaluating navigation agents' spatial perception and reasoning capabilities. In this work, we introduce the NavSpace benchmark, which contains six task categories and 1,228 trajectory-instruction pairs designed to probe the spatial intelligence of navigation agents. On this benchmark, we comprehensively evaluate 22 navigation agents, including state-of-the-art navigation models and multimodal large language models. The evaluation results lift the veil on spatial intelligence in embodied navigation. Furthermore, we propose SNav, a new spatially intelligent navigation model. SNav outperforms existing navigation agents on NavSpace and real robot tests, establishing a strong baseline for future work.
Authors: Yi Lu, Jianing Wang, Linsen Guo, Wei He, Hongyin Tang, Tao Gui, Xuanjing Huang, Xuezhi Cao, Wei Wang, Xunliang Cai
Abstract: Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models' ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning behaviors in LRMs through query composition. Based on R-HORIZON, we construct a long-horizon reasoning benchmark, comprising complex multi-step reasoning tasks with interdependent problems that span long reasoning horizons. Through comprehensive evaluation of LRMs using the R-HORIZON benchmark, we find that even the most advanced LRMs suffer significant performance degradation. Our analysis reveals that LRMs exhibit limited effective reasoning length and struggle to allocate thinking budget across multiple problems appropriately. Recognizing these limitations, we use R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). Compared to training with single-horizon data, RLVR with R-HORIZON not only substantially improves performance on the multi-horizon reasoning tasks, but also promotes accuracy on standard reasoning tasks, with an increase of 7.5 on AIME2024. These results position R-HORIZON as a scalable, controllable, and low-cost paradigm for enhancing and evaluating the long-horizon reasoning capabilities of LRMs.
Authors: Lirui Guo, Michael G. Burke, Wynita M. Griggs
Abstract: Shared Autonomous Vehicles (SAVs) are likely to become an important part of the transportation system, making effective human-SAV interactions an important area of research. This paper introduces a dataset of 200 human-SAV interactions to further this area of study. We present an open-source human-SAV conversational dataset, comprising both textual data (e.g., 2,136 human-SAV exchanges) and empirical data (e.g., post-interaction survey results on a range of psychological factors). The dataset's utility is demonstrated through two benchmark case studies: First, using random forest modeling and chord diagrams, we identify key predictors of SAV acceptance and perceived service quality, highlighting the critical influence of response sentiment polarity (i.e., perceived positivity). Second, we benchmark the performance of an LLM-based sentiment analysis tool against the traditional lexicon-based TextBlob method. Results indicate that even simple zero-shot LLM prompts more closely align with user-reported sentiment, though limitations remain. This study provides novel insights for designing conversational SAV interfaces and establishes a foundation for further exploration into advanced sentiment modeling, adaptive user interactions, and multimodal conversational systems.
Authors: Jianlyu Chen, Junwei Lan, Chaofan Li, Defu Lian, Zheng Liu
Abstract: In this paper, we introduce ReasonEmbed, a novel text embedding model developed for reasoning-intensive document retrieval. Our work includes three key technical contributions. First, we propose ReMixer, a new data synthesis method that overcomes the triviality problem prevalent in previous synthetic datasets, enabling large-scale production of 82K high-quality training samples. Second, we design Redapter, a self-adaptive learning algorithm that dynamically adjusts training each sample's weight based on its reasoning intensity. This allows the model to effectively capture the complex semantic relationships between queries and documents. Third, we implement ReasonEmbed across multiple backbones of varying sizes, all of which achieve superior performance on reasoning-intensive retrieval tasks. Notably, our ReasonEmbed-Qwen3-8B model offers a record-high nDCG@10 score of 38.1 on the BRIGHT benchmark, which significantly outperforms existing text embedding models. We will fully open-source our created resources in ReasonEmbed to push forward the research advancement in this field.
Authors: Marta Emili Garcia Segura, Stephen Hailes, Mirco Musolesi
Abstract: Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players' learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner's Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner's Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.
Authors: Jason Bohne, Pawel Polak, David Rosenberg, Brian Bloniarz, Gary Kazantsev
Abstract: Direct Preference Optimization (DPO) has recently emerged as a simple and effective alternative to reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with user preferences. However, existing DPO formulations rely on a single monolithic model, which limits their expressivity in multi-task settings and their adaptability to heterogeneous or diverse preference distributions. In this work, we propose Mix- and MoE-DPO, a framework that extends DPO with both soft mixture models and mixture-of-experts (MoE) architectures, using a stochastic variational inference approach. Our method introduces a latent-variable model over expert assignments and optimizes a variational evidence lower bound (ELBO), enabling stable and efficient learning of specialized expert policies from preference data. Mix- and MoE-DPO provides three key advantages over standard DPO: (i) generalization via universal function approximation through mixtures; (ii) reward and policy specialization through expert components tailored to distinct preference modes; and (iii) contextual alignment through input-dependent soft gating that enables user-specific mixture policies. Our framework supports both shared base architectures with expert-specific policy heads and fully independent expert models, allowing flexible trade-offs between parameter efficiency and specialization. We validate our approach on a variety of model sizes and multi-preference datasets, demonstrating that Mix- and MoE-DPO offers a powerful and scalable method for preference-based LLM alignment.
Authors: Marius Dragoi, Ioana Pintilie, Florin Gogianu, Florin Brad
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm to improve Large Language Models on reasoning tasks such as coding, math or logic. To assess the reasoning boundary (the fraction of problems a model can solve) researchers often report Pass@k at large sampling budgets. Recent results reveal a crossover phenomenon: while RLVR models outperform the base model at small k values, the base model usually outperforms them when sampling a very large number of completions. This has been interpreted as evidence that base models have a larger reasoning boundary. We argue that on tasks with discrete answer spaces, such as math with numeric outputs, Pass@k at large k reflects the increasingly higher chance of success in the limit of the number of trials rather than genuine reasoning, and can therefore be misleading. We propose Cover@tau, which measures the fraction of problems that a model can solve for which at least a tau proportion of completions are correct. Unlike Pass@k, Cover@tau captures reasoning under an explicit reliability threshold: models that rely on random guessing degrade rapidly as tau increases. We evaluate several RLVR models using Cover@tau-based metrics and illustrate how the relative rankings of popular algorithms change compared to Pass@1, offering a different perspective on reasoning boundaries.
Authors: Heming Zou, Yunliang Zang, Wutong Xu, Yao Zhu, Xiangyang Ji
Abstract: Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. Inspired by the fly olfactory circuit, we propose FlyLoRA, an implicit MoE-based LoRA variant that introduces: (1) rank-wise expert activation in the up-projection matrix, and (2) an implicit router that unifies expert routing and down-projection, where a frozen sparse random projection matrix replaces the traditional dense trainable version. This design resolves the trade-off between intra-task decorrelation and computational efficiency by eliminating the need for an explicit router, while inherently mitigating inter-task interference due to the orthogonality property of random matrices. Extensive experiments across four domains -- general knowledge understanding, scientific question answering, mathematical reasoning, and code generation -- demonstrate consistent performance improvements over existing methods. Beyond empirical gains, FlyLoRA highlights how biological structures can inspire innovations in AI technologies. Code is available at https://github.com/gfyddha/FlyLoRA.
Authors: Cheng Qian, Zuxin Liu, Shirley Kokane, Akshara Prabhakar, Jielin Qiu, Haolin Chen, Zhiwei Liu, Heng Ji, Weiran Yao, Shelby Heinecke, Silvio Savarese, Caiming Xiong, Huan Wang
Abstract: Modern LLM deployments confront a widening cost-performance spectrum: premium models deliver strong reasoning but are expensive, while lightweight models are economical yet brittle on complex tasks. Static escalation rules and keyword heuristics under-utilize this spectrum and fail to adapt across task types. We present xRouter, a tool-calling-based routing system in which a learned router can either answer directly or invoke one or more external models. The router is trained end-to-end with reinforcement learning using an explicit, cost-aware reward that encodes cost-performance trade-offs, eliminating the need for hand-engineered routing rules. Our implementation encompasses the full reinforcement learning framework, including reward and cost accounting, as well as the deployment and evaluation pipelines. Across diverse benchmarks, xRouter achieves strong cost-performance trade-offs (e.g., substantial cost reductions at comparable task completion rates), and provides empirical insights into what reliably helps learned routing and what does not, ranging from model trainability to the difficulty of eliciting sophisticated orchestration behaviors in small open models. We hope these findings and our open implementation will serve as a practical substrate for advancing learned, cost-aware LLM orchestration.
Authors: Bianca-Mihaela Ganescu, Suchir Salhan, Andrew Caines, Paula Buttery
Abstract: Training vision-language models on cognitively-plausible amounts of data requires rethinking how models integrate multimodal information. Within the constraints of the Vision track for the BabyLM Challenge 2025, we propose a lightweight decoder-based architecture with (1) token-wise dynamic gating for adaptive fusion of linguistic and visual cues, (2) feature modulation and channel attention to maximise the utility of limited visual information and (3) auxiliary contrastive objectives for visual grounding. Evaluation on five benchmarks (BLiMP, BLiMP Supplement, EWoK, Winoground and VQA) shows competitive or superior performance to multimodal baselines. More notably, our dynamic gate discovers interpretable patterns without explicit supervision, favouring visual cues for content words and linguistic cues for function words. While we identify limitations in the Challenge constraints, such as the information bottleneck created by global image embeddings and training instability from the dataset split, our findings establish dynamic gating as a powerful tool for efficient multimodal learning, offering both interpretability and performance even under severe constraints.
Authors: Onur Kele\c{s}, Asl{\i} \"Ozy\"urek, Gerardo Ortega, Kadir G\"okg\"o, Esam Ghaleb
Abstract: Iconicity, the resemblance between linguistic form and meaning, is pervasive in signed languages, offering a natural testbed for visual grounding. For vision-language models (VLMs), the challenge is to recover such essential mappings from dynamic human motion rather than static context. We introduce the \textit{Visual Iconicity Challenge}, a novel video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction (e.g., handshape, location), (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess $13$ state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. On \textit{phonological form prediction}, VLMs recover some handshape and location detail but remain below human performance; on \textit{transparency}, they are far from human baselines; and only top models correlate moderately with human \textit{iconicity ratings}. Interestingly, \textit{models with stronger phonological form prediction correlate better with human iconicity judgment}, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks and motivate human-centric signals and embodied learning methods for modelling iconicity and improving visual grounding in multimodal models.
Authors: Jiayun Luo, Wan-Cyuan Fan, Lyuyang Wang, Xiangteng He, Tanzila Rahman, Purang Abolmaesumi, Leonid Sigal
Abstract: Large Vision Language Models (LVLMs) have recently emerged as powerful architectures capable of understanding and reasoning over both visual and textual information. These models typically rely on two key components: a Vision Transformer (ViT) and a Large Language Model (LLM). ViT encodes visual content into a sequence of image tokens and serves as the perceptual front-end -- the eyes of the model. In contrast, the LLM interprets these tokens to perform high-level reasoning, generates responses, and functions as the cognitive core -- the brain of the model. However, it remains unclear which visual tokens contribute most significantly to understanding and reasoning, and how effectively these signals are propagated from ViT to the LLM. While most existing works have focused on identifying attention sinks, low-semantic tokens receiving disproportionately high attention, within the LLM, we shift the focus to the vision encoder by identifying a class of high-norm visual tokens from ViT, referred to as ViT attention sinks -- a problem that has been rarely studied but is indeed very important for LVLMs. Our findings show that these ViT sinks encapsulate high-level semantic concepts from images, allowing the LLM to perform more effective understanding and reasoning. Despite their importance, these sink tokens are often overlooked in existing LVLM architectures. To explore their contribution, we present both qualitative and quantitative analyses of the information embedded in these sink tokens. We also propose both training-free and training-based approaches to better leverage how this information is interpreted by the LLM, and to what extent. By explicitly utilizing these tokens, we demonstrate substantial improvements across a range of LVLMs and visual reasoning tasks, highlighting the untapped potential of ViT attention sinks in enhancing visual reasoning.
Authors: Shangheng Du, Xiangchao Yan, Dengyang Jiang, Jiakang Yuan, Yusong Hu, Xin Li, Liang He, Bo Zhang, Lei Bai
Abstract: Large language models (LLMs) have shown impressive performance in general programming tasks. However, in Machine Learning Engineering (MLE) scenarios such as AutoML and Kaggle competitions, achieving high performance depends heavily on expert intervention and repeated adjustments rather than simply generating correct code. When applied directly to these tasks, LLMs often lack fine-grained domain priors, and existing MLE approaches that use linear or tree-structured searches limit knowledge transfer to adjacent hierarchical links. As a result, they cannot leverage past full trajectories or share information across branches, limiting self-evolving ability and search space diversity. To address these limitations, we introduce AutoMLGen, an LLM-based coding agent that integrates a domain knowledge base for high-quality prior guidance and Monte Carlo Graph Search (MCGS) for efficient exploration. MCGS retains the tree-guided exploration of MCTS while embedding a graph structure into the expansion stage to enable dynamic path reorganization, historical trajectory reuse, and multi-solution fusion to support both self-evolution and collaborative learning. Combined with fine-grained operator sets, this design improves stability and accelerates convergence. Evaluation on the MLE-Bench shows that AutoMLGen achieves state-of-the-art performance in numerous dimensions, such as the average medal rate and the valid submission rate, under a 12-hour budget (half the standard runtime). The code is available at https://github.com/Alpha-Innovator/InternAgent.
Authors: Md Kowsher, Ali O. Polat, Ehsan Mohammady Ardehaly, Mehrdad Salehi, Zia Ghiasi, Prasanth Murali, Chen Chen
Abstract: This paper presents a theoretical framework explaining why fine tuning small, randomly selected subnetworks (slices) within pre trained models can be sufficient for downstream adaptation. We prove that pretrained networks exhibit a universal winning slice property arising from two phenomena: (1) spectral balance the eigenspectra of different weight matrix slices are remarkably similar; and (2) high task energy their backbone representations retain rich, task relevant features. This leads to the Universal Winning Slice Hypothesis, which provides a theoretical foundation for parameter efficient fine tuning (PEFT) in large scale models. Inspired by this, we propose SliceFine, a PEFT method that exploits this inherent redundancy by updating only selected slices of the original weights introducing zero new parameters, unlike adapter-based approaches. Empirically, SliceFine matches the performance of state of the art PEFT methods across language and vision tasks, while significantly improving training speed, memory efficiency, and model compactness. Our work bridges theory and practice, offering a theoretically grounded alternative to existing PEFT techniques.
Authors: Grace Liu, Yuxiao Qu, Jeff Schneider, Aarti Singh, Aviral Kumar
Abstract: Many tasks require learned models to strategically gather relevant information over multiple rounds of interaction before actually acting on a task. Strategic information gathering requires models to know not only how to effectively acquire information, but also when to stop gathering information and make a decision, in order to avoid overthinking or getting derailed when acting. In this paper, we formalize this problem and introduce Counterfactuals and Reasoning for Termination (CaRT), an approach for teaching LLMs when to stop seeking information. To appropriately learn when to terminate, CaRT fine-tunes LLMs using counterfactual pairs of trajectories, one where termination is appropriate and a minimally modified version of the same trajectory where it is not. It trains the LLM to explain the rationale for the termination decision in either case via verbal reasoning, and imbues this capability into the base LLM via fine-tuning. We instantiate CaRT in two domains: interactive medical diagnosis and math problem solving. In both domains, we find that CaRT improves the efficiency of information gathering and task success rate compared to other fine-tuning methods.
Authors: Hongxing Li, Dingming Li, Zixuan Wang, Yuchen Yan, Hang Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu, Jun Xiao, Yueting Zhuang
Abstract: Spatial reasoning remains a fundamental challenge for Vision-Language Models (VLMs), with current approaches struggling to achieve robust performance despite recent advances. We identify that this limitation stems from a critical gap: existing methods attempt to learn spatial reasoning directly without establishing the hierarchical foundations of perception and understanding. To address this challenge, we present a comprehensive methodology for building spatial intelligence progressively. We introduce SpatialLadder-26k, a multimodal dataset containing 26,610 samples spanning object localization, single image, multi-view, and video spatial reasoning tasks, constructed through a standardized pipeline that ensures systematic coverage across modalities. Building on this dataset, we design a three-stage progressive training framework that (1) establishes spatial perception through object localization, (2) develops spatial understanding through multi-dimensional spatial tasks, and (3) strengthens complex reasoning via reinforcement learning with verifiable rewards. This approach yields SpatialLadder, a 3B-parameter model that achieves state-of-the-art performance on spatial reasoning benchmarks, with 23.4% average improvement over the base model, surpassing GPT-4o by 20.8% and Gemini-2.0-Flash by 10.1%. Notably, SpatialLadder maintains strong generalization with 7.2% improvement on out-of-domain benchmarks, demonstrating that progressive training from perception to reasoning is essential for robust spatial intelligence.
Authors: Nikhil Reddy Varimalla, Yunfei Xu, Arkadiy Saakyan, Meng Fan Wang, Smaranda Muresan
Abstract: As Video Large Language Models (VideoLLMs) are deployed globally, they require understanding of and grounding in the relevant cultural background. To properly assess these models' cultural awareness, adequate benchmarks are needed. We introduce VideoNorms, a benchmark of over 1000 (video clip, norm) pairs from US and Chinese cultures annotated with socio-cultural norms grounded in speech act theory, norm adherence and violations labels, and verbal and non-verbal evidence. To build VideoNorms, we use a human-AI collaboration framework, where a teacher model using theoretically-grounded prompting provides candidate annotations and a set of trained human experts validate and correct the annotations. We benchmark a variety of open-weight VideoLLMs on the new dataset which highlight several common trends: 1) models performs worse on norm violation than adherence; 2) models perform worse w.r.t Chinese culture compared to the US culture; 3) models have more difficulty in providing non-verbal evidence compared to verbal for the norm adhere/violation label and struggle to identify the exact norm corresponding to a speech-act; and 4) unlike humans, models perform worse in formal, non-humorous contexts. Our findings emphasize the need for culturally-grounded video language model training - a gap our benchmark and framework begin to address.
Authors: Kai Zhang, Xiangchao Chen, Bo Liu, Tianci Xue, Zeyi Liao, Zhihan Liu, Xiyao Wang, Yuting Ning, Zhaorun Chen, Xiaohan Fu, Jian Xie, Yuxuan Sun, Boyu Gou, Qi Qi, Zihang Meng, Jianwei Yang, Ning Zhang, Xian Li, Ashish Shah, Dat Huynh, Hengduo Li, Zi Yang, Sara Cao, Lawrence Jang, Shuyan Zhou, Jiacheng Zhu, Huan Sun, Jason Weston, Yu Su, Yifan Wu
Abstract: A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
Authors: Tajamul Ashraf, Umair Nawaz, Abdelrahman M. Shaker, Rao Anwer, Philip Torr, Fahad Shahbaz Khan, Salman Khan
Abstract: Vision language models (VLMs) are increasingly deployed as controllers with access to external tools for complex reasoning and decision-making, yet their effectiveness remains limited by the scarcity of high-quality multimodal trajectories and the cost of manual annotation. We address this challenge with a vision-centric agent tuning framework that automatically synthesizes multimodal trajectories, generates step-wise preference pairs, and trains a VLM controller for robust tool-use reasoning. Our pipeline first constructs M-TRACE, a large-scale dataset of 28.5K multimodal tasks with 177K verified trajectories, enabling imitation-based trajectory tuning. Building on this, we develop MATRIX Agent, a controller finetuned on M-TRACE for step-wise tool reasoning. To achieve finer alignment, we further introduce Pref-X, a set of 11K automatically generated preference pairs, and optimize MATRIX on it via step-wise preference learning. Across three benchmarks, Agent-X, GTA, and GAIA, MATRIX consistently surpasses both open- and closed-source VLMs, demonstrating scalable and effective multimodal tool use. Our data and code is avaliable at https://github.com/mbzuai-oryx/MATRIX.
Authors: Pragya Srivastava, Manuj Malik, Vivek Gupta, Tanuja Ganu, Dan Roth
Abstract: Large Language Models (LLMs), excel in natural language understanding, but their capability for complex mathematical reasoning with an amalgamation of structured tables and unstructured text is uncertain. This study explores LLMs' mathematical reasoning on four financial tabular question-answering datasets: TATQA, FinQA, ConvFinQA, and Multihiertt. Through extensive experiments with various models and prompting techniques, we assess how LLMs adapt to complex tables and mathematical tasks. We focus on sensitivity to table complexity and performance variations with an increasing number of arithmetic reasoning steps. The results provide insights into LLMs' capabilities and limitations in handling complex mathematical scenarios for semi-structured tables. Ultimately, we introduce a novel prompting technique tailored to semi-structured documents, matching or outperforming other baselines in performance while providing a nuanced understanding of LLMs abilities for such a task.
Authors: Zhipeng Xu, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Chaojun Xiao, Zhiyuan Liu, Ge Yu, Chenyan Xiong
Abstract: Large Language Models (LLMs) have demonstrated strong performance across a wide range of NLP tasks. However, they often exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. Inspired by constructivist learning theory, we propose ThinkNote, a novel framework that enhances the external knowledge utilization of LLMs through a two-stage constructivist cognitive modeling process. Specifically, ThinkNote performs knowledge assimilation to align new information with the model's parametric memory, forming a coherent internal representation. It then applies thought accommodation to adapt internal reasoning, thereby promoting more consistent and reliable outputs. Extensive experimental results demonstrate that ThinkNote achieves a 10% improvement over strong baseline methods on various question-answering benchmarks. Further analysis indicates that ThinkNote effectively integrates and utilizes external knowledge to help LLMs generate accurate responses and improves their self-consistency. All data and codes are available at https://github.com/OpenMatch/ThinkNote.
Authors: Xiaochong Lan, Zhiguang Han, Yiming Cheng, Li Sheng, Jie Feng, Chen Gao, Yong Li
Abstract: Limited access to mental healthcare resources hinders timely depression diagnosis, leading to detrimental outcomes. Social media platforms present a valuable data source for early detection, yet this task faces two significant challenges: 1) the need for medical knowledge to distinguish clinical depression from transient mood changes, and 2) the dual requirement for high accuracy and model explainability. To address this, we propose DORIS, a framework that leverages Large Language Models (LLMs). To integrate medical knowledge, DORIS utilizes LLMs to annotate user texts against established medical diagnostic criteria and to summarize historical posts into temporal mood courses. These medically-informed features are then used to train an accurate Gradient Boosting Tree (GBT) classifier. Explainability is achieved by generating justifications for predictions based on the LLM-derived symptom annotations and mood course analyses. Extensive experimental results validate the effectiveness as well as interpretability of our method, highlighting its potential as a supportive clinical tool.
Authors: Jing Li, Zhijie Sun, Dachao Lin, Xuan He, Binfan Zheng, Yi Lin, Rongqian Zhao, Xin Chen
Abstract: Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models by activating only a subset of parameters per input. However, existing MoE models suffer from two critical limitations: (1) inefficient token-to-expert routing that causes excessive communication overhead, and (2) expert homogenization that leads to redundant computations. Current approaches address these challenges separately, failing to achieve simultaneous improvements in both training efficiency and model performance. We present Expert-Token Resonance (ETR), a theoretically-grounded bidirectional routing mechanism that fundamentally reimagines expert-token interactions in MoE architectures. Our key insight is that optimal routing requires adaptive coordination between token-choice routing (TCR) during early training phases and expert-choice routing (ECR) in later stages. We prove that this dynamic approach maximizes training success rate (the probability of correct token-expert assignments) while reducing the expert capacity lower bound by up to 40%. ETR incorporates three technical innovations: (1) an affinity-based routing architecture using Grouped Average Pooling (GrAP) that reduces computational complexity from O(d^2) to O(d^2/D) while maintaining orthogonality to prevent expert homogenization; (2) a bidirectional selection mechanism that enables both tokens and experts to actively participate in the routing process based on cosine similarity scores; and (3) an adaptive capacity strategy that dynamically adjusts expert bounds based on training progress, eliminating communication bubbles in All-to-All operations. Extensive experiments on Ascend NPU clusters demonstrate that ETR achieves 5.4%-46.6% improvements in end-to-end training efficiency compared to baseline MoE implementations, with 9.7%-14.5% performance gains across GDAD, GPQA, HumanEval, and TeleQnA benchmarks.
Authors: Shuo Yu (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Mingyue Cheng (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Qi Liu (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Daoyu Wang (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Jiqian Yang (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Jie Ouyang (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Yucong Luo (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China), Chenyi Lei (Kuaishou Technology, Beijing, China), Enhong Chen (State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China)
Abstract: Retrieval-augmented generation (RAG) is increasingly recognized as an effective approach to mitigating the hallucination of large language models (LLMs) through the integration of external knowledge. While numerous efforts, most studies focus on a single type of external knowledge source. However, in real-world applications, most situations involve diverse knowledge from various sources, yet this area has been less explored. The main dilemma is the lack of a suitable dataset containing multiple knowledge sources and pre-exploration of the associated issues. To address these challenges, we standardize a benchmark dataset that combines structured and unstructured knowledge across diverse and complementary domains. Based on this dataset, we further develop a plug-and-play RAG framework, \textbf{PruningRAG}, whose main characteristic is the use of multi-granularity pruning strategies to optimize the integration of relevant information while minimizing misleading context. It consistently improves performance across various existing RAG variants, demonstrating its robustness and broad applicability. Building upon the standardized dataset and PruningRAG, we also report a series of experimental results, as well as insightful findings. Our dataset and code are publicly available\footnote{https://github.com/USTCAGI/PruningRAG}, with the aim of advancing future research in the RAG community.
Authors: Wei Wu, Zhuoshi Pan, Chao Wang, Liyi Chen, Yunchu Bai, Tianfu Wang, Kun Fu, Zheng Wang, Hui Xiong
Abstract: Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of TokenSelect demonstrates up to $23.84\times$ speedup in attention computation and up to $2.28\times$ acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.
Authors: Yaoxiang Wang, Haoling Li, Xin Zhang, Jie Wu, Xiao Liu, Wenxiang Hu, Zhongxin Guo, Yangyu Huang, Ying Xin, Yujiu Yang, Jinsong Su, Qi Chen, Scarlett Li
Abstract: Existing methods for code generation use code snippets as seed data, restricting the complexity and diversity of the synthesized data. In this paper, we introduce a novel feature tree-based synthesis framework, which revolves around hierarchical code features derived from high-level abstractions of code. The feature tree is constructed from raw data and refined iteratively to increase the quantity and diversity of the extracted features, which captures and recognizes more complex patterns and relationships within the code. By adjusting the depth and breadth of the sampled subtrees, our framework provides precise control over the complexity of the generated code, enabling functionalities that range from function-level operations to multi-file scenarios. We fine-tuned widely-used base models to obtain EpiCoder series, achieving state-of-the-art performance on multiple benchmarks at both the function and file levels. In particular, empirical evidence indicates that our approach shows significant potential in the synthesizing of repository-level code data. Our code and data are publicly available at https://github.com/microsoft/EpiCoder.
Authors: Keer Lu, Zheng Liang, Da Pan, Shusen Zhang, Guosheng Dong, Zhonghai Wu, Huang Leng, Bin Cui, Wentao Zhang
Abstract: Large Language Models (LLMs) have exhibited remarkable capabilities in clinical scenarios. Despite their potential, existing works face challenges when applying LLMs to medical settings. Strategies relying on training with medical datasets are highly cost-intensive and may suffer from outdated training data. Leveraging external knowledge bases is a suitable alternative, yet it faces obstacles such as limited retrieval precision and poor effectiveness in answer extraction. These issues collectively prevent LLMs from demonstrating the expected level of proficiency in mastering medical expertise. To address these challenges, we introduce Med-R^2, a novel LLM physician framework that adheres to the Evidence-Based Medicine (EBM) process, efficiently integrating retrieval mechanisms as well as the selection and reasoning processes of evidence, thereby enhancing the problem-solving capabilities of LLMs in healthcare scenarios and fostering a trustworthy LLM physician. Our comprehensive experiments indicate that Med-R^2 achieves a 13.27\% improvement over vanilla RAG methods and even a 4.55\% enhancement compared to fine-tuning strategies, without incurring additional training costs. Furthermore, we find that our LLaMA3.1-70B + Med-R$^2$ surpasses frontier models, including GPT-4o, Claude3.5-Sonnet and DeepSeek-V3 by 1.05\%, 6.14\% and 1.91\%. Med-R$^2$ effectively enhances the capabilities of LLMs in the medical domain.
Authors: Chao-Chung Wu, Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, Shao-Hua Sun, Hung-yi Lee
Abstract: Maintaining consistent model performance across domains is a fundamental challenge in machine learning. While recent work has explored using LLM-generated data for fine-tuning, its impact on cross-domain generalization remains poorly understood. This paper presents a systematic analysis revealing that fine-tuning with LLM-generated data not only improves target task performance but also reduces non-target task degradation compared to fine-tuning with ground truth data. Through analyzing the data sequence in tasks of various domains, we demonstrate that this enhancement of non-target task robustness stems from the reduction of high perplexity tokens found in LLM-generated sequences. Following our findings, we showed that masking high perplexity tokens in ground truth training data achieves similar non-target task performance preservation, comparable to using LLM-generated data. Extensive experiments across different model families and scales, including Gemma 2 IT 2B, Llama 3 8B Instruct, and three additional models, agree with our findings. To the best of our knowledge, this is the first work to provide an empirical explanation based on token perplexity reduction to mitigate catastrophic forgetting in LLMs after fine-tuning, offering valuable insights for developing more robust fine-tuning strategies.
Authors: Andrianos Michail, Simon Clematide, Rico Sennrich
Abstract: The evaluation of cross-lingual semantic search models is often limited to existing datasets from tasks such as information retrieval and semantic textual similarity. We introduce Cross-Lingual Semantic Discrimination (CLSD), a lightweight evaluation task that requires only parallel sentences and a Large Language Model (LLM) to generate adversarial distractors. CLSD measures an embedding model's ability to rank the true parallel sentence above semantically misleading but lexically similar alternatives. As a case study, we construct CLSD datasets for German--French in the news domain. Our experiments show that models fine-tuned for retrieval tasks benefit from pivoting through English, whereas bitext mining models perform best in direct cross-lingual settings. A fine-grained similarity analysis further reveals that embedding models differ in their sensitivity to linguistic perturbations. We release our code and datasets under AGPL-3.0: https://github.com/impresso/cross_lingual_semantic_discrimination
URLs: https://github.com/impresso/cross_lingual_semantic_discrimination
Authors: Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, Yongxin Tong, Zhiming Zheng
Abstract: Current RAG retrievers are designed primarily for human readers, emphasizing complete, readable, and coherent paragraphs. However, LLMs benefit more from precise, compact, and well-structured input, which enhances reasoning quality and efficiency. Existing methods often rely on reranking or summarization to identify key sentences, but may suffer from semantic breaks and unfaithfulness. Thus, efficiently extracting and organizing answer-relevant clues from large-scale documents while reducing LLM reasoning costs remains a challenge for RAG. Inspired by Occam's razor, we frame LLM-centric retrieval as a MinMax optimization: maximizing the extraction of potential clues and reranking them for well-organization, while minimizing reasoning costs by truncating to the smallest sufficient clues set. In this paper, we propose CompSelect, a Compact clue Selection mechanism for LLM-centric RAG, consisting of a clue extractor, a reranker, and a truncator. (1) The clue extractor first uses answer-containing sentences as fine-tuning targets, aiming to extract sufficient potential clues; (2) The reranker is trained to prioritize effective clues based on real LLM feedback; (3) The truncator uses the truncated text containing the minimum sufficient clues for answering the question as fine-tuning targets, thereby enabling efficient RAG reasoning. Experiments on three QA datasets show that CompSelect improves QA performance by approximately 11\% and reduces Total Latency and Online Latency by approximately 17\% and 67\% compared to various baseline methods on both LLaMA3 and Qwen3. Further analysis confirms its robustness to unreliable retrieval and generalization across different scenarios, offering a scalable and cost-efficient solution for web-scale RAG applications.
Authors: Jusen Du, Weigao Sun, Disen Lan, Jiaxi Hu, Yu Cheng
Abstract: Linear sequence modeling methods, such as linear attention, state space modeling, and linear RNNs, offer significant efficiency improvements by reducing the complexity of training and inference. However, these methods typically compress the entire input sequence into a single fixed-size memory state, which leads to suboptimal performance on recall-intensive tasks. To address this limitation, we introduce a novel architecture called Mixture-of-Memories (MoM). MoM utilizes multiple independent memory states, with a router network directing input tokens to specific memory states. This approach greatly enhances the overall memory capacity while minimizing memory interference. MoM serves as a general framework that can be seamlessly combined with diverse memory update mechanisms across linear models. As a result, MoM performs exceptionally well on recall-intensive tasks, surpassing existing linear sequence modeling techniques. Despite incorporating multiple memory states, the computation of each memory state remains linear in complexity, allowing MoM to retain the linear-complexity advantage during training, while constant-complexity during inference. Our experimental results show that MoM outperforms current linear sequence models on downstream language tasks, particularly recall-intensive tasks, and even achieves performance comparable to Transformer models. The code is released at https://github.com/OpenSparseLLMs/MoM and is also released as a part of https://github.com/OpenSparseLLMs/Linear-MoE.
URLs: https://github.com/OpenSparseLLMs/MoM, https://github.com/OpenSparseLLMs/Linear-MoE.
Authors: Xiaochen Wang, Heming Xia, Jialin Song, Longyu Guan, Yixin Yang, Qingxiu Dong, Weiyao Luo, Yifan Pu, Yiru Wang, Xiangdi Meng, Wenjie Li, Zhifang Sui
Abstract: Large Multimodal Models (LMMs) have achieved remarkable success across various visual-language tasks. However, existing benchmarks predominantly focus on single-image understanding, leaving the analysis of image sequences largely unexplored. To address this limitation, we introduce StripCipher, a comprehensive benchmark designed to evaluate capabilities of LMMs to comprehend and reason over sequential images. StripCipher comprises a human-annotated dataset and three challenging subtasks: visual narrative comprehension, contextual frame prediction, and temporal narrative reordering. Our evaluation of 16 state-of-the-art LMMs, including GPT-4o and Qwen2.5VL, reveals a significant performance gap compared to human capabilities, particularly in tasks that require reordering shuffled sequential images. For instance, GPT-4o achieves only 23.93% accuracy in the reordering subtask, which is 56.07% lower than human performance. Further quantitative analysis discuss several factors, such as input format of images, affecting the performance of LLMs in sequential understanding, underscoring the fundamental challenges that remain in the development of LMMs.
Authors: Huazheng Wang, Yongcheng Jing, Haifeng Sun, Yingjie Wang, Jingyu Wang, Jianxin Liao, Dacheng Tao
Abstract: In this paper, we investigate knowledge forgetting in large language models with a focus on its generalisation, ensuring that models forget not only specific training samples but also related implicit knowledge. To this end, we begin by identifying a broader unlearning scope that includes both target data and logically associated samples, including rephrased, subject-replaced, relation-reversed, and one-hop reasoned data. We then conduct a rigorous evaluation of 15 state-of-the-art methods across three datasets, revealing that unlearned models still recall paraphrased answers and retain target facts in their intermediate layers. This motivates us to take a preliminary step toward more generalised implicit knowledge forgetting by proposing PerMU, a novel probability perturbation-based unlearning paradigm. PerMU simulates adversarial unlearning samples to eliminate fact-related tokens from the logit distribution, collectively reducing the probabilities of all answer-associated tokens. Experiments are conducted on a diverse range of datasets, including TOFU, Harry Potter, ZsRE, WMDP, and MUSE, using models ranging from 1.3B to 13B in scale. The results demonstrate that PerMU delivers up to a 50.40% improvement in unlearning vanilla target data while maintaining a 40.73% boost in forgetting implicit knowledge. Our code can be found in https://github.com/MaybeLizzy/PERMU.
Authors: Moritz Altemeyer, Steffen Eger, Johannes Daxenberger, Yanran Chen, Tim Altendorf, Philipp Cimiano, Benjamin Schiller
Abstract: Large Language Models (LLMs) have revolutionized various Natural Language Generation (NLG) tasks, including Argument Summarization (ArgSum), a key subfield of Argument Mining. This paper investigates the integration of state-of-the-art LLMs into ArgSum systems and their evaluation. In particular, we propose a novel prompt-based evaluation scheme, and validate it through a novel human benchmark dataset. Our work makes three main contributions: (i) the integration of LLMs into existing ArgSum systems, (ii) the development of two new LLM-based ArgSum systems, benchmarked against prior methods, and (iii) the introduction of an advanced LLM-based evaluation scheme. We demonstrate that the use of LLMs substantially improves both the generation and evaluation of argument summaries, achieving state-of-the-art results and advancing the field of ArgSum. We also show that among the four LLMs integrated in (i) and (ii), Qwen-3-32B, despite having the fewest parameters, performs best, even surpassing GPT-4o.
Authors: Fajri Koto, Rituraj Joshi, Nurdaulet Mukhituly, Yuxia Wang, Zhuohan Xie, Rahul Pal, Daniil Orel, Parvez Mullah, Diana Turmakhan, Maiya Goloburda, Mohammed Kamran, Samujjwal Ghosh, Bokang Jia, Jonibek Mansurov, Mukhammed Togmanov, Debopriyo Banerjee, Nurkhan Laiyk, Akhmed Sakip, Xudong Han, Ekaterina Kochmar, Alham Fikri Aji, Aaryamonvikram Singh, Alok Anil Jadhav, Satheesh Katipomu, Samta Kamboj, Monojit Choudhury, Gurpreet Gosal, Gokulakrishnan Ramakrishnan, Biswajit Mishra, Sarath Chandran, Avraham Sheinin, Natalia Vassilieva, Neha Sengupta, Preslav Nakov
Abstract: Llama-3.1-Sherkala-8B-Chat, or Sherkala-Chat (8B) for short, is a state-of-the-art instruction-tuned open generative large language model (LLM) designed for Kazakh. Sherkala-Chat (8B) aims to enhance the inclusivity of LLM advancements for Kazakh speakers. Adapted from the LLaMA-3.1-8B model, Sherkala-Chat (8B) is trained on 45.3B tokens across Kazakh, English, Russian, and Turkish. With 8 billion parameters, it demonstrates strong knowledge and reasoning abilities in Kazakh, significantly outper-forming existing open Kazakh and multilingual models of similar scale while achieving competitive performance in English. To ensure effective and responsible alignment, we leverage translated instruction datasets, a Kazakhstan-specific instruction dataset that is automatically constructed and manually verified, and Kazakh-specific safety data. We release Sherkala-Chat (8B) as an open-weight model, along with a detailed description of its training, alignment, and evaluation, to support research and real-world applications for Kazakh speakers.
Authors: Hang Zheng, Hongshen Xu, Yuncong Liu, Lu Chen, Pascale Fung, Kai Yu
Abstract: Large language models (LLMs) are prone to hallucination stemming from misaligned self-awareness, particularly when processing queries exceeding their knowledge boundaries. While existing mitigation strategies employ uncertainty estimation or query rejection mechanisms, they suffer from computational efficiency and sacrificed helpfulness. To address these issues, we propose the Explicit Knowledge Boundary Modeling (EKBM) framework, integrating fast and slow reasoning systems to harmonize reliability and usability. The framework first employs a fast-thinking model to generate confidence-labeled responses, enabling immediate utilization of high-confidence outputs, whereas uncertain predictions trigger a slow refinement model for accuracy improvement. To align model behavior with our proposed object, we propose a hybrid training pipeline, enhancing self-awareness without degrading task performance. Evaluations on dialogue state tracking tasks demonstrate that EKBM achieves superior model reliability over uncertainty-based baselines. Further analysis reveals that refinement substantially boosts accuracy while maintaining low computational overhead. The framework establishes a scalable paradigm for deploying reliable LLMs in error-sensitive applications, effectively balancing accuracy and practical utility.
Authors: Jie Wu, Haoling Li, Xin Zhang, Xiao Liu, Yangyu Huang, Jianwen Luo, Yizhen Zhang, Zuchao Li, Ruihang Chu, Yujiu Yang, Scarlett Li
Abstract: Preference learning extends the performance of Code LLMs beyond traditional supervised fine-tuning by leveraging relative quality comparisons. In existing approaches, a set of n candidate solutions is evaluated based on test case success rates, with the candidate demonstrating a higher pass rate being labeled as positive and its counterpart with a lower pass rate as negative. However, because this approach aligns entire failing code blocks rather than pinpointing specific errors, it lacks the granularity necessary to capture meaningful error-correction relationships. As a result, the model is unable to learn more informative error-correction patterns. To address these issues, we propose Target-DPO, a new preference alignment framework that mimics human iterative debugging to refine Code LLMs. Target-DPO explicitly locates error regions and aligns the corresponding tokens via a tailored DPO algorithm. To facilitate it, we introduce the CodeFlow dataset, where samples are iteratively refined until passing tests, with modifications capturing error corrections. Extensive experiments show that a diverse suite of Code LLMs equipped with Target-DPO achieves significant performance gains in code generation and improves on challenging tasks like BigCodeBench. In-depth analysis reveals that Target-DPO yields fewer errors. Code, model and datasets are in: https://github.com/JieWu02/Target-DPO.
Authors: Yansong Ning, Shuowei Cai, Wei Li, Jun Fang, Naiqiang Tan, Hua Chai, Hao Liu
Abstract: On-demand ride-hailing services like DiDi, Uber, and Lyft have transformed urban transportation, offering unmatched convenience and flexibility. In this paper, we introduce DiMA, an LLM-powered ride-hailing assistant deployed in DiDi Chuxing. Its goal is to provide seamless ride-hailing services and beyond through a natural and efficient conversational interface under dynamic and complex spatiotemporal urban contexts. To achieve this, we propose a spatiotemporal-aware order planning module that leverages external tools for precise spatiotemporal reasoning and progressive order planning. Additionally, we develop a cost-effective dialogue system that integrates multi-type dialog repliers with cost-aware LLM configurations to handle diverse conversation goals and trade-off response quality and latency. Furthermore, we introduce a continual fine-tuning scheme that utilizes real-world interactions and simulated dialogues to align the assistant's behavior with human preferred decision-making processes. Since its deployment in the DiDi application, DiMA has demonstrated exceptional performance, achieving 93% accuracy in order planning and 92% in response generation during real-world interactions. Offline experiments further validate DiMA capabilities, showing improvements of up to 70.23% in order planning and 321.27% in response generation compared to three state-of-the-art agent frameworks, while reducing latency by $0.72\times$ to $5.47\times$. These results establish DiMA as an effective, efficient, and intelligent mobile assistant for ride-hailing services. Our project is released at https://github.com/usail-hkust/DiMA and we also release the MCP service (https://mcp.didichuxing.com/api) to foster the ride-hailing research community.
URLs: https://github.com/usail-hkust/DiMA, https://mcp.didichuxing.com/api)
Authors: Hojun Cho, Donghu Kim, Soyoung Yang, Chan Lee, Hunjoo Lee, Jaegul Choo
Abstract: Language agents powered by large language models (LLMs) face significant deployment challenges in resource-constrained environments, particularly for specialized domains and less-common languages. This paper presents Tox-chat, a Korean chemical toxicity information agent devised within these limitations. We propose two key innovations: a context-efficient architecture that reduces token consumption through hierarchical section search, and a scenario-based dialogue generation methodology that effectively distills tool-using capabilities from larger models. Experimental evaluations demonstrate that our fine-tuned 8B parameter model substantially outperforms both untuned models and baseline approaches, in terms of DB faithfulness and preference. Our work offers valuable insights for researchers developing domain-specific language agents under practical constraints.
Authors: Zhendong Chu, Jian Xie, Shen Wang, Zichao Wang, Qingsong Wen
Abstract: Education materials for K-12 students often consist of multiple modalities, such as text and images, posing challenges for models to fully understand nuanced information in these materials. In this paper, we propose a unified language and vision assistant UniEDU designed for various educational applications, including knowledge recommendation, knowledge tracing, time cost prediction, and user answer prediction, all within a single model. Unlike conventional task-specific models, UniEDU offers a unified solution that excels across multiple educational tasks while maintaining strong generalization capabilities. Its adaptability makes it well-suited for real-world deployment in diverse learning environments. Furthermore, UniEDU is optimized for industry-scale deployment by significantly reducing computational overhead-achieving approximately a 300\% increase in efficiency-while maintaining competitive performance with minimal degradation compared to fully fine-tuned models. This work represents a significant step toward creating versatile AI systems tailored to the evolving demands of education.
Authors: Xuan Luo, Weizhi Wang, Xifeng Yan
Abstract: Various layer-skipping methods have been proposed to accelerate token generation in large language models (LLMs). However, limited attention has been paid to a fundamental question: How do computational demands vary across the generation of different tokens? In this work, we introduce FlexiDepth, a method that dynamically adjusts the number of Transformer layers used in text generation. By incorporating a plug-in router and adapter, FlexiDepth enables adaptive computation in LLMs without modifying their original parameters. Applied to Llama-3-8B, it skips 8 out of 32 layers while maintaining full benchmark performance. Our experiments reveal that computational demands in LLMs significantly vary based on token type. Specifically, generating repetitive tokens or fixed phrases requires fewer layers, whereas producing tokens involving computation or high uncertainty requires more layers. Despite the computational savings, FlexiDepth does not yet achieve wall-clock speedup due to varied skipping patterns and I/O overhead. To inspire future work and advance research on practical speedup, we open-sourced FlexiDepth and a dataset documenting its layer allocation patterns.
Authors: Ziyi Liu, Priyanka Dey, Jen-tse Huang, Zhenyu Zhao, Bowen Jiang, Rahul Gupta, Yang Liu, Yao Du, Jieyu Zhao
Abstract: Cultural Intelligence (CQ) refers to the ability to understand unfamiliar cultural contexts, a crucial skill for large language models (LLMs) to effectively engage with globally diverse users. Existing studies often focus on explicitly stated cultural norms, but fail to capture the subtle, implicit values that are common in daily conversation. To address this gap, we introduce CQBench, a benchmark specifically designed to assess LLMs' capability to infer implicit cultural values from natural conversational contexts. CQBench consists of multi character conversation based stories using values from the World Value Survey and the GlobalOpinions, with topics including ethical, religious, social, etc. Our automatic dataset construction pipeline integrates rigorous validation procedures (incorporation, consistency, and implicitness checks), achieving a 94.5% human model agreement in the final validation. To leverage CQBench data, we design three tasks of increasing complexity: attitude detection, value selection, and value extraction. These tasks evaluate whether models can detect attitude and recognize values embedded within natural dialogues rather than relying on explicit cultural knowledge. We find that while frontier models like o1 reach human level performance in value selection (0.809 F1), they still fall short in nuanced attitude detection (0.622 F1). Notably, finetuning a smaller LLaMA-3.2-3B on only 500 culturally rich examples improves performance by over 10%, even outperforming o3-mini in some cases. Using CQ-Bench, we provide insights into the current challenges in LLMs' CQ research and suggest practical pathways for enhancing LLMs' cross-cultural reasoning abilities.
Authors: Alexandra Bazarova, Aleksandr Yugay, Andrey Shulga, Alina Ermilova, Andrei Volodichev, Konstantin Polev, Julia Belikova, Rauf Parchiev, Dmitry Simakov, Maxim Savchenko, Andrey Savchenko, Serguei Barannikov, Alexey Zaytsev
Abstract: Hallucination, i.e., generating factually incorrect content, remains a critical challenge for large language models (LLMs). We introduce TOHA, a TOpology-based HAllucination detector in the RAG setting, which leverages a topological divergence metric to quantify the structural properties of graphs induced by attention matrices. Examining the topological divergence between prompt and response subgraphs reveals consistent patterns: higher divergence values in specific attention heads correlate with hallucinated outputs, independent of the dataset. Extensive experiments - including evaluation on question answering and summarization tasks - show that our approach achieves state-of-the-art or competitive results on several benchmarks while requiring minimal annotated data and computational resources. Our findings suggest that analyzing the topological structure of attention matrices can serve as an efficient and robust indicator of factual reliability in LLMs.
Authors: Muhan Gao, Jash Shah, Weiqi Wang, Daniel Khashabi
Abstract: Scientific knowledge is growing rapidly, making it difficult to track progress and high-level conceptual links across broad disciplines. While tools like citation networks and search engines help retrieve related papers, they lack the abstraction needed to capture the needed to represent the density and structure of activity across subfields. We motivate SCIENCE HIERARCHOGRAPHY, the goal of organizing scientific literature into a high-quality hierarchical structure that spans multiple levels of abstraction -- from broad domains to specific studies. Such a representation can provide insights into which fields are well-explored and which are under-explored. To achieve this goal, we develop a hybrid approach that combines efficient embedding-based clustering with LLM-based prompting, striking a balance between scalability and semantic precision. Compared to LLM-heavy methods like iterative tree construction, our approach achieves superior quality-speed trade-offs. Our hierarchies capture different dimensions of research contributions, reflecting the interdisciplinary and multifaceted nature of modern science. We evaluate its utility by measuring how effectively an LLM-based agent can navigate the hierarchy to locate target papers. Results show that our method improves interpretability and offers an alternative pathway for exploring scientific literature beyond traditional search methods. Code, data and demo are available: https://github.com/JHU-CLSP/science-hierarchography
Authors: Vignesh Ethiraj, Ashwath David, Sidhanth Menon, Divya Vijay, Vidhyakshaya Kannan
Abstract: The specialized vocabulary and nuanced concepts of the telecommunications industry pose persistent challenges for standard Natural Language Processing (NLP) models. Generic embedding models often struggle to represent telecom-specific semantics, limiting their utility in retrieval and downstream tasks. We present T-VEC (Telecom Vectorization Model), a domain-adapted embedding model fine-tuned from the gte-Qwen2-1.5B-instruct backbone using a triplet loss objective. Fine-tuning was performed on T-Embed, a high-quality, large-scale dataset covering diverse telecom concepts, standards, and operational scenarios. Although T-Embed contains some proprietary material and cannot be fully released, we open source 75% of the dataset to support continued research in domain-specific representation learning. On a custom benchmark comprising 1500 query-passage pairs from IETF RFCs and vendor manuals, T-VEC surpasses MPNet, BGE, Jina and E5, demonstrating superior domain grounding and semantic precision in telecom-specific retrieval. Embedding visualizations further showcase tight clustering of telecom-relevant concepts. We release T-VEC and its tokenizer to support semantically faithful NLP applications within the telecom domain.
Authors: Atharva Kulkarni, Yuan Zhang, Joel Ruben Antony Moniz, Xiou Ge, Bo-Hsiang Tseng, Dhivya Piraviperumal, Swabha Swayamdipta, Hong Yu
Abstract: Hallucinations pose a significant obstacle to the reliability and widespread adoption of language models, yet their accurate measurement remains a persistent challenge. While many task- and domain-specific metrics have been proposed to assess faithfulness and factuality concerns, the robustness and generalization of these metrics are still untested. In this paper, we conduct a large-scale empirical evaluation of 6 diverse sets of hallucination detection metrics across 4 datasets, 37 language models from 5 families, and 5 decoding methods. Our extensive investigation reveals concerning gaps in current hallucination evaluation: metrics often fail to align with human judgments, take an overtly myopic view of the problem, and show inconsistent gains with parameter scaling. Encouragingly, LLM-based evaluation, particularly with GPT-4, yields the best overall results, and mode-seeking decoding methods seem to reduce hallucinations, especially in knowledge-grounded settings. These findings underscore the need for more robust metrics to understand and quantify hallucinations, and better strategies to mitigate them.
Authors: Cl\'ea Chataigner, Rebecca Ma, Prakhar Ganesh, Yuhao Chen, Afaf Ta\"ik, Elliot Creager, Golnoosh Farnadi
Abstract: Large language models (LLMs) are highly sensitive to subtle changes in prompt phrasing, posing challenges for reliable auditing. Prior methods often apply unconstrained prompt paraphrasing, which risk missing linguistic and demographic factors that shape authentic user interactions. We introduce AUGMENT (Automated User-Grounded Modeling and Evaluation of Natural Language Transformations), a framework for generating controlled paraphrases, grounded in user behaviors. AUGMENT leverages linguistically informed rules and enforces quality through checks on instruction adherence, semantic similarity, and realism, ensuring paraphrases are both reliable and meaningful for auditing. Through case studies on the BBQ and MMLU datasets, we show that controlled paraphrases uncover systematic weaknesses that remain obscured under unconstrained variation. These results highlight the value of the AUGMENT framework for reliable auditing.
Authors: Mehran Sarmadi, Morteza Alikhani, Erfan Zinvandi, Zahra Pourbahman
Abstract: Recent advancements in text embedding have significantly improved natural language understanding across many languages, yet Persian remains notably underrepresented in large-scale embedding research. In this paper, we present Hakim, a novel state-of-the-art Persian text embedding model that achieves a 8.5% performance improvement over existing approaches on the FaMTEB benchmark, outperforming all previously developed Persian language models. As part of this work, we introduce three new datasets - Corpesia, Pairsia-sup, and Pairsia-unsup - to support supervised and unsupervised training scenarios. Additionally, Hakim is designed for applications in chatbots and retrieval-augmented generation (RAG) systems, particularly addressing retrieval tasks that require incorporating message history within these systems. We also propose a new baseline model built on the BERT architecture. Our language model consistently achieves higher accuracy across various Persian NLP tasks, while the RetroMAE-based model proves particularly effective for textual information retrieval applications. Together, these contributions establish a new foundation for advancing Persian language understanding.
Authors: Jingyu Peng, Maolin Wang, Nan Wang, Jiatong Li, Yuchen Li, Yuyang Ye, Wanyu Wang, Pengyue Jia, Kai Zhang, Xiangyu Zhao
Abstract: Despite substantial advancements in aligning large language models (LLMs) with human values, current safety mechanisms remain susceptible to jailbreak attacks. We hypothesize that this vulnerability stems from distributional discrepancies between alignment-oriented prompts and malicious prompts. To investigate this, we introduce LogiBreak, a novel and universal black-box jailbreak method that leverages logical expression translation to circumvent LLM safety systems. By converting harmful natural language prompts into formal logical expressions, LogiBreak exploits the distributional gap between alignment data and logic-based inputs, preserving the underlying semantic intent and readability while evading safety constraints. We evaluate LogiBreak on a multilingual jailbreak dataset spanning three languages, demonstrating its effectiveness across various evaluation settings and linguistic contexts.
Authors: Zhepei Wei, Wenlin Yao, Yao Liu, Weizhi Zhang, Qin Lu, Liang Qiu, Changlong Yu, Puyang Xu, Chao Zhang, Bing Yin, Hyokun Yun, Lihong Li
Abstract: While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.
Authors: Shijia Zhou, Siyao Peng, Simon M. Luebke, J\"org Ha{\ss}ler, Mario Haim, Saif M. Mohammad, Barbara Plank
Abstract: Media framing refers to the emphasis on specific aspects of perceived reality to shape how an issue is defined and understood. Its primary purpose is to shape public perceptions often in alignment with the authors' opinions and stances. However, the interaction between stance and media frame remains largely unexplored. In this work, we apply an interdisciplinary approach to conceptualize and computationally explore this interaction with internet memes on climate change. We curate CLIMATEMEMES, the first dataset of climate-change memes annotated with both stance and media frames, inspired by research in communication science. CLIMATEMEMES includes 1,184 memes sourced from 47 subreddits, enabling analysis of frame prominence over time and communities, and sheds light on the framing preferences of different stance holders. We propose two meme understanding tasks: stance detection and media frame detection. We evaluate LLaVA-NeXT and Molmo in various setups, and report the corresponding results on their LLM backbone. Human captions consistently enhance performance. Synthetic captions and human-corrected OCR also help occasionally. Our findings highlight that VLMs perform well on stance, but struggle on frames, where LLMs outperform VLMs. Finally, we analyze VLMs' limitations in handling nuanced frames and stance expressions on climate change internet memes.
Authors: Ruihan Yang, Caiqi Zhang, Zhisong Zhang, Xinting Huang, Dong Yu, Nigel Collier, Deqing Yang
Abstract: Large Language Models (LLMs) are prone to hallucination, particularly in long-form generations. A promising direction to mitigate hallucination is to teach LLMs to express uncertainty explicitly when they lack sufficient knowledge. However, existing work lacks direct and fair evaluation of LLMs' ability to express uncertainty effectively in long-form generation. To address this gap, we first introduce UNCLE, a benchmark designed to evaluate uncertainty expression in both long- and short-form question answering (QA). UNCLE covers five domains and includes more than 1,000 entities, each with paired short- and long-form QA items. Our dataset is the first to directly link short- and long-form QA through aligned questions and gold-standard answers. Along with UNCLE, we propose a suite of new metrics to assess the models' capabilities to selectively express uncertainty. We then demonstrate that current models fail to convey uncertainty appropriately in long-form generation. We further explore both prompt-based and training-based methods to improve models' performance, with the training-based methods yielding greater gains. Further analysis of alignment gaps between short- and long-form uncertainty expression highlights promising directions for future research using UNCLE.
Authors: Peilin Wu, Mian Zhang, Xinlu Zhang, Xinya Du, Zhiyu Zoey Chen
Abstract: Agentic Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) by enabling dynamic, multi-step reasoning and information retrieval. However, these systems often exhibit sub-optimal search behaviors like over-search (retrieving redundant information) and under-search (failing to retrieve necessary information), which hinder efficiency and reliability. This work formally defines and quantifies these behaviors, revealing their prevalence across multiple QA datasets and agentic RAG systems (e.g., one model could have avoided searching in 27.7% of its search steps). Furthermore, we demonstrate a crucial link between these inefficiencies and the models' uncertainty regarding their own knowledge boundaries, where response accuracy correlates with model's uncertainty in its search decisions. To address this, we propose $\beta$-GRPO, a reinforcement learning-based training method that incorporates confidence threshold to reward high-certainty search decisions. Experiments on seven QA benchmarks show that $\beta$-GRPO enable a 3B model with better agentic RAG ability, outperforming other strong baselines with a 4% higher average exact match score.
Authors: Yige Yuan, Teng Xiao, Li Yunfan, Bingbing Xu, Shuchang Tao, Yunqi Qiu, Huawei Shen, Xueqi Cheng
Abstract: Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51%}$ on AdvBench and $\textbf{16.36%}$ on MATH. Our code is publicly available at https://github.com/yuanyige/sea
Authors: Mathew J. Koretsky, Maya Willey, Adi Asija, Owen Bianchi, Chelsea X. Alvarado, Tanay Nayak, Nicole Kuznetsov, Sungwon Kim, Mike A. Nalls, Daniel Khashabi, Faraz Faghri
Abstract: Biomedical researchers increasingly rely on large-scale structured databases for complex analytical tasks. However, current text-to-SQL systems often struggle to map qualitative scientific questions into executable SQL, particularly when implicit domain reasoning is required. We introduce BiomedSQL, the first benchmark explicitly designed to evaluate scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base. BiomedSQL comprises 68,000 question/SQL query/answer triples generated from templates and grounded in a harmonized BigQuery knowledge base that integrates gene-disease associations, causal inference from omics data, and drug approval records. Each question requires models to infer domain-specific criteria, such as genome-wide significance thresholds, effect directionality, or trial phase filtering, rather than rely on syntactic translation alone. We evaluate a range of open- and closed-source LLMs across prompting strategies and interaction paradigms. Our results reveal a substantial performance gap: GPT-o3-mini achieves 59.0% execution accuracy, while our custom multi-step agent, BMSQL, reaches 62.6%, both well below the expert baseline of 90.0%. BiomedSQL provides a new foundation for advancing text-to-SQL systems capable of supporting scientific discovery through robust reasoning over structured biomedical knowledge bases. Our dataset is publicly available at https://huggingface.co/datasets/NIH-CARD/BiomedSQL, and our code is open-source at https://github.com/NIH-CARD/biomedsql.
URLs: https://huggingface.co/datasets/NIH-CARD/BiomedSQL,, https://github.com/NIH-CARD/biomedsql.
Authors: Jiyoung Lee, Seungho Kim, Jieun Han, Jun-Min Lee, Kitaek Kim, Alice Oh, Edward Choi
Abstract: Large Language Models (LLMs) are predominantly evaluated on Standard American English (SAE), often overlooking the diversity of global English varieties. This narrow focus may raise fairness concerns as degraded performance on non-standard varieties can lead to unequal benefits for users worldwide. Therefore, it is critical to extensively evaluate the linguistic robustness of LLMs on multiple non-standard English varieties. We introduce Trans-EnV, a framework that automatically transforms SAE datasets into multiple English varieties to evaluate the linguistic robustness. Our framework combines (1) linguistics expert knowledge to curate variety-specific features and transformation guidelines from linguistic literature and corpora, and (2) LLM-based transformations to ensure both linguistic validity and scalability. Using Trans-EnV, we transform six benchmark datasets into 38 English varieties and evaluate seven state-of-the-art LLMs. Our results reveal significant performance disparities, with accuracy decreasing by up to 46.3% on non-standard varieties. These findings highlight the importance of comprehensive linguistic robustness evaluation across diverse English varieties. Each construction of Trans-EnV was validated through rigorous statistical testing and consultation with a researcher in the field of second language acquisition, ensuring its linguistic validity. Our code and datasets are publicly available at https://github.com/jiyounglee-0523/TransEnV and https://huggingface.co/collections/jiyounglee0523/transenv-681eadb3c0c8cf363b363fb1.
URLs: https://github.com/jiyounglee-0523/TransEnV, https://huggingface.co/collections/jiyounglee0523/transenv-681eadb3c0c8cf363b363fb1.
Authors: Zhanqiu Hu, Jian Meng, Yash Akhauri, Mohamed S. Abdelfattah, Jae-sun Seo, Zhiru Zhang, Udit Gupta
Abstract: Diffusion language models offer parallel token generation and inherent bidirectionality, promising more efficient and powerful sequence modeling compared to autoregressive approaches. However, state-of-the-art diffusion models (e.g., Dream 7B, LLaDA 8B) suffer from slow inference. While they match the quality of similarly sized autoregressive (AR) models (e.g., Qwen2.5 7B, Llama3 8B), their iterative denoising requires multiple full-sequence forward passes, resulting in high computational costs and latency, particularly for long input prompts and long-context scenarios. Furthermore, parallel token generation introduces token incoherence problems, and current sampling heuristics suffer from significant quality drops with decreasing denoising steps. We address these limitations with two training-free techniques. First, we propose FreeCache, a Key-Value (KV) approximation caching technique that reuses stable KV projections across denoising steps, effectively reducing the computational cost of DLM inference. Second, we introduce Guided Diffusion, a training-free method that uses a lightweight pretrained autoregressive model to supervise token unmasking, dramatically reducing the total number of denoising iterations without sacrificing quality. We conduct extensive evaluations on open-source reasoning benchmarks, and our combined methods deliver an average of 12.14x end-to-end speedup across various tasks with negligible accuracy degradation. For the first time, diffusion language models achieve a comparable and even faster latency as the widely adopted autoregressive models. Our work successfully paved the way for scaling up the diffusion language model to a broader scope of applications across different domains.
Authors: Md Kowsher, Nusrat Jahan Prottasha, Shiyun Xu, Shetu Mohanto, Ozlem Garibay, Niloofar Yousefi, Chen Chen
Abstract: Bidirectional language models have better context understanding and perform better than unidirectional models on natural language understanding tasks, yet the theoretical reasons behind this advantage remain unclear. In this work, we investigate this disparity through the lens of the Information Bottleneck (IB) principle, which formalizes a trade-off between compressing input information and preserving task-relevant content. We propose FlowNIB, a dynamic and scalable method for estimating mutual information during training that addresses key limitations of classical IB approaches, including computational intractability and fixed trade-off schedules. Theoretically, we show that bidirectional models retain more mutual information and exhibit higher effective dimensionality than unidirectional models. To support this, we present a generalized framework for measuring representational complexity and prove that bidirectional representations are strictly more informative under mild conditions. We further validate our findings through extensive experiments across multiple models and tasks using FlowNIB, revealing how information is encoded and compressed throughout training. Together, our work provides a principled explanation for the effectiveness of bidirectional architectures and introduces a practical tool for analyzing information flow in deep language models.
Authors: Soyoung Oh, Xinting Huang, Mathis Pink, Michael Hahn, Vera Demberg
Abstract: Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom's literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve an idiom's figurative interpretation, while suppressing its literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway prioritizes the figurative interpretation and a parallel direct route favors the literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.
Authors: Yujun Zhou, Jiayi Ye, Zipeng Ling, Yufei Han, Yue Huang, Haomin Zhuang, Zhenwen Liang, Kehan Guo, Taicheng Guo, Xiangqi Wang, Xiangliang Zhang
Abstract: Logical reasoning is a core capability for large language models (LLMs), yet existing benchmarks that rely solely on final-answer accuracy fail to capture the quality of the reasoning process. To address this, we introduce FineLogic, a fine-grained evaluation framework that assesses logical reasoning across three dimensions: overall accuracy, stepwise soundness, and representation-level probing. Leveraging this framework, we conduct a comprehensive study on how different supervision formats in fine-tuning shape reasoning abilities. We fine-tune LLMs on four supervision styles: one in natural language and three symbolic variants. We find a key trade-off: natural language supervision excels at generalization to out-of-distribution and long-chain problems, whereas symbolic supervision is superior at instilling structurally sound, atomic reasoning steps. Furthermore, our probing analysis indicates that fine-tuning primarily refines the model's step-by-step generation process, rather than improving its ability to converge on an answer early. Together, our framework and analysis provide a more rigorous lens for evaluating and improving logical reasoning in LLMs. The code is available at https://github.com/YujunZhou/FineLogic.
Authors: Nurul Aisyah, Muhammad Dehan Al Kautsar, Arif Hidayat, Raqib Chowdhury, Fajri Koto
Abstract: Despite rapid progress in vision-language and large language models (VLMs and LLMs), their effectiveness for AI-driven educational assessment in real-world, underrepresented classrooms remains largely unexplored. We evaluate state-of-the-art VLMs and LLMs on over 14K handwritten answers from grade-4 classrooms in Indonesia, covering Mathematics and English aligned with the local national curriculum. Unlike prior work on clean digital text, our dataset features naturally curly, diverse handwriting from real classrooms, posing realistic visual and linguistic challenges. Assessment tasks include grading and generating personalized Indonesian feedback guided by rubric-based evaluation. Results show that the VLM struggles with handwriting recognition, causing error propagation in LLM grading, yet LLM feedback remains pedagogically useful despite imperfect visual inputs, revealing limits in personalization and contextual relevance.
Authors: Yu Sun, Xingyu Qian, Weiwen Xu, Hao Zhang, Chenghao Xiao, Long Li, Deli Zhao, Wenbing Huang, Tingyang Xu, Qifeng Bai, Yu Rong
Abstract: Reasoning-based large language models have excelled in mathematics and programming, yet their potential in knowledge-intensive medical question answering remains underexplored and insufficiently validated in clinical contexts. To bridge this gap, we introduce ReasonMed, the largest medical reasoning dataset to date, comprising 370k high-quality examples distilled from 1.75 million initial reasoning paths generated by complementary LLMs and curated through a cost-efficient easy-medium-difficult (EMD) pipeline. ReasonMed is built through a multi-agent generation, verification, and refinement process, in which an Error Refiner improves reasoning paths by correcting error-prone steps identified by a verifier. Using ReasonMed, we investigate effective strategies for training medical reasoning models and find that integrating detailed CoT reasoning with concise answer summaries yields the most robust fine-tuning results. Models trained on ReasonMed set a new benchmark: ReasonMed-7B surpasses the prior best sub-10B models by 4.17% and even exceeds LLaMA3.1-70B on PubMedQA by 4.60%. When scaled to ReasonMed-14B, it remains highly competitive, underscoring consistent scaling potential. The codes and datasets are available at https://github.com/YuSun-Work/ReasonMed.
Authors: Tzu-Ling Lin, Wei-Chih Chen, Teng-Fang Hsiao, Hou-I Liu, Ya-Hsin Yeh, Yu Kai Chan, Wen-Sheng Lien, Po-Yen Kuo, Philip S. Yu, Hong-Han Shuai
Abstract: Peer review is essential for maintaining academic quality, but the increasing volume of submissions places a significant burden on reviewers. Large language models (LLMs) offer potential assistance in this process, yet their susceptibility to textual adversarial attacks raises reliability concerns. This paper investigates the robustness of LLMs used as automated reviewers in the presence of such attacks. We focus on three key questions: (1) The effectiveness of LLMs in generating reviews compared to human reviewers. (2) The impact of adversarial attacks on the reliability of LLM-generated reviews. (3) Challenges and potential mitigation strategies for LLM-based review. Our evaluation reveals significant vulnerabilities, as text manipulations can distort LLM assessments. We offer a comprehensive evaluation of LLM performance in automated peer reviewing and analyze its robustness against adversarial attacks. Our findings emphasize the importance of addressing adversarial risks to ensure AI strengthens, rather than compromises, the integrity of scholarly communication.
Authors: Joanito Agili Lopo, Muhammad Ravi Shulthan Habibi, Tack Hwa Wong, Muhammad Ilham Ghozali, Fajri Koto, Genta Indra Winata, Peerat Limkonchotiwat, Alham Fikri Aji, Samuel Cahyawijaya
Abstract: Large Language Models (LLMs) have demonstrated remarkable generalization capabilities across tasks and languages, revolutionizing natural language processing. This paper investigates the naturally emerging representation alignment in LLMs, particularly in the middle layers, and its implications for disentangling language-specific and language-agnostic information. We empirically confirm the existence of this alignment, analyze its behavior in comparison to explicitly designed alignment models, and demonstrate its potential for language-specific manipulation without semantic degradation. Building on these findings, we propose Inference-Time Language Control (ITLC), a novel method that leverages latent injection to enable precise cross-lingual language control and mitigate language confusion in LLMs. Our experiments highlight ITLC's strong cross-lingual control capabilities while preserving semantic integrity in target languages. Furthermore, we demonstrate its effectiveness in alleviating the cross-lingual language confusion problem, which persists even in current large-scale LLMs, leading to inconsistent language generation. This work advances our understanding of representation alignment in LLMs and introduces a practical solution for enhancing their monolingual and cross-lingual performance.
Authors: William Walden, Kathryn Ricci, Miriam Wanner, Zhengping Jiang, Chandler May, Rongkun Zhou, Benjamin Van Durme
Abstract: Wikipedia is a critical resource for modern NLP, serving as a rich repository of up-to-date and citation-backed information on a wide variety of subjects. The reliability of Wikipedia -- its groundedness in its cited sources -- is vital to this purpose. This work analyzes both how grounded Wikipedia is and how readily fine-grained grounding evidence can be retrieved. To this end, we introduce PeopleProfiles -- a large-scale, multi-level dataset of claim support annotations on biographical Wikipedia articles. We show that: (1) ~22% of claims in Wikipedia lead sections are unsupported by the article body; (2) ~30% of claims in the article body are unsupported by their publicly accessible sources; and (3) real-world Wikipedia citation practices often differ from documented standards. Finally, we show that complex evidence retrieval remains a challenge -- even for recent reasoning rerankers.
Authors: Xinyi Liu, Weiguang Wang, Hangfeng He
Abstract: With the growing adoption of Large Language Models (LLMs) for open-ended tasks, accurately assessing epistemic uncertainty, which reflects a model's lack of knowledge, has become crucial to ensuring reliable outcomes. However, quantifying epistemic uncertainty in such tasks is challenging due to the presence of aleatoric uncertainty, which arises from multiple valid answers. While bias can introduce noise into epistemic uncertainty estimation, it may also reduce noise from aleatoric uncertainty. To investigate this trade-off, we conduct experiments on Visual Question Answering (VQA) tasks and find that mitigating prompt-introduced bias improves uncertainty quantification in GPT-4o. Building on prior work showing that LLMs tend to copy input information when model confidence is low, we further analyze how these prompt biases affect measured epistemic and aleatoric uncertainty across varying bias-free confidence levels with GPT-4o and Qwen2-VL. We find that all considered biases have greater effects in both uncertainties when bias-free model confidence is lower. Moreover, lower bias-free model confidence is associated with greater bias-induced underestimation of epistemic uncertainty, resulting in overconfident estimates, whereas it has no significant effect on the direction of bias effect in aleatoric uncertainty estimation. These distinct effects deepen our understanding of bias mitigation for uncertainty quantification and potentially inform the development of more advanced techniques.
Authors: Nicholas Lourie, Michael Y. Hu, Kyunghyun Cho
Abstract: Downstream scaling laws aim to predict task performance at larger scales from the model's performance at smaller scales. Whether such prediction should be possible is unclear: some works discover clear linear scaling trends after simple transformations of the performance metric, whereas others point out fundamental challenges to downstream scaling laws, such as emergence and inverse scaling. In this work, we conduct a meta-analysis of existing data on downstream scaling laws, and we find that predictable scaling only occurs in a minority of cases: 39% of the time. Moreover, seemingly benign changes to the experimental setting can completely change the scaling behavior. Our analysis underscores the need to understand the conditions under which scaling laws succeed. To accurately model the relationship between pretraining loss and task performance, we must embrace the cases in which scaling behavior deviates from linear trends.
Authors: Mohammad Anas Azeez, Rafiq Ali, Ebad Shabbir, Zohaib Hasan Siddiqui, Gautam Siddharth Kashyap, Jiechao Gao, Usman Naseem
Abstract: Large Language Models (LLMs) hold significant promise for transforming digital health by enabling automated medical question answering. However, ensuring these models meet critical industry standards for factual accuracy, usefulness, and safety remains a challenge, especially for open-source solutions. We present a rigorous benchmarking framework using a dataset of over 1,000 health questions. We assess model performance across honesty, helpfulness, and harmlessness. Our results highlight trade-offs between factual reliability and safety among evaluated models -- Mistral-7B, BioMistral-7B-DARE, and AlpaCare-13B. AlpaCare-13B achieves the highest accuracy (91.7%) and harmlessness (0.92), while domain-specific tuning in BioMistral-7B-DARE boosts safety (0.90) despite its smaller scale. Few-shot prompting improves accuracy from 78% to 85%, and all models show reduced helpfulness on complex queries, highlighting ongoing challenges in clinical QA.
Authors: Zhiyuan Peng, Ting-ruen Wei, Tingyu Song, Yilun Zhao
Abstract: Large Language Models (LLMs) have recently been applied to reranking tasks in information retrieval, achieving strong performance. However, their high computational demands often hinder practical deployment. Existing studies evaluate the efficiency of LLM-based rerankers using proxy metrics such as latency, the number of forward passes, input tokens, and output tokens. However, these metrics depend on hardware and running-time choices (\eg parallel or not, batch size, etc), and often fail to account for model size, making it difficult to interpret and obscuring the evaluation of the efficiency-effectiveness tradeoff. To address this issue, we propose \ours\footnote{https://github.com/zhiyuanpeng/EER-FLOPs.} for LLM-based rerankers: RPP (ranking metrics per PetaFLOP), measuring how much ranking quality (e.g., NDCG or MRR) a method achieves per PetaFLOP, and QPP (queries per PetaFLOP), measuring how many queries can be processed per PetaFLOP. Accompanied by the new metrics, an interpretable FLOPs estimator is developed to estimate the FLOPs of an LLM-based reranker even without running any experiments. Based on the proposed metrics, we conduct comprehensive experiments to evaluate a wide range of LLM-based rerankers with different architectures, studying the efficiency-effectiveness trade-off and bringing this issue to the attention of the research community.
Authors: Sanhanat Sivapiromrat, Caiqi Zhang, Marco Basaldella, Nigel Collier
Abstract: Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.
Authors: Jiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau, Weiyan Shi
Abstract: LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.
Authors: Michael Carl, Takanori Mizowaki, Aishvarya Ray, Masaru Yamada, Devi Sri Bandaru, Xinyue Ren
Abstract: The paper introduces a novel behavioural translation style space (BTSS) that describes possible behavioural translation patterns. The suggested BTSS is organized as a hierarchical structure that entails various embedded processing layers. We posit that observable translation behaviour - i.e. eye and finger movements - is fundamental when executing the physical act of translation but it is caused and shaped by higher-order cognitive processes and affective translation states. We analyse records of keystrokes and gaze data as indicators of the hidden mental processing structure and organize the behavioural patterns as a multi-layered embedded BTSS. We develop a perspective in which the BTSS serves as the basis for a computational translation agent to simulate the temporal dynamics of affect, behavioural routines and cognition during human translation production.
Authors: Abraham Toluwase Owodunni, Orevaoghene Ahia, Sachin Kumar
Abstract: Language models (LMs) are challenging to adapt to new data distributions by simple finetuning. This is due to the rigidity of their subword tokenizers, which typically remain unchanged during adaptation. This inflexibility often leads to inefficient tokenization, causing overfragmentation of out-of-distribution domains, unseen languages, or scripts. In this work, we develop byte-level LMs with learnable tokenizers to make tokenization adaptive. Our models include a submodule that learns to predict boundaries between the input byte sequence, encoding it into variable-length segments. Existing tokenizer-free methods train this boundary predictor using an auxiliary loss that enforces a fixed compression rate across the training corpus, introducing a new kind of rigidity. We propose FLEXITOKENS, a simplified training objective that enables significantly greater flexibility during adaptation. Evaluating across multiple multilingual benchmarks, morphologically diverse tasks, and domains, we demonstrate that FLEXITOKENS consistently reduces token over-fragmentation and achieves up to 10% improvements on downstream task performance compared to subword and other gradient-based tokenizers. Code and data for our experiments will be released at https://github.com/owos/flexitokens
Authors: Karen Zhou, John Giorgi, Pranav Mani, Peng Xu, Davis Liang, Chenhao Tan
Abstract: AI-generated clinical notes are increasingly used in healthcare, but evaluating their quality remains a challenge due to high subjectivity and limited scalability of expert review. Existing automated metrics often fail to align with real-world physician preferences. To address this, we propose a pipeline that systematically distills real user feedback into structured checklists for note evaluation. These checklists are designed to be interpretable, grounded in human feedback, and enforceable by LLM-based evaluators. Using deidentified data from over 21,000 clinical encounters (prepared in accordance with the HIPAA safe harbor standard) from a deployed AI medical scribe system, we show that our feedback-derived checklist outperforms a baseline approach in our offline evaluations in coverage, diversity, and predictive power for human ratings. Extensive experiments confirm the checklist's robustness to quality-degrading perturbations, significant alignment with clinician preferences, and practical value as an evaluation methodology. In offline research settings, our checklist offers a practical tool for flagging notes that may fall short of our defined quality standards.
Authors: Chaymaa Abbas, Mariette Awad, Razane Tajeddine
Abstract: Style-conditioned data poisoning is identified as a covert vector for amplifying sociolinguistic bias in large language models. Using small poisoned budgets that pair dialectal prompts -- principally African American Vernacular English (AAVE) and a Southern dialect -- with toxic or stereotyped completions during instruction tuning, this work probes whether linguistic style can act as a latent trigger for harmful behavior. Across multiple model families and scales, poisoned exposure elevates toxicity and stereotype expression for dialectal inputs -- most consistently for AAVE -- while Standard American English remains comparatively lower yet not immune. A multi-metric audit combining classifier-based toxicity with an LLM-as-a-judge reveals stereotype-laden content even when lexical toxicity appears muted, indicating that conventional detectors under-estimate sociolinguistic harms. Additionally, poisoned models exhibit emergent jailbreaking despite the absence of explicit slurs in the poison, suggesting weakened alignment rather than memorization. These findings underscore the need for dialect-aware evaluation, content-level stereotype auditing, and training protocols that explicitly decouple style from toxicity to prevent bias amplification through seemingly minor, style-based contamination.
Authors: Tianxiang Chen, Zhentao Tan, Xiaofan Bo, Yue Wu, Tao Gong, Qi Chu, Jieping Ye, Nenghai Yu
Abstract: Effectively handling long contexts is challenging for Large Language Models (LLMs) due to the rarity of long texts, high computational demands, and substantial forgetting of short-context abilities. Recent approaches have attempted to construct long contexts for instruction tuning, but these methods often require LLMs or human interventions, which are both costly and limited in length and diversity. Also, the drop in short-context performances of present long-context LLMs remains significant. In this paper, we introduce Flora, an effortless (human/LLM-free) long-context construction strategy. Flora can markedly enhance the long-context performance of LLMs by arbitrarily assembling short instructions based on categories and instructing LLMs to generate responses based on long-context meta-instructions. This enables Flora to produce contexts of arbitrary length and scale with rich diversity, while only slightly compromising short-context performance. Experiments on Llama3-8B-Instruct and QwQ-32B show that LLMs enhanced by Flora excel in three long-context benchmarks while maintaining strong performances in short-context tasks. Our data-construction code is available at \href{https://github.com/txchen-USTC/Flora}{https://github.com/txchen-USTC/Flora}.
URLs: https://github.com/txchen-USTC/Flora, https://github.com/txchen-USTC/Flora
Authors: Jungyeon Lee, Kangmin Lee, Taeuk Kim
Abstract: Knowledge conflict often arises in retrieval-augmented generation (RAG) systems, where retrieved documents may be inconsistent with one another or contradict the model's parametric knowledge. Existing benchmarks for investigating the phenomenon have notable limitations, including a narrow focus on the question answering setup, heavy reliance on entity substitution techniques, and a restricted range of conflict types. To address these issues, we propose a knowledge graph (KG)-based framework that generates varied and subtle conflicts between two similar yet distinct contexts, while ensuring interpretability through the explicit relational structure of KGs. Experimental results on our benchmark, MAGIC, provide intriguing insights into the inner workings of LLMs regarding knowledge conflict: both open-source and proprietary models struggle with conflict detection -- especially when multi-hop reasoning is required -- and often fail to pinpoint the exact source of contradictions. Finally, we present in-depth analyses that serve as a foundation for improving LLMs in integrating diverse, sometimes even conflicting, information.
Authors: Keer Lu, Zheng Liang, Youquan Li, Jiejun Tan, Da Pan, Shusen Zhang, Guosheng Dong, Zhonghai Wu, Huang Leng, Bin Cui, Wentao Zhang
Abstract: In medical scenarios, effectively retrieving external knowledge and leveraging it for rigorous logical reasoning is of significant importance. Despite their potential, existing work has predominantly focused on enhancing either retrieval or reasoning capabilities of the models in isolation, with little attention given to their joint optimization, which leads to limited coordination between the two processes. Additionally, current methods rely heavily on supervised fine-tuning (SFT), which can cause models to memorize existing problem-solving pathways, thereby restricting their generalization ability when confronted with novel problem contexts. Furthermore, while some studies have explored to improve retrieval-augmented reasoning in general domains via reinforcement learning, their reward function designs do not adequately capture the specific demands of the medical domain. To address these challenges, we introduce **Med-R$^3$**, a **Med**ical **R**etrieval-augmented **R**easoning framework driven by progressive **R**einforcement learning. In this framework, we first develop the model's ability to perform logical reasoning over medical problems. Subsequently, on the basis of this foundation, we adaptively optimize the retrieval capability to better align with the characteristics of knowledge corpus and external information utilization throughout the reasoning process. Finally, we conduct joint optimization of the model's retrieval and reasoning coordination. Extensive experiments indicate that **Med-R$^3$** could achieve state-of-the-art performances, with LLaMA3.1-8B-Instruct + Med-R$^3$ surpassing closed-sourced GPT-4o-mini by 3.93\% at a comparable parameter scale, while Qwen2.5-14B augmented with Med-R$^3$ shows a more substantial gain of 13.53\%.
Authors: Yi Jiang, Sendong Zhao, Jianbo Li, Haochun Wang, Lizhe Zhang, Yan Liu, Bing Qin
Abstract: Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs), especially for knowledge-intensive tasks. Despite its advantages, current RAG methods often struggle to fully exploit knowledge during generation. In particular, the synergy between the model's internal parametric knowledge and external retrieved knowledge remains limited. Retrieved contents may sometimes mislead generation, while certain generated content can guide the model toward more accurate outputs. In this work, we propose Collaborative Chain-of-Agents, a framework designed to enhance explicitly synergy over both parametric and retrieved knowledge. Specifically, we first introduce CoCoA-zero, a multi-agent RAG framework that first performs conditional knowledge induction and then reasons answers. Building on this, we develop CoCoA, a long-chain training strategy that synthesizes extended multi-agent reasoning trajectories from CoCoA-zero to fine-tune the LLM. This strategy enhances the model's capability to explicitly integrate and jointly leverage parametric and retrieved knowledge. Experimental results demonstrate the superiority of CoCoA in open-domain QA and multi-hop QA.
Authors: Stefanie Urchs, Veronika Thurner, Matthias A{\ss}enmacher, Christian Heumann, Stephanie Thiemichen
Abstract: Language corpora are the foundation of most natural language processing research, yet they often reproduce structural inequalities. One such inequality is gender discrimination in how actors are represented, which can distort analyses and perpetuate discriminatory outcomes. This paper introduces a user-centric, actor-level pipeline for detecting and mitigating gender discrimination in large-scale text corpora. By combining discourse-aware analysis with metrics for sentiment, syntactic agency, and quotation styles, our method enables both fine-grained auditing and exclusion-based balancing. Applied to the taz2024full corpus of German newspaper articles (1980-2024), the pipeline yields a more gender-balanced dataset while preserving core dynamics of the source material. Our findings show that structural asymmetries can be reduced through systematic filtering, though subtler biases in sentiment and framing remain. We release the tools and reports to support further research in discourse-based fairness auditing and equitable corpus construction.
Authors: Josh Barua, Seun Eisape, Kayo Yin, Alane Suhr
Abstract: While large reasoning models have shown remarkable ability to generate long chains-of-thought (CoTs) in English, we still lack understanding of how these long-form reasoning abilities transfer to the vast majority of the world's languages. In this work, we systematically investigate four key stages of model development--scaling, pretraining, post-training, and inference--to understand how long CoT capabilities extend beyond English. We compare two reasoning settings across nine non-English target languages: En-CoT, where models process target-language inputs, but reason in English; and Target-CoT, where models both process inputs and generate long CoTs in the target language. We find that scaling reasoning model size improves multilingual task performance in En-CoT, but Target-CoT performance lags behind. This gap widens for tasks requiring long, multi-step CoTs such as mathematical reasoning. Shifting to pretraining, we find that adding a specialized reasoning stage enhances En-CoT performance but degrades Target-CoT, whereas broad multilingual pretraining improves both modes simultaneously. Given the scarcity of high-quality reasoning traces in languages other than English, we explore synthetic data curation approaches for post-training. We demonstrate that fine-tuning on reasoning traces automatically translated from gold English traces outperforms fine-tuning on target-language traces distilled from large reasoning models. Finally, we report disparities in inference efficiency between languages and uncover language-specific failure modes in CoTs. We release models, datasets, and code to foster further research.
Authors: Hyunjun Kim, Junwoo Ha, Sangyoon Yu, Haon Park
Abstract: LLM-as-a-Judge (LLMaaJ) enables scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover the hidden objective of a conversation and know when that inference is reliable? Large language models degrade with irrelevant or lengthy context, and multi-turn jailbreaks can scatter goals across turns. We present ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must output a one-sentence base objective and a self-reported confidence. Accuracy is scored by semantic similarity to gold objectives, then thresholded once on 300 calibration items ($\tau^\star = 0.66$; $F_1@\tau^\star = 0.891$). Metacognition is assessed with expected calibration error, Brier score, Wrong@High-Confidence (0.80 / 0.90 / 0.95), and risk--coverage curves. Across six models (gpt-4.1, claude-sonnet-4, Qwen3-235B-A22B-FP8, kimi-k2, deepseek-v3.1, gemini-2.5-flash) evaluated on SafeMTData\_Attack600, SafeMTData\_1K, and MHJ, kimi-k2 achieves the highest objective-extraction accuracy (0.612; 95\% CI [0.594, 0.630]), while claude-sonnet-4 (0.603) and deepseek-v3.1 (0.599) are statistically tied. claude-sonnet-4 offers the best selective risk and calibration (AURC 0.242; ECE 0.206; Brier 0.254). Performance varies sharply across datasets (16--82\% accuracy), showing that automated obfuscation imposes challenges beyond model choice. High-confidence errors remain: Wrong@0.90 ranges from 14.9\% (claude-sonnet-4) to 47.7\% (Qwen3-235B-A22B-FP8). ObjexMT therefore supplies an actionable test for LLM judges: when objectives are implicit, judges often misinfer them; exposing objectives or gating decisions by confidence is advisable. All experimental data are in the Supplementary Material and at https://github.com/hyunjun1121/ObjexMT_dataset.
Authors: Duy Le, Kent Ziti, Evan Girard-Sun, Bakr Bouhaya, Sean O'Brien, Vasu Sharma, Kevin Zhu
Abstract: Language models are increasingly tested on multilingual creativity, demanding culturally grounded, abstract generations. Standard prompting methods often produce repetitive or shallow outputs. We introduce Adaptive Originality Filtering (AOF), a prompting strategy that enforces novelty and cultural fidelity via semantic rejection. To assess quality, we propose RiddleScore, a metric combining novelty, diversity, fluency, and answer alignment. AOF improves Distinct-2 (0.915 in Japanese), reduces Self-BLEU (0.177), and raises RiddleScore (up to +57.1% in Arabic). Human evaluations confirm fluency, creativity, and cultural fit gains. However, improvements vary: Arabic shows greater RiddleScore gains than Distinct-2; Japanese sees similar changes. Though focused on riddles, our method may apply to broader creative tasks. Overall, semantic filtering with composite evaluation offers a lightweight path to culturally rich generation without fine-tuning.
Authors: Wenhao Li, Yuxin Zhang, Gen Luo, Haiyuan Wan, Ziyang Gong, Fei Chao, Rongrong Ji
Abstract: Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$\mu$s on a single A100 GPU, with end-to-end throughput up to 3$\times$ higher than vanilla decoding.
Authors: Zinan Tang, Xin Gao, Qizhi Pei, Zhuoshi Pan, Mengzhang Cai, Jiang Wu, Conghui He, Lijun Wu
Abstract: Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to improve data quality, existing approaches often face limitations in static dataset curation that fail to adapt to evolving model capabilities. In this paper, we introduce Middo, a self-evolving Model-informed dynamic data optimization framework that uses model-aware data selection and context-preserving data refinement. Unlike conventional one-off filtering/synthesis methods, our framework establishes a closed-loop optimization system: (1) A self-referential diagnostic module proactively identifies suboptimal samples through tri-axial model signals - loss patterns (complexity), embedding cluster dynamics (diversity), and self-alignment scores (quality); (2) An adaptive optimization engine then transforms suboptimal samples into pedagogically valuable training points while preserving semantic integrity; (3) This optimization process continuously evolves with model capability through dynamic learning principles. Experiments on multiple benchmarks demonstrate that our Middo consistently enhances the quality of seed data and boosts LLM's performance with improving accuracy by 7.15% on average while maintaining the original dataset scale. This work establishes a new paradigm for sustainable LLM training through dynamic human-AI co-evolution of data and models. Our datasets, models, and code are publicly available at https://github.com/Word2VecT/Middo.
Authors: Eunjung Cho, Alexander Hoyle, Yoan Hermstr\"uwer
Abstract: Large Language Models (LLMs) are increasingly used to generate user-tailored summaries, adapting outputs to specific stakeholders. In legal contexts, this raises important questions about motivated reasoning -- how models strategically frame information to align with a stakeholder's position within the legal system. Building on theories of legal realism and recent trends in legal practice, we investigate how LLMs respond to prompts conditioned on different legal roles (e.g., judges, prosecutors, attorneys) when summarizing judicial decisions. We introduce an evaluation framework grounded in legal fact and reasoning inclusion, also considering favorability towards stakeholders. Our results show that even when prompts include balancing instructions, models exhibit selective inclusion patterns that reflect role-consistent perspectives. These findings raise broader concerns about how similar alignment may emerge as LLMs begin to infer user roles from prior interactions or context, even without explicit role instructions. Our results underscore the need for role-aware evaluation of LLM summarization behavior in high-stakes legal settings.
Authors: Chang Su, Dengliang Shi, Siyuan Huang, Jintao Du, Changhua Meng, Yu Cheng, Weiqiang Wang, Zhouhan Lin
Abstract: Large language models (LLMs) have increasingly been explored as powerful text embedders. Existing LLM-based text embedding approaches often leverage the embedding of the final token, typically a reserved special token such as [EOS]. However, these tokens have not been intentionally trained to capture the semantics of the whole context, limiting their capacity as text embeddings, especially for retrieval and re-ranking tasks. We propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding. This stage employs bidirectional generative reconstruction tasks, namely EBQ2D (Embedding-Based Query-to-Document) and EBD2Q (Embedding-Based Document-to-Query), which interleave to anchor the [EOS] embedding and reconstruct either side of Query-Document pairs. Experimental results demonstrate that our additional training stage significantly improves LLM performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.
Authors: Hyunjun Kim, Junwoo Ha, Sangyoon Yu, Haon Park
Abstract: Multi-turn-to-single-turn (M2S) compresses iterative red-teaming into one structured prompt, but prior work relied on a handful of manually written templates. We present X-Teaming Evolutionary M2S, an automated framework that discovers and optimizes M2S templates through language-model-guided evolution. The system pairs smart sampling from 12 sources with an LLM-as-judge inspired by StrongREJECT and records fully auditable logs. Maintaining selection pressure by setting the success threshold to $\theta = 0.70$, we obtain five evolutionary generations, two new template families, and 44.8% overall success (103/230) on GPT-4.1. A balanced cross-model panel of 2,500 trials (judge fixed) shows that structural gains transfer but vary by target; two models score zero at the same threshold. We also find a positive coupling between prompt length and score, motivating length-aware judging. Our results demonstrate that structure-level search is a reproducible route to stronger single-turn probes and underscore the importance of threshold calibration and cross-model evaluation. Code, configurations, and artifacts are available at https://github.com/hyunjun1121/M2S-x-teaming.
Authors: Kaiyan Zhang, Yuxin Zuo, Bingxiang He, Youbang Sun, Runze Liu, Che Jiang, Yuchen Fan, Kai Tian, Guoli Jia, Pengfei Li, Yu Fu, Xingtai Lv, Yuchen Zhang, Sihang Zeng, Shang Qu, Haozhan Li, Shijie Wang, Yuru Wang, Xinwei Long, Fangfu Liu, Xiang Xu, Jiaze Ma, Xuekai Zhu, Ermo Hua, Yihao Liu, Zonglin Li, Huayu Chen, Xiaoye Qu, Yafu Li, Weize Chen, Zhenzhao Yuan, Junqi Gao, Dong Li, Zhiyuan Ma, Ganqu Cui, Zhiyuan Liu, Biqing Qi, Ning Ding, Bowen Zhou
Abstract: In this paper, we survey recent advances in Reinforcement Learning (RL) for reasoning with Large Language Models (LLMs). RL has achieved remarkable success in advancing the frontier of LLM capabilities, particularly in addressing complex logical tasks such as mathematics and coding. As a result, RL has emerged as a foundational methodology for transforming LLMs into LRMs. With the rapid progress of the field, further scaling of RL for LRMs now faces foundational challenges not only in computational resources but also in algorithm design, training data, and infrastructure. To this end, it is timely to revisit the development of this domain, reassess its trajectory, and explore strategies to enhance the scalability of RL toward Artificial SuperIntelligence (ASI). In particular, we examine research applying RL to LLMs and LRMs for reasoning abilities, especially since the release of DeepSeek-R1, including foundational components, core problems, training resources, and downstream applications, to identify future opportunities and directions for this rapidly evolving area. We hope this review will promote future research on RL for broader reasoning models. Github: https://github.com/TsinghuaC3I/Awesome-RL-for-LRMs
Authors: Wensheng Lu, Keyu Chen, Ruizhi Qiao, Xing Sun
Abstract: Retrieval-Augmented Generation (RAG) enhances the response capabilities of language models by integrating external knowledge sources. However, document chunking as an important part of RAG system often lacks effective evaluation tools. This paper first analyzes why existing RAG evaluation benchmarks are inadequate for assessing document chunking quality, specifically due to evidence sparsity. Based on this conclusion, we propose HiCBench, which includes manually annotated multi-level document chunking points, synthesized evidence-dense quetion answer(QA) pairs, and their corresponding evidence sources. Additionally, we introduce the HiChunk framework, a multi-level document structuring framework based on fine-tuned LLMs, combined with the Auto-Merge retrieval algorithm to improve retrieval quality. Experiments demonstrate that HiCBench effectively evaluates the impact of different chunking methods across the entire RAG pipeline. Moreover, HiChunk achieves better chunking quality within reasonable time consumption, thereby enhancing the overall performance of RAG systems.
Authors: Yuanjie Lyu, Chengyu Wang, Jun Huang, Tong Xu
Abstract: Large Language Model agents excel at solving complex tasks through iterative reasoning and tool use, but typically depend on ultra-large, costly backbones. Existing distillation approaches train smaller students to imitate full teacher trajectories, yet reasoning and knowledge gaps between the teacher and student can cause compounding errors. We propose SCoRe, a student-centered framework in which the student generates training trajectories and the teacher corrects only the earliest error, producing training data matched to the student's ability and exposing specific weaknesses. The student is first fine-tuned on corrected trajectories. Subsequently, short-horizon reinforcement learning starts from the verified prefix preceding the earliest error, with target rewards assigned at that step. This design encourages autonomous problem-solving beyond imitation and enhances training stability. On 12 challenging benchmarks, a 7B-parameter student distilled with SCoRe matches the agentic performance of a 72B-parameter teacher.
Authors: Xiaowei Zhu, Yubing Ren, Fang Fang, Qingfeng Tan, Shi Wang, Yanan Cao
Abstract: The rapid advancement of large language models (LLMs) has blurred the line between AI-generated and human-written text. This progress brings societal risks such as misinformation, authorship ambiguity, and intellectual property concerns, highlighting the urgent need for reliable AI-generated text detection methods. However, recent advances in generative language modeling have resulted in significant overlap between the feature distributions of human-written and AI-generated text, blurring classification boundaries and making accurate detection increasingly challenging. To address the above challenges, we propose a DNA-inspired perspective, leveraging a repair-based process to directly and interpretably capture the intrinsic differences between human-written and AI-generated text. Building on this perspective, we introduce DNA-DetectLLM, a zero-shot detection method for distinguishing AI-generated and human-written text. The method constructs an ideal AI-generated sequence for each input, iteratively repairs non-optimal tokens, and quantifies the cumulative repair effort as an interpretable detection signal. Empirical evaluations demonstrate that our method achieves state-of-the-art detection performance and exhibits strong robustness against various adversarial attacks and input lengths. Specifically, DNA-DetectLLM achieves relative improvements of 5.55% in AUROC and 2.08% in F1 score across multiple public benchmark datasets. Code and data are available at https://github.com/Xiaoweizhu57/DNA-DetectLLM.
Authors: Dongxu Lu, Johan Jeuring, Albert Gatt
Abstract: Evaluating large language models (LLMs) in long-form, knowledge-grounded role-play dialogues remains challenging. This study compares LLM-generated and human-authored responses in multi-turn professional training simulations through human evaluation ($N=38$) and automated LLM-as-a-judge assessment. Human evaluation revealed significant degradation in LLM-generated response quality across turns, particularly in naturalness, context maintenance and overall quality, while human-authored responses progressively improved. In line with this finding, participants also indicated a consistent preference for human-authored dialogue. These human judgements were validated by our automated LLM-as-a-judge evaluation, where Gemini 2.0 Flash achieved strong alignment with human evaluators on both zero-shot pairwise preference and stochastic 6-shot construct ratings, confirming the widening quality gap between LLM and human responses over time. Our work contributes a multi-turn benchmark exposing LLM degradation in knowledge-grounded role-play dialogues and provides a validated hybrid evaluation framework to guide the reliable integration of LLMs in training simulations.
Authors: Pei-Shuo Wang, Jian-Jia Chen, Chun-Che Yang, Chi-Chih Chang, Ning-Chi Huang, Mohamed S. Abdelfattah, Kai-Chiang Wu
Abstract: The immense model sizes of large language models (LLMs) challenge deployment on memory-limited consumer GPUs. Although model compression and parameter offloading are common strategies to address memory limitations, compression can degrade quality, and offloading maintains quality but suffers from slow inference. Speculative decoding presents a promising avenue to accelerate parameter offloading, utilizing a fast draft model to propose multiple draft tokens, which are then verified by the target LLM in parallel with a single forward pass. This method reduces the time-consuming data transfers in forward passes that involve offloaded weight transfers. Existing methods often rely on pretrained weights of the same family, but require additional training to align with custom-trained models. Moreover, approaches that involve draft model training usually yield only modest speedups. This limitation arises from insufficient alignment with the target model, preventing higher token acceptance lengths. To address these challenges and achieve greater speedups, we propose SubSpec, a plug-and-play method to accelerate parameter offloading that is lossless and training-free. SubSpec constructs a highly aligned draft model by generating low-bit quantized substitute layers from offloaded target LLM portions. Additionally, our method shares the remaining GPU-resident layers and the KV-Cache, further reducing memory overhead and enhance alignment. SubSpec achieves a high average acceptance length, delivering 9.1x speedup for Qwen2.5 7B on MT-Bench (8GB VRAM limit) and an average of 12.5x speedup for Qwen2.5 32B on popular generation benchmarks (24GB VRAM limit).
Authors: Yijin Guo, Zicheng Zhang, Ye Shen, Farong Wen, Junying Wang, Qi Jia, Guangtao Zhai
Abstract: Existing evaluations of foundation models, including recent human-centric approaches, fail to capture what truly matters: user's experience during interaction. Current methods treat evaluation as a matter of output correctness alone, overlooking that user satisfaction emerges from the interplay between response quality and interaction, which limits their ability to account for the mechanisms underlying user experience. To address this gap, we introduce QoNext, the first framework that adapts Quality of Experience (QoE) principles from networking and multimedia to the assessment of foundation models. QoNext identifies experiential factors that shape user experience and incorporates them into controlled experiments, where human ratings are collected under varied configurations. From these studies we construct a QoE-oriented database and train predictive models that estimate perceived user experience from measurable system parameters. Our results demonstrate that QoNext not only enables proactive and fine-grained evaluation but also provides actionable guidance for productized services of optimizing foundation models in practice.
Authors: Wenjun Wang, Shuo Cai, Congkai Xie, Mingfa Feng, Yiming Zhang, Zhen Li, Kejing Yang, Ming Li, Jiannong Cao, Yuan Xie, Hongxia Yang
Abstract: The immense computational cost of training Large Language Models (LLMs) presents a major barrier to innovation. While FP8 training offers a promising solution with significant theoretical efficiency gains, its widespread adoption has been hindered by the lack of a comprehensive, open-source training recipe. To bridge this gap, we introduce an end-to-end FP8 training recipe that seamlessly integrates continual pre-training and supervised fine-tuning. Our methodology employs a fine-grained, hybrid-granularity quantization strategy to maintain numerical fidelity while maximizing computational efficiency. Through extensive experiments, including the continue pre-training of models on a 160B-token corpus, we demonstrate that our recipe is not only remarkably stable but also essentially lossless, achieving performance on par with the BF16 baseline across a suite of reasoning benchmarks. Crucially, this is achieved with substantial efficiency improvements, including up to a 22% reduction in training time, a 14% decrease in peak memory usage, and a 19% increase in throughput. Our results establish FP8 as a practical and robust alternative to BF16, and we will release the accompanying code to further democratize large-scale model training.
Authors: Huacan Chai, Zijie Cao, Maolin Ran, Yingxuan Yang, Jianghao Lin, Xin Peng, Hairui Wang, Renjie Ding, Ziyu Wan, Muning Wen, Weiwen Liu, Weinan Zhang, Fei Huang, Ying Wen
Abstract: Large language models (LLMs) have achieved impressive success in single-turn function calling, yet real-world applications such as travel planning or multi-stage data analysis typically unfold across multi-turn conversations. In these settings, LLMs must not only issue accurate function calls at each step but also maintain progress awareness, the ability to summarize past interactions and plan future actions to ensure coherent, long-horizon task execution. Existing approaches, however, either reduce multi-turn training to isolated single-turn samples, which neglects task-level planning, or employ end-to-end reinforcement learning (RL) that struggles with redundancy and lacks explicit integration of progress awareness. To overcome these limitations, we introduce PARL-MT, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. PARL-MT combines (i) a Progress Awareness Generation (PAG) pipeline, which automatically constructs datasets coupling conversation summaries with future task planning, and (ii) a Progress Awareness-Guided Reinforcement Learning (PAG-RL) algorithm, which integrates progress awareness into RL training to reduce contextual redundancy and improve alignment between local actions and global task completion. Empirical results on two public benchmarks demonstrate that PARL-MT significantly outperforms existing methods, highlighting the effectiveness of progress awareness in enabling robust and efficient multi-turn function calling.
Authors: Jinyeop Song, Song Wang, Julian Shun, Yada Zhu
Abstract: Knowledge-graph retrieval-augmented generation (KG-RAG) couples large language models (LLMs) with structured, verifiable knowledge graphs (KGs) to reduce hallucinations and expose reasoning traces. However, many KG-RAG systems compose multiple LLM modules (e.g planning, reasoning, and responding), inflating inference cost and binding behavior to a specific target KG. To address this, we introduce KG-R1, an agentic KG retrieval-augmented generation (KG-RAG) framework through reinforcement learning (RL). KG-R1 utilizes a single agent that interacts with KGs as its environment, learning to retrieve at each step and incorporating the retrieved information into its reasoning and generation. The process is optimized through end-to-end RL. In controlled experiments across Knowledge-Graph Question Answering (KGQA) benchmarks, our method demonstrates both efficiency and transferability: Using Qwen-2.5-3B, KG-R1 improves answer accuracy with fewer generation tokens than prior multi-module workflow methods that use larger foundation or fine-tuned models. Furthermore, KG-R1 enables plug and play: after training, it maintains strong accuracy on new KGs without modification. These properties make KG-R1 a promising KG-RAG framework for real-world deployment. Our code is publicly available at https://github.com/Jinyeop3110/KG-R1.
Authors: Hsuan Su, Ting-Yao Hu, Hema Swetha Koppula, Kundan Krishna, Hadi Pouransari, Cheng-Yu Hsieh, Cem Koc, Joseph Yitan Cheng, Oncel Tuzel, Raviteja Vemulapalli
Abstract: Large language models (LLMs) often generate hallucinations -- unsupported content that undermines reliability. While most prior works frame hallucination detection as a binary task, many real-world applications require identifying hallucinated spans, which is a multi-step decision making process. This naturally raises the question of whether explicit reasoning can help the complex task of detecting hallucination spans. To answer this question, we first evaluate pretrained models with and without Chain-of-Thought (CoT) reasoning, and show that CoT reasoning has the potential to generate at least one correct answer when sampled multiple times. Motivated by this, we propose RL4HS, a reinforcement learning framework that incentivizes reasoning with a span-level reward function. RL4HS builds on Group Relative Policy Optimization and introduces Class-Aware Policy Optimization to mitigate reward imbalance issue. Experiments on the RAGTruth benchmark (summarization, question answering, data-to-text) show that RL4HS surpasses pretrained reasoning models and supervised fine-tuning, demonstrating the necessity of reinforcement learning with span-level rewards for detecting hallucination spans.
Authors: Xiaoyang Yuan, Yujuan Ding, Yi Bin, Wenqi Shao, Jinyu Cai, Jingkuan Song, Yang Yang, Heng Tao Shen
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is a promising paradigm for enhancing the reasoning ability in Large Language Models (LLMs). However, prevailing methods primarily rely on self-exploration or a single off-policy teacher to elicit long chain-of-thought (LongCoT) reasoning, which may introduce intrinsic model biases and restrict exploration, ultimately limiting reasoning diversity and performance. Drawing inspiration from multi-teacher strategies in knowledge distillation, we introduce Adaptive Multi-Guidance Policy Optimization (AMPO), a novel framework that adaptively leverages guidance from multiple proficient teacher models, but only when the on-policy model fails to generate correct solutions. This "guidance-on-demand" approach expands exploration while preserving the value of self-discovery. Moreover, AMPO incorporates a comprehension-based selection mechanism, prompting the student to learn from the reasoning paths that it is most likely to comprehend, thus balancing broad exploration with effective exploitation. Extensive experiments show AMPO substantially outperforms a strong baseline (GRPO), with a 4.3% improvement on mathematical reasoning tasks and 12.2% on out-of-distribution tasks, while significantly boosting Pass@k performance and enabling more diverse exploration. Notably, using four peer-sized teachers, our method achieves comparable results to approaches that leverage a single, more powerful teacher (e.g., DeepSeek-R1) with more data. These results demonstrate a more efficient and scalable path to superior reasoning and generalizability. Our code is available at https://github.com/SII-Enigma/AMPO.
Authors: Xiangyu Peng, Can Qin, Zeyuan Chen, Ran Xu, Caiming Xiong, Chien-Sheng Wu
Abstract: Multimodal retrieval-augmented generation (MM-RAG) is a key approach for applying large language models (LLMs) and agents to real-world knowledge bases, yet current evaluations are fragmented, focusing on either text or images in isolation or on simplified multimodal setups that fail to capture document-centric multimodal use cases. In this paper, we introduce UniDoc-Bench, the first large-scale, realistic benchmark for MM-RAG built from 70k real-world PDF pages across eight domains. Our pipeline extracts and links evidence from text, tables, and figures, then generates 1,600 multimodal QA pairs spanning factual retrieval, comparison, summarization, and logical reasoning queries. To ensure reliability, 20% of QA pairs are validated by multiple annotators and expert adjudication. UniDoc-Bench supports apples-to-apples comparison across four paradigms: (1) text-only, (2) image-only, (3) multimodal text-image fusion, and (4) multimodal joint retrieval -- under a unified protocol with standardized candidate pools, prompts, and evaluation metrics. Our experiments show that multimodal text-image fusion RAG systems consistently outperform both unimodal and jointly multimodal embedding-based retrieval, indicating that neither text nor images alone are sufficient and that current multimodal embeddings remain inadequate. Beyond benchmarking, our analysis reveals when and how visual context complements textual evidence, uncovers systematic failure modes, and offers actionable guidance for developing more robust MM-RAG pipelines.
Authors: Shaoyi Zheng, Canyu Zhang, Tianyi Zhou, Shengjie Wang
Abstract: In-context learning (ICL) enables efficient few-shot learning in large language models (LLMs) without training, but suffers from the quadratic input complexity of transformers, limiting the maximum number of exemplars. While various efficient ICL approaches partition the context into blocks to process (e.g., ensembling, compression, cross-attention), they often ignore the information redundancy or under-representation caused by different partition strategies, leading to suboptimal performance. To tackle this problem, we propose Sub-CP, a block-aware context selection framework that leverages submodular objectives to control block diversity. Sub-CP supports a flexible spectrum of selection strategies, allowing each block to range from globally diverse to locally coherent. This allows fine-grained control over semantic structure while enabling precomputation. Extensive experiments across diverse tasks on multiple datasets show that Sub-CP consistently improves performance across model scales.
Authors: Donghang Wu, Haoyang Zhang, Chen Chen, Tianyu Zhang, Fei Tian, Xuerui Yang, Gang Yu, Hexin Liu, Nana Hou, Yuchen Hu, Eng Siong Chng
Abstract: Recent advances in spoken dialogue language models (SDLMs) reflect growing interest in shifting from turn-based to full-duplex systems, where the models continuously perceive user speech streams while generating responses. This simultaneous listening and speaking design enables real-time interaction and the agent can handle dynamic conversational behaviors like user barge-in. However, during the listening phase, existing systems keep the agent idle by repeatedly predicting the silence token, which departs from human behavior: we usually engage in lightweight thinking during conversation rather than remaining absent-minded. Inspired by this, we propose Chronological Thinking, a on-the-fly conversational thinking mechanism that aims to improve response quality in full-duplex SDLMs. Specifically, chronological thinking presents a paradigm shift from conventional LLM thinking approaches, such as Chain-of-Thought, purpose-built for streaming acoustic input. (1) Strictly causal: the agent reasons incrementally while listening, updating internal hypotheses only from past audio with no lookahead. (2) No additional latency: reasoning is amortized during the listening window; once the user stops speaking, the agent halts thinking and begins speaking without further delay. Experiments demonstrate the effectiveness of chronological thinking through both objective metrics and human evaluations show consistent improvements in response quality. Furthermore, chronological thinking robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.
Authors: Yao Dou, Michel Galley, Baolin Peng, Chris Kedzie, Weixin Cai, Alan Ritter, Chris Quirk, Wei Xu, Jianfeng Gao
Abstract: Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce SimulatorArena, a benchmark of 909 annotated human-LLM conversations on two interactive tasks -- math tutoring and document creation. SimulatorArena evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman's $\rho$ of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation. Using the best simulator for each task, we benchmark 18 assistants, including the latest LLMs such as GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro.
Authors: Xue-Yong Fu, Elena Khasanova, Md Tahmid Rahman Laskar, Harsh Saini, Shashi Bhushan TN
Abstract: Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
Authors: Gagan Bhatia, Somayajulu G Sripada, Kevin Allan, Jacobo Azcona
Abstract: Large Language Models (LLMs) are prone to hallucination, the generation of plausible yet factually incorrect statements. This work investigates the intrinsic, architectural origins of this failure mode through three primary contributions. First, to enable the reliable tracing of internal semantic failures, we propose Distributional Semantics Tracing (DST), a unified framework that integrates established interpretability techniques to produce a causal map of a model's reasoning, treating meaning as a function of context (distributional semantics). Second, we pinpoint the model's layer at which a hallucination becomes inevitable, identifying a specific commitment layer where a model's internal representations irreversibly diverge from factuality. Third, we identify the underlying mechanism for these failures. We observe a conflict between distinct computational pathways, which we interpret using the lens of dual-process theory: a fast, heuristic associative pathway (akin to System 1) and a slow, deliberate, contextual pathway (akin to System 2), leading to predictable failure modes such as Reasoning Shortcut Hijacks. Our framework's ability to quantify the coherence of the contextual pathway reveals a strong negative correlation ($\rho = -0.863$) with hallucination rates, implying that these failures are predictable consequences of internal semantic weakness. The result is a mechanistic account of how, when, and why hallucinations occur within the Transformer architecture.
Authors: Aisha Alansari, Hamzah Luqman
Abstract: Large language models (LLMs) have transformed natural language processing, achieving remarkable performance across diverse tasks. However, their impressive fluency often comes at the cost of producing false or fabricated information, a phenomenon known as hallucination. Hallucination refers to the generation of content by an LLM that is fluent and syntactically correct but factually inaccurate or unsupported by external evidence. Hallucinations undermine the reliability and trustworthiness of LLMs, especially in domains requiring factual accuracy. This survey provides a comprehensive review of research on hallucination in LLMs, with a focus on causes, detection, and mitigation. We first present a taxonomy of hallucination types and analyze their root causes across the entire LLM development lifecycle, from data collection and architecture design to inference. We further examine how hallucinations emerge in key natural language generation tasks. Building on this foundation, we introduce a structured taxonomy of detection approaches and another taxonomy of mitigation strategies. We also analyze the strengths and limitations of current detection and mitigation approaches and review existing evaluation benchmarks and metrics used to quantify LLMs hallucinations. Finally, we outline key open challenges and promising directions for future research, providing a foundation for the development of more truthful and trustworthy LLMs.
Authors: Yisha Wu, Cen Mia Zhao, Yuanpei Cao, Xiaoqing Su, Yashar Mehdad, Mindy Ji, Claire Na Cheng
Abstract: We introduce an incremental summarization system for customer support agents that intelligently determines when to generate concise bullet notes during conversations, reducing agents' context-switching effort and redundant review. Our approach combines a fine-tuned Mixtral-8x7B model for continuous note generation with a DeBERTa-based classifier to filter trivial content. Agent edits refine the online notes generation and regularly inform offline model retraining, closing the agent edits feedback loop. Deployed in production, our system achieved a 3% reduction in case handling time compared to bulk summarization (with reductions of up to 9% in highly complex cases), alongside high agent satisfaction ratings from surveys. These results demonstrate that incremental summarization with continuous feedback effectively enhances summary quality and agent productivity at scale.
Authors: Luca Giordano, Simon Razniewski
Abstract: Large Language Models (LLMs) encode substantial factual knowledge, yet measuring and systematizing this knowledge remains challenging. Converting it into structured format, for example through recursive extraction approaches such as the GPTKB methodology (Hu et al., 2025b), is still underexplored. Key open questions include whether such extraction can terminate, whether its outputs are reproducible, and how robust they are to variations. We systematically study LLM knowledge materialization using miniGPTKBs (domain-specific, tractable subcrawls), analyzing termination, reproducibility, and robustness across three categories of metrics: yield, lexical similarity, and semantic similarity. We experiment with four variations (seed, language, randomness, model) and three illustrative domains (from history, entertainment, and finance). Our findings show (i) high termination rates, though model-dependent; (ii) mixed reproducibility; and (iii) robustness that varies by perturbation type: high for seeds and temperature, lower for languages and models. These results suggest that LLM knowledge materialization can reliably surface core knowledge, while also revealing important limitations.
Authors: Chenpeng Wang, Xiaojie Cheng, Chunye Wang, Linfeng Yang, Lei Zhang
Abstract: Tool-augmented language models have demonstrated strong capabilities, but their reliance on live API access creates scalability and reliability challenges during training and deployment. We propose MTR, a simulation-first training framework for tool-augmented reasoning. Instead of relying on live APIs, MTR learns from complete ReAct traces with schema-validated, simulated observations. Our approach operates through a multi-agent architecture where a ToolMaker generates task-specific, OpenAI-compatible tool interfaces, an AutoAgent produces structured think-act-observe sequences, and a ToolActor simulates realistic responses. Training proceeds in two stages: Stage-1 Supervised Fine-Tuning (SFT) teaches 'trace grammar' from complete reasoning sequences; Stage-2 Group Relative Policy Optimization (GRPO) optimizes strategy with a composite trace reward that balances answer correctness and internal consistency. Across four multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle), MTR attains competitive Exact Match (EM) scores to live-API systems and excels on reasoning-intensive tasks, suggesting that effective tool reasoning can be learned from structured traces without live interactions.
Authors: Yining Wang, Jinman Zhao, Chuangxin Zhao, Shuhao Guan, Gerald Penn, Shinan Liu
Abstract: Reinforcement Learning with Human Feedback (RLHF) has been the dominant approach for improving the reasoning capabilities of Large Language Models (LLMs). Recently, Reinforcement Learning with Verifiable Rewards (RLVR) has simplified this paradigm by replacing the reward and value models with rule-based verifiers. A prominent example is Group Relative Policy Optimization (GRPO). However, GRPO inherently suffers from a length bias, since the same advantage is uniformly assigned to all tokens of a response. As a result, longer responses distribute the reward over more tokens and thus contribute disproportionately to gradient updates. Several variants, such as DAPO and Dr. GRPO, modify the token-level aggregation of the loss, yet these methods remain heuristic and offer limited interpretability regarding their implicit token preferences. In this work, we explore the possibility of allowing the model to learn its own token preference during optimization. We unify existing frameworks under a single formulation and introduce a learnable parameter $\lambda$ that adaptively controls token-level weighting. We use $\lambda$-GRPO to denote our method, and we find that $\lambda$-GRPO achieves consistent improvements over vanilla GRPO and DAPO on multiple mathematical reasoning benchmarks. On Qwen2.5 models with 1.5B, 3B, and 7B parameters, $\lambda$-GRPO improves average accuracy by $+1.9\%$, $+1.0\%$, and $+1.7\%$ compared to GRPO, respectively. Importantly, these gains come without any modifications to the training data or additional computational cost, highlighting the effectiveness and practicality of learning token preferences.
Authors: Vaibhav Srivastav, Steven Zheng, Eric Bezzam, Eustache Le Bihan, Nithin Koluguri, Piotr \.Zelasko, Somshubra Majumdar, Adel Moumen, Sanchit Gandhi
Abstract: Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.
Authors: Shrestha Ghosh, Luca Giordano, Yujia Hu, Tuan-Phong Nguyen, Simon Razniewski
Abstract: LLMs are remarkable artifacts that have revolutionized a range of NLP and AI tasks. A significant contributor is their factual knowledge, which, to date, remains poorly understood, and is usually analyzed from biased samples. In this paper, we take a deep tour into the factual knowledge (or beliefs) of a frontier LLM, based on GPTKB v1.5 (Hu et al., 2025a), a recursively elicited set of 100 million beliefs of one of the strongest currently available frontier LLMs, GPT-4.1. We find that the models' factual knowledge differs quite significantly from established knowledge bases, and that its accuracy is significantly lower than indicated by previous benchmarks. We also find that inconsistency, ambiguity and hallucinations are major issues, shedding light on future research opportunities concerning factual LLM knowledge.
Authors: Rajvee Sheth, Samridhi Raj Sinha, Mahavir Patil, Himanshu Beniwal, Mayank Singh
Abstract: Code-switching (CSW), the alternation of languages and scripts within a single utterance, remains a fundamental challenge for multiling ual NLP, even amidst the rapid advances of large language models (LLMs). Most LLMs still struggle with mixed-language inputs, limited CSW datasets, and evaluation biases, hindering deployment in multilingual societies. This survey provides the first comprehensive analysis of CSW-aware LLM research, reviewing 308 studies spanning five research areas, 12 NLP tasks, 30+ datasets, and 80+ languages. We classify recent advances by architecture, training strategy, and evaluation methodology, outlining how LLMs have reshaped CSW modeling and what challenges persist. The paper concludes with a roadmap emphasizing the need for inclusive datasets, fair evaluation, and linguistically grounded models to achieve truly multilingual intelligence. A curated collection of all resources is maintained at https://github.com/lingo-iitgn/awesome-code-mixing/.
Authors: Donggyu Lee, Sungwon Park, Yerin Hwang, Hyoshin Kim, Hyunwoo Oh, Jungwon Kim, Meeyoung Cha, Sangyoon Park, Jihee Kim
Abstract: Causal reasoning is fundamental for Large Language Models (LLMs) to understand genuine cause-and-effect relationships beyond pattern matching. Existing benchmarks suffer from critical limitations such as reliance on synthetic data and narrow domain coverage. We introduce a novel benchmark constructed from casually identified relationships extracted from top-tier economics and finance journals, drawing on rigorous methodologies including instrumental variables, difference-in-differences, and regression discontinuity designs. Our benchmark comprises 40,379 evaluation items covering five task types across domains such as health, environment, technology, law, and culture. Experimental results on eight state-of-the-art LLMs reveal substantial limitations, with the best model achieving only 57.6\% accuracy. Moreover, model scale does not consistently translate to superior performance, and even advanced reasoning models struggle with fundamental causal relationship identification. These findings underscore a critical gap between current LLM capabilities and demands of reliable causal reasoning in high-stakes applications.
Authors: Leitian Tao, Ilia Kulikov, Swarnadeep Saha, Tianlu Wang, Jing Xu, Yixuan Li, Jason E Weston, Ping Yu
Abstract: Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.
Authors: Guangliang Liu, Haitao Mao, Bochuan Cao, Zhiyu Xue, Xitong Zhang, Rongrong Wang, Kristen Marie Johnson
Abstract: Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.
Authors: Yue Li, Ran Tao, Derek Hommel, Yusuf Denizay D\"onder, Sungyong Chang, David Mimno, Unso Eun Seo Jo
Abstract: In the business domain, where data-driven decision making is crucial, text-to-SQL is fundamental for easy natural language access to structured data. While recent LLMs have achieved strong performance in code generation, existing text-to-SQL benchmarks remain focused on factual retrieval of past records. We introduce CORGI, a new benchmark specifically designed for real-world business contexts. CORGI is composed of synthetic databases inspired by enterprises such as Doordash, Airbnb, and Lululemon. It provides questions across four increasingly complex categories of business queries: descriptive, explanatory, predictive, and recommendational. This challenge calls for causal reasoning, temporal forecasting, and strategic recommendation, reflecting multi-level and multi-step agentic intelligence. We find that LLM performance drops on high-level questions, struggling to make accurate predictions and offer actionable plans. Based on execution success rate, the CORGI benchmark is about 21% more difficult than the BIRD benchmark. This highlights the gap between popular LLMs and the need for real-world business intelligence. We release a public dataset and evaluation framework, and a website for public submissions.
Authors: Changhao Li, Yuchen Zhuang, Rushi Qiang, Haotian Sun, Hanjun Dai, Chao Zhang, Bo Dai
Abstract: Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Matryoshka Pilot (M-Pilot), a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with M-Pilot serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. M-Pilot is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on diverse tasks demonstrate that our method effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks.
Authors: Jiabin Tang, Tianyu Fan, Chao Huang
Abstract: Large Language Model (LLM) Agents have demonstrated remarkable capabilities in task automation and intelligent decision-making, driving the widespread adoption of agent development frameworks such as LangChain and AutoGen. However, these frameworks predominantly serve developers with extensive technical expertise - a significant limitation considering that only 0.03 % of the global population possesses the necessary programming skills. This stark accessibility gap raises a fundamental question: Can we enable everyone, regardless of technical background, to build their own LLM agents using natural language alone? To address this challenge, we introduce AutoAgent-a Fully-Automated and highly Self-Developing framework that enables users to create and deploy LLM agents through Natural Language Alone. Operating as an autonomous Agent Operating System, AutoAgent comprises four key components: i) Agentic System Utilities, ii) LLM-powered Actionable Engine, iii) Self-Managing File System, and iv) Self-Play Agent Customization module. This lightweight yet powerful system enables efficient and dynamic creation and modification of tools, agents, and workflows without coding requirements or manual intervention. Beyond its code-free agent development capabilities, AutoAgent also serves as a versatile multi-agent system for General AI Assistants. Comprehensive evaluations on the GAIA benchmark demonstrate AutoAgent's effectiveness in generalist multi-agent tasks, surpassing existing state-of-the-art methods. Furthermore, AutoAgent's Retrieval-Augmented Generation (RAG)-related capabilities have shown consistently superior performance compared to many alternative LLM-based solutions.
Authors: Kyle Gao, Dening Lu, Liangzhi Li, Nan Chen, Hongjie He, Linlin Xu, Jonathan Li
Abstract: The Los Angeles wildfires of January 2025 caused more than 250 billion dollars in damage and lasted for nearly an entire month before containment. Following our previous work, the Digital Twin Building, we modify and leverage the multi-agent large language model framework as well as the cloud-mapping integration to study the air quality during the Los Angeles wildfires. Recent advances in large language models have allowed for out-of-the-box automated large-scale data analysis. We use a multi-agent large language system comprised of an Instructor agent and Worker agents. Upon receiving the users' instructions, the Instructor agent retrieves the data from the cloud platform and produces instruction prompts to the Worker agents. The Worker agents then analyze the data and provide summaries. The summaries are finally input back into the Instructor agent, which then provides the final data analysis. We test this system's capability for data-based policy recommendation by assessing our Instructor-Worker LLM system's health recommendations based on air quality during the Los Angeles wildfires.
Authors: Lang Cao, Renhong Chen, Yingtian Zou, Chao Peng, Huacong Xu, Yuxian Wang, Wu Ning, Qian Chen, Mofan Peng, Zijie Chen, Peishuo Su, Sirui Han, Yitong Li
Abstract: We introduce the Entropy-Driven Uncertainty Process Reward Model (EDU-PRM), a novel entropy-driven training framework for process reward modeling that enables dynamic, uncertainty-aligned segmentation of complex reasoning steps, eliminating the need for costly manual step annotations. Unlike previous Process Reward Models (PRMs) that rely on static partitioning and human labeling, EDU-PRM automatically anchors step boundaries at tokens with high predictive entropy, effectively capturing intrinsic logical transitions and facilitating efficient exploration of diverse reasoning paths. On the ProcessBench benchmark, EDU-PRM outperforms strong public PRM baselines, such as Math-Shepherd PRM and Omega PRM, and EDU-PRM achieves comparable results with SOTA models while only using 1.5% training data. Furthermore, by leveraging our proposed EDU sampling strategy, we observe accuracy boosts from 64.7% to 67.3% for generative reasoning tasks, accompanied by a reduction of 32% in token usage. These findings underscore the potential of EDU-PRM as a scalable and annotation-efficient paradigm for process supervision in mathematical reasoning, paving the way for more efficient and robust approaches to complex mathematical problem solving.
Authors: Rana Muhammad Shahroz Khan, Zhen Tan, Sukwon Yun, Charles Fleming, Tianlong Chen
Abstract: Most discussions about Large Language Model (LLM) safety have focused on single-agent settings but multi-agent LLM systems now create novel adversarial risks because their behavior depends on communication between agents and decentralized reasoning. In this work, we innovatively focus on attacking pragmatic systems that have constrains such as limited token bandwidth, latency between message delivery, and defense mechanisms. We design a $\textit{permutation-invariant adversarial attack}$ that optimizes prompt distribution across latency and bandwidth-constraint network topologies to bypass distributed safety mechanisms within the system. Formulating the attack path as a problem of $\textit{maximum-flow minimum-cost}$, coupled with the novel $\textit{Permutation-Invariant Evasion Loss (PIEL)}$, we leverage graph-based optimization to maximize attack success rate while minimizing detection risk. Evaluating across models including $\texttt{Llama}$, $\texttt{Mistral}$, $\texttt{Gemma}$, $\texttt{DeepSeek}$ and other variants on various datasets like $\texttt{JailBreakBench}$ and $\texttt{AdversarialBench}$, our method outperforms conventional attacks by up to $7\times$, exposing critical vulnerabilities in multi-agent systems. Moreover, we demonstrate that existing defenses, including variants of $\texttt{Llama-Guard}$ and $\texttt{PromptGuard}$, fail to prohibit our attack, emphasizing the urgent need for multi-agent specific safety mechanisms.
Authors: Hengran Zhang, Minghao Tang, Keping Bi, Jiafeng Guo, Shihao Liu, Daiting Shi, Dawei Yin, Xueqi Cheng
Abstract: This paper explores the use of large language models (LLMs) for annotating document utility in training retrieval and retrieval-augmented generation (RAG) systems, aiming to reduce dependence on costly human annotations. We address the gap between retrieval relevance and generative utility by employing LLMs to annotate document utility. To effectively utilize multiple positive samples per query, we introduce a novel loss that maximizes their summed marginal likelihood. Using the Qwen-2.5-32B model, we annotate utility on the MS MARCO dataset and conduct retrieval experiments on MS MARCO and BEIR, as well as RAG experiments on MS MARCO QA, NQ, and HotpotQA. Our results show that LLM-generated annotations enhance out-of-domain retrieval performance and improve RAG outcomes compared to models trained solely on human annotations or downstream QA metrics. Furthermore, combining LLM annotations with just 20% of human labels achieves performance comparable to using full human annotations. Our study offers a comprehensive approach to utilizing LLM annotations for initializing QA systems on new corpora.
Authors: Ke Hong, Xiuhong Li, Minxu Liu, Qiuli Mao, Tianqi Wu, Zixiao Huang, Lufang Chen, Zhong Wang, Yichong Zhang, Zhenhua Zhu, Guohao Dai, Yu Wang
Abstract: Generative models have achieved remarkable success across various applications, driving the demand for multi-GPU computing. Inter-GPU communication becomes a bottleneck in multi-GPU computing systems, particularly on consumer-grade GPUs. By exploiting concurrent hardware execution, overlapping computation and communication latency becomes an effective technique for mitigating the communication overhead. We identify that an efficient and adaptable overlapping design should satisfy (1) tile-wise overlapping to maximize the overlapping opportunity, (2) interference-free computation to maintain the original computational performance, and (3) communication agnosticism to reduce the development burden against varying communication primitives. Nevertheless, current designs fail to simultaneously optimize for all of those features. To address the issue, we propose FlashOverlap, which utilizes a novel signaling mechanism: when part of the output finishes, the computation kernel sends a signal to trigger the communication of that part, while continuing the computation of the remaining part (interference-free computation). Consequently, the communication of the finished part and the computation of the remaining part can be overlapped. On top of the signaling mechanism, FlashOverlap comprises two key components: (1) the determination of the signaling timing to boost the overlap efficiency (tile-wise overlapping), and (2) a pre-communication reordering to create the contiguous address for finished data, enabling communication by simply calling NCCL APIs (communication agnosticism), and a post-communication reordering to correct the data order. Experiments show that FlashOverlap achieves up to 1.65x speedup through overlap, outperforming existing works in most cases. Code is available at https://github.com/infinigence/FlashOverlap.
Authors: Tianyu Liu, Simeng Han, Xiao Luo, Hanchen Wang, Pan Lu, Biqing Zhu, Yuge Wang, Keyi Li, Jiapeng Chen, Rihao Qu, Yufeng Liu, Xinyue Cui, Aviv Yaish, Yuhang Chen, Minsheng Hao, Chuhan Li, Kexing Li, Arman Cohan, Hua Xu, Mark Gerstein, James Zou, Hongyu Zhao
Abstract: Large language models (LLMs) and large multimodal models (LMMs) promise to accelerate biomedical discovery, yet their reliability remains unclear. We introduce ARIEL (AI Research Assistant for Expert-in-the-Loop Learning), an open-source evaluation and optimization framework that pairs a curated multimodal biomedical corpus with expert-vetted tasks to probe two capabilities: full-length article summarization and fine-grained figure interpretation. Using uniform protocols and blinded PhD-level evaluation, we find that state-of-the-art models generate fluent but incomplete summaries, whereas LMMs struggle with detailed visual reasoning. We later observe that prompt engineering and lightweight fine-tuning substantially improve textual coverage, and a compute-scaled inference strategy enhances visual question answering. We build an ARIEL agent that integrates textual and visual cues, and we show it can propose testable mechanistic hypotheses. ARIEL delineates current strengths and limitations of foundation models, and provides a reproducible platform for advancing trustworthy AI in biomedicine.
Authors: Xinyan Hu, Kayo Yin, Michael I. Jordan, Jacob Steinhardt, Lijie Chen
Abstract: To perform few-shot learning, language models extract signals from a few input-label pairs, aggregate these into a learned prediction rule, and apply this rule to new inputs. How is this implemented in the forward pass of modern transformer models? To explore this question, we study a structured family of few-shot learning tasks for which the true prediction rule is to add an integer $k$ to the input. We introduce a novel optimization method that localizes the model's few-shot ability to only a few attention heads. We then perform an in-depth analysis of individual heads, via dimensionality reduction and decomposition. As an example, on Llama-3-8B-instruct, we reduce its mechanism on our tasks to just three attention heads with six-dimensional subspaces, where four dimensions track the unit digit with trigonometric functions at periods $2$, $5$, and $10$, and two dimensions track magnitude with low-frequency components. To deepen our understanding of the mechanism, we also derive a mathematical identity relating ``aggregation'' and ``extraction'' subspaces for attention heads, allowing us to track the flow of information from individual examples to a final aggregated concept. Using this, we identify a self-correction mechanism where mistakes learned from earlier demonstrations are suppressed by later demonstrations. Our results demonstrate how tracking low-dimensional subspaces of localized heads across a forward pass can provide insight into fine-grained computational structures in language models.
Authors: Honglin Bao, Siyang Wu, Jiwoong Choi, Yingrong Mao, James A. Evans
Abstract: This paper calls on the research community not only to investigate how human biases are inherited by large language models (LLMs) but also to explore how these biases in LLMs can be leveraged to make society's "unwritten code" - such as implicit stereotypes and heuristics - visible and accessible for critique. We introduce a conceptual framework through a case study in science: uncovering hidden rules in peer review - the factors that reviewers care about but rarely state explicitly due to normative scientific expectations. The idea of the framework is to push LLMs to speak out their heuristics through generating self-consistent hypotheses - why one paper appeared stronger in reviewer scoring - among paired papers submitted to 45 academic conferences, while iteratively searching deeper hypotheses from remaining pairs where existing hypotheses cannot explain. We observed that LLMs' normative priors about the internal characteristics of good science extracted from their self-talk, e.g., theoretical rigor, were systematically updated toward posteriors that emphasize storytelling about external connections, such as how the work is positioned and connected within and across literatures. Human reviewers tend to explicitly reward aspects that moderately align with LLMs' normative priors (correlation = 0.49) but avoid articulating contextualization and storytelling posteriors in their review comments (correlation = -0.14), despite giving implicit reward to them with positive scores. These patterns are robust across different models and out-of-sample judgments. We discuss the broad applicability of our proposed framework, leveraging LLMs as diagnostic tools to amplify and surface the tacit codes underlying human society, enabling public discussion of revealed values and more precisely targeted responsible AI.
Authors: Ruida Wang, Yuxin Li, Yi R. Fung, Tong Zhang
Abstract: Enhancing the mathematical reasoning capabilities of LLMs has garnered significant attention in both the mathematical and computer science communities. Recent works have made substantial progress in both Natural Language (NL) reasoning and Formal Language (FL) reasoning by leveraging the potential of pure Reinforcement Learning (RL) methods on base models. However, RL approaches struggle to impart new capabilities not presented in the base model, highlighting the need to integrate more knowledge like FL into NL math reasoning effectively. Yet, this integration is challenging due to inherent disparities in problem structure and reasoning format between NL and FL. To address these challenges, we introduce **NL-FL HybridReasoning (NFL-HR)**, an end-to-end framework designed to incorporate the FL expert into NL math problem-solving. To bridge the NL and FL input format gap, we propose the NL-FL Problem Alignment method, which reformulates the Question-Answering (QA) problems in NL as existence theorems in FL. Subsequently, the Mixed Problem Input technique we provide enables the FL reasoner to handle both QA and existence problems concurrently. Lastly, we mitigate the NL and FL output format gap in reasoning through an LLM-based Answer Extraction mechanism. Comprehensive experiments demonstrate that the NFL-HR framework achieves **89.80**% and **84.34%** accuracy rates on the MATH-500 and the AMC benchmarks, surpassing the NL baseline by **4.60%** and **4.82%**, respectively. Notably, some problems resolved by our framework remain unsolved by the NL baseline model even under a larger number of trials.
Authors: Zory Zhang, Pinyuan Feng, Bingyang Wang, Tianwei Zhao, Suyang Yu, Qingying Gao, Hokin Deng, Ziqiao Ma, Yijiang Li, Dezhi Luo
Abstract: The ability to infer what others are looking at is a critical component of a theory of mind that underpins natural human-AI interaction. We characterized this skill in 111 Vision Language Models (VLMs) and human participants (N = 65) using photos taken with manipulated difficulty and variability. We found that 94 of the 111 VLMs were not better than random guessing, while humans achieved near-ceiling accuracy. VLMs respond with each choice almost equally frequently. Are they randomly guessing? At least for five top-tier VLMs, their performance was above chance, declined with increasing task difficulty, but barely varied across different prompts and scene objects. These behavioral patterns cannot be explained by considering VLMs as random guessers. Instead, they likely utilize head orientation but not eye appearance to infer gaze direction, such that their performance is imperfect, subject to the task difficulty, but robust to superficial perceptual variations. This suggests that VLMs, lacking effective gaze inference skills, have yet to become technologies that can naturally interact with humans, but the potential remains.
Authors: Yunfei Xie, Yinsong Ma, Shiyi Lan, Alan Yuille, Junfei Xiao, Chen Wei
Abstract: Developing reasoning capabilities in multimodal large language models (MLLMs) remains challenging. Motivated by literature suggesting that gameplay promotes transferable reasoning skills, we propose a novel post-training method, Visual Game Learning (ViGaL), where MLLMs develop generalizable reasoning skills through playing arcade-like games. Specifically, we show that training a 7B-parameter MLLM via reinforcement learning (RL) on simple games like Snake significantly enhances the downstream performance on multimodal math benchmarks like MathVista, on multi-discipline questions like MMMU and on 3D spatial reasoning benchmarks like VSI-Bench, without seeing any worked solutions, equations, or diagrams during RL. Remarkably, our model outperforms specialist models post-trained on benchmark-oriented multimodal reasoning data, while preserving the model's performance on general visual benchmarks, a challenge where specialist models often fall short. Our findings suggest that multimodal reasoning can emerge from gameplay, pointing to a promising strategy of designing surrogate tasks for RL post-training.
Authors: Chenxi Liu, Tianyi Xiong, Yanshuo Chen, 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: Huaying Yuan, Zheng Liu, Junjie Zhou, Hongjin Qian, Yan Shu, Nicu Sebe, Ji-Rong Wen, Zhicheng Dou
Abstract: Long-video understanding~(LVU) is a challenging problem in computer vision. Existing methods either downsample frames for single-pass reasoning, sacrificing fine-grained details, or depend on textual reasoning over task-agnostic representations, hindering task-specific perception and exploration. In this paper, we propose VideoExplorer, a framework grounded in the principle of ``thinking with video'', which naturally intertwines planning, temporal grounding, and scalable perception into a coherent reasoning process. Rather than reasoning over a static context, VideoExplorer iteratively formulates sub-questions, locates relevant moments, and performs task-oriented, temporally scalable video understanding until reaching the final answer, enabling faithful, efficient, and interpretable reasoning. To address the lack of LVU training resources, we construct a long-video reasoning dataset using difficulty-adaptive sampling to ensure high-quality trajectories on complex tasks. Building on this dataset, we design a two-stage training pipeline: supervised trajectory initialization followed by trajectory-level preference optimization, encouraging adaptive temporal grounding and iterative information integration guided by downstream rewards. Extensive evaluations on popular long-video understanding and reasoning benchmarks demonstrate VideoExplorer's significant advantage over existing baselines, highlighting its robustness, adaptability, and efficiency. Our code is made publicly available in this repository(https://github.com/yhy-2000/VideoDeepResearch).
Authors: Zohaib Hasan Siddiqui, Jiechao Gao, Ebad Shabbir, Mohammad Anas Azeez, Rafiq Ali, Gautam Siddharth Kashyap, Usman Naseem
Abstract: Running Large Language Models (LLMs) on edge devices is constrained by high compute and memory demands posing a barrier for real-time applications in sectors like healthcare, education, and embedded systems. Current solutions such as quantization, pruning, and retrieval-augmented generation (RAG) offer only partial optimizations and often compromise on speed or accuracy. We introduce HOLA, an end-to-end optimization framework for efficient LLM deployment. Internally, it leverages Hierarchical Speculative Decoding (HSD) for faster inference without quality loss. Externally, AdaComp-RAG adjusts retrieval complexity based on context needs. Together with LoBi, which blends structured pruning (LoRA) and quantization, HOLA delivers significant gains: 17.6% EMA on GSM8K, 10.5% MCA on ARC, and reduced latency and memory on edge devices like Jetson Nano--proving both scalable and production-ready.
Authors: Ashutosh Hathidara, Julien Yu, Sebastian Schreiber
Abstract: Large language models (LLMs) are increasingly tasked with invoking enterprise APIs, yet they routinely falter when near-duplicate tools vie for the same user intent or when required arguments are left underspecified. We introduce DiaFORGE (Dialogue Framework for Organic Response Generation & Evaluation), a disambiguation-centric, three-stage pipeline that (i) synthesizes persona-driven, multi-turn dialogues in which the assistant must distinguish among highly similar tools, (ii) performs supervised fine-tuning of open-source models with reasoning traces across 3B - 70B parameters, and (iii) evaluates real-world readiness via a dynamic suite that redeploys each model in a live agentic loop and reports end-to-end goal completion alongside conventional static metrics. On our dynamic benchmark DiaBENCH, models trained with DiaFORGE raise tool-invocation success by 27 pp over GPT-4o and by 49 pp over Claude-3.5-Sonnet, both under optimized prompting. To spur further research, we release an open corpus of 5000 production-grade enterprise API specifications paired with rigorously validated, disambiguation-focused dialogues, offering a practical blueprint for building reliable, enterprise-ready tool-calling agents.
Authors: Ang Li, Charles Wang, Deqing Fu, Kaiyu Yue, Zikui Cai, Wang Bill Zhu, Ollie Liu, Peng Guo, Willie Neiswanger, Furong Huang, Tom Goldstein, Micah Goldblum
Abstract: Humans often use visual aids, for example diagrams or sketches, when solving complex problems. Training multimodal models to do the same, known as Visual Chain of Thought (Visual CoT), is challenging due to: (1) poor off-the-shelf visual CoT performance, which hinders reinforcement learning, and (2) the lack of high-quality visual CoT training data. We introduce $\textbf{Zebra-CoT}$, a diverse large-scale dataset with 182,384 samples, containing logically coherent interleaved text-image reasoning traces. We focus on four categories of tasks where sketching or visual reasoning is especially natural, spanning scientific questions such as geometry, physics, and algorithms; 2D visual reasoning tasks like visual search and jigsaw puzzles; 3D reasoning tasks including 3D multi-hop inference, embodied and robot planning; visual logic problems and strategic games like chess. Fine-tuning the Anole-7B model on the Zebra-CoT training corpus results in an improvement of +12% in our test-set accuracy and yields up to +13% performance gain on standard VLM benchmark evaluations. Fine-tuning Bagel-7B yields a model that generates high-quality interleaved visual reasoning chains, underscoring Zebra-CoT's effectiveness for developing multimodal reasoning abilities. We open-source our dataset and models to support development and evaluation of visual CoT.
Authors: Hengran Zhang, Keping Bi, Jiafeng Guo, Jiaming Zhang, Shuaiqiang Wang, Dawei Yin, Xueqi Cheng
Abstract: Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating retrieved information. Standard retrieval process prioritized relevance, focusing on topical alignment between queries and passages. In contrast, in RAG, the emphasis has shifted to utility, which considers the usefulness of passages for generating accurate answers. Despite empirical evidence showing the benefits of utility-based retrieval in RAG, the high computational cost of using LLMs for utility judgments limits the number of passages evaluated. This restriction is problematic for complex queries requiring extensive information. To address this, we propose a method to distill the utility judgment capabilities of LLMs into smaller, more efficient models. Our approach focuses on utility-based selection rather than ranking, enabling dynamic passage selection tailored to specific queries without the need for fixed thresholds. We train student models to learn pseudo-answer generation and utility judgments from teacher LLMs, using a sliding window method that dynamically selects useful passages. Our experiments demonstrate that utility-based selection provides a flexible and cost-effective solution for RAG, significantly reducing computational costs while improving answer quality. We present the distillation results using Qwen3-32B as the teacher model for both relevance ranking and utility-based selection, distilled into RankQwen1.7B and UtilityQwen1.7B. Our findings indicate that for complex questions, utility-based selection is more effective than relevance ranking in enhancing answer generation performance. We will release the relevance ranking and utility-based selection annotations for the MS MARCO dataset, supporting further research in this area.
Authors: Gaoke Zhang, Bo Wang, Yunlong Ma, Dongming Zhao, Zifei Yu
Abstract: An agent powered by large language models have achieved impressive results, but effectively handling the vast amounts of historical data generated during interactions remains a challenge. The current approach is to design a memory module for the agent to process these data. However, existing methods, such as MemoryBank and A-MEM, have poor quality of stored memory content, which affects recall performance and response quality. In order to better construct high-quality long-term memory content, we have designed a multiple memory system (MMS) inspired by cognitive psychology theory. The system processes short-term memory to multiple long-term memory fragments, and constructs retrieval memory units and contextual memory units based on these fragments, with a one-to-one correspondence between the two. During the retrieval phase, MMS will match the most relevant retrieval memory units based on the user's query. Then, the corresponding contextual memory units is obtained as the context for the response stage to enhance knowledge, thereby effectively utilizing historical data. Experiments on LoCoMo dataset compared our method with three others, proving its effectiveness. Ablation studies confirmed the rationality of our memory units. We also analyzed the robustness regarding the number of selected memory segments and the storage overhead, demonstrating its practical value.
Authors: Sudhanshu Agrawal, Risheek Garrepalli, Raghavv Goel, Mingu Lee, Christopher Lott, Fatih Porikli
Abstract: Diffusion LLMs (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs (AR-LLMs) with the potential to operate at significantly higher token generation rates. However, currently available open-source dLLMs often generate at much lower rates, typically decoding only a single token at every denoising timestep in order to maximize output quality. We present Spiffy, a speculative decoding algorithm that accelerates dLLM inference by $\mathbf{2.8{-}3.1\times}$ while provably preserving the model's output distribution. This work addresses the unique challenges involved in applying ideas from speculative decoding of AR-LLMs to the dLLM setting. Spiffy proposes draft states by leveraging the dLLM's distribution itself in an auto-speculative manner. This approach is efficient and effective, and eliminates the overheads of training and running an independent draft model. To structure the candidate draft states, we propose a novel directed draft graph which is uniquely designed to take advantage of the bidirectional, block-wise nature of dLLM generation and can be verified in parallel by the dLLM. To further optimize the structure of these draft graphs, we introduce an efficient, offline calibration algorithm that procedurally determines high-quality graph configurations. These optimized draft graphs, enabling increased acceptance rates, lead to a significant boost in the overall speedup achieved by the system. Crucially, Spiffy is also complementary to other recent innovations in improving dLLM generation speeds such as KV-caching and multi-token unmasking. We demonstrate that when combined with such parallel decoding algorithms, Spiffy is able to effectively multiply the benefits of these methods leading to total speedups of up to $\mathbf{7.9\times}$.
Authors: Yulei Qin, Xiaoyu Tan, Zhengbao He, Gang Li, Haojia Lin, Zongyi Li, Zihan Xu, Yuchen Shi, Siqi Cai, Renting Rui, Shaofei Cai, Yuzheng Cai, Xuan Zhang, Sheng Ye, Ke Li, Xing Sun
Abstract: Reinforcement learning (RL) is the dominant paradigm for sharpening strategic tool use capabilities of LLMs on long-horizon, sparsely-rewarded agent tasks, yet it faces a fundamental challenge of exploration-exploitation trade-off. Existing studies stimulate exploration through the lens of policy entropy, but such mechanical entropy maximization is prone to RL training instability due to the multi-turn distribution shifting. In this paper, we target the progressive exploration-exploitation balance under the guidance of the agent own experiences without succumbing to either entropy collapsing or runaway divergence. We propose SPEAR, a curriculum-based self-imitation learning (SIL) recipe for training agentic LLMs. It extends the vanilla SIL framework, where a replay buffer stores self-generated promising trajectories for off-policy update, by gradually steering the policy evolution within a well-balanced range of entropy across stages. Specifically, our approach incorporates a curriculum to manage the exploration process, utilizing intrinsic rewards to foster skill-level exploration and facilitating action-level exploration through SIL. At first, the auxiliary tool call reward plays a critical role in the accumulation of tool-use skills, enabling broad exposure to the unfamiliar distributions of the environment feedback with an upward entropy trend. As training progresses, self-imitation gets strengthened to exploit existing successful patterns from replayed experiences for comparative action-level exploration, accelerating solution iteration without unbounded entropy growth. To further stabilize training, we recalibrate the advantages of experiences in the replay buffer to address the potential policy drift. Reugularizations such as the clipping of tokens with high covariance between probability and advantage are introduced to the trajectory-level entropy control to curb over-confidence.
Authors: Zonghan Yang, Shengjie Wang, Kelin Fu, Wenyang He, Weimin Xiong, Yibo Liu, Yibo Miao, Bofei Gao, Yejie Wang, Yingwei Ma, Yanhao Li, Yue Liu, Zhenxing Hu, Kaitai Zhang, Shuyi Wang, Huarong Chen, Flood Sung, Yang Liu, Yang Gao, Zhilin Yang, Tianyu Liu
Abstract: Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. We argue these paradigms are not mutually exclusive: reasoning-intensive Agentless training induces skill priors, including localization, code edit, and self-reflection that enable efficient and effective SWE-Agent adaptation. In this work, we first curate the Agentless training recipe and present Kimi-Dev, an open-source SWE LLM achieving 60.4\% on SWE-bench Verified, the best among workflow approaches. With additional SFT adaptation on 5k publicly-available trajectories, Kimi-Dev powers SWE-Agents to 48.6\% pass@1, on par with that of Claude 3.5 Sonnet (241022 version). These results show that structured skill priors from Agentless training can bridge workflow and agentic frameworks for transferable coding agents.
Authors: Runyan Tan, Shuang Wu, Phillip Howard
Abstract: Obtaining high-quality outputs from Large Language Models (LLMs) often depends upon the choice of a sampling-based decoding strategy to probabilistically choose the next token at each generation step. While a variety of such sampling methods have been proposed, their performance can be sensitive to the selection of hyperparameters which may require different settings depending upon the generation task and temperature configuration. In this work, we introduce $p$-less sampling: an information-theoretic approach to sampling which dynamically sets a truncation threshold at each decoding step based on the entire token probability distribution. Unlike existing methods, $p$-less sampling has no hyperparameters and consistently produces high-quality outputs as temperature increases. We provide theoretical perspectives on $p$-less sampling to ground our proposed method and conduct experiments to empirically validate its effectiveness across a range of math, logical reasoning, and creative writing tasks. Our results demonstrate how $p$-less sampling consistently outperforms existing sampling approaches while exhibiting much less degradation in text quality at higher temperature values. We further show how $p$-less achieves greater inference-time efficiency than alternative methods through lower average token sampling times and shorter generation lengths, without sacrificing accuracy. Finally, we provide analyses to highlight the benefits of $p$-less through qualitative examples, case studies, and diversity assessments.
Authors: Hyunjun Kim, Sejong Kim
Abstract: We introduce MacroBench, a code-first benchmark that evaluates whether LLMs can synthesize reusable browser-automation programs (macros) from natural-language goals by reading HTML/DOM and emitting Selenium. MacroBench instantiates seven self-hosted sites covering 681 tasks across interaction complexity and targeting difficulty. Our end-to-end protocol validates generated code via static checks, sandboxed execution, and outcome verification (DOM assertions, database snapshots), and includes a safety suite for scraping, spam/abuse, and credential/privacy prompts. Across 2,636 model-task runs, we observe stratified success: GPT-4o-mini (96.8%), GPT-4o (95.3%), Gemini (89.0%), DeepSeek (83.4%). Models handle simple tasks reliably (91.7%) but fail on complex workflows (0.0%), and none meet production-quality coding practices despite functional completion. We release our complete benchmark pipeline, evaluation framework, and experimental results at https://github.com/hyunjun1121/MacroBench to enable reproducible assessment of macro synthesis for web automation.
Authors: Zeyu Zhu, Kevin Qinghong Lin, Mike Zheng Shou
Abstract: Academic presentation videos have become an essential medium for research communication, yet producing them remains highly labor-intensive, often requiring hours of slide design, recording, and editing for a short 2 to 10 minutes video. Unlike natural video, presentation video generation involves distinctive challenges: inputs from research papers, dense multi-modal information (text, figures, tables), and the need to coordinate multiple aligned channels such as slides, subtitles, speech, and human talker. To address these challenges, we introduce Paper2Video, the first benchmark of 101 research papers paired with author-created presentation videos, slides, and speaker metadata. We further design four tailored evaluation metrics--Meta Similarity, PresentArena, PresentQuiz, and IP Memory--to measure how videos convey the paper's information to the audience. Building on this foundation, we propose PaperTalker, the first multi-agent framework for academic presentation video generation. It integrates slide generation with effective layout refinement by a novel effective tree search visual choice, cursor grounding, subtitling, speech synthesis, and talking-head rendering, while parallelizing slide-wise generation for efficiency. Experiments on Paper2Video demonstrate that the presentation videos produced by our approach are more faithful and informative than existing baselines, establishing a practical step toward automated and ready-to-use academic video generation. Our dataset, agent, and code are available at https://github.com/showlab/Paper2Video.