论文检索

Kuai Yu, Naicheng Yu, Han Wang, Rui Yang, Huan Zhang

Web agents have demonstrated strong performance on a wide range of web-based tasks. However, existing research on the effect of environmental variation has mostly focused on robustness to adversarial attacks, with less attention to agents’ preferences in benign scenarios. Although early studies have examined how textual attributes influence agent behavior, a systematic understanding of how visual attributes shape agent decision-making remains limited. To address this, we introduce VAF, a controlled evaluation pipeline for quantifying how webpage Visual Attribute Factors influence web-agent decision-making. Specifically, VAF consists of three stages: (i) variant generation, which ensures the variants share identical semantics as the original item while only differ in visual attributes; (ii) browsing interaction, where agents navigate the page via scrolling and clicking the interested item, mirroring how human users browse online; (iii) validating through both click action and reasoning from agents, which we use the Target Click Rate and Target Mention Rate to jointly evaluate the effect of visual attributes. By quantitatively measuring the decision-making difference between the original and variant, we identify which visual attributes influence agents’ behavior most. Extensive experiments, across 8 variant families (48 variants total), 5 real-world websites (including shopping, travel, and news browsing), and 4 representative web agents, show that background color contrast, item size, position, and card clarity have a strong influence on agents’ actions, whereas font styling, text color, and item image clarity exhibit minor effects.

Luca Zhou, Pratham Yashwante, Marshall Fisher, Alessio Sampieri, Zihao Zhou, Fabio Galasso, Rose Yu

Time series captioning, the task of describing time series in natural language, requires numeric and temporal reasoning, trend interpretation, and contextual understanding. Existing benchmarks, however, often rely on fully synthetic or generic captions, and typically neglect metadata and visual representations. We introduce CaTS-Bench, a comprehensive benchmark for Context-aware Time Series reasoning across 11 diverse domains, centered on a gold-standard evaluation set of 1746 human-rewritten captions that measure how effectively models translate numeric trends into immediately interpretable narratives. To address the scarcity of human-annotated data, we also propose a scalable pipeline for generating high-fidelity synthetic captions, the quality of which we validate. We evaluate leading Vision-Language Models on our benchmark, revealing that even proprietary models struggle to capture numeric nuances in temporal descriptions, while finetuning open-source models on synthetic data yields substantial performance gains. Finally, we release a diagnostic suite of 910 multiple-choice questions and use tailored numeric metrics to gauge time-series-specific reasoning capabilities, establishing CaTS-Bench as a reliable foundation for grounded, multimodal text generation in numeric domains.

Thennal D K, Chris Biemann, Hans Ole Hatzel

Label projection is an effective technique for cross-lingual transfer, extending span-annotated datasets from a high-resource language to low-resource ones. Most approaches perform label projection as a separate step after machine translation, and prior work that combines the two reports degraded translation quality. We re-evaluate this claim with LabelPigeon, a novel framework that jointly performs translation and label projection via XML tags. We design a direct evaluation scheme for label projection, and find that LabelPigeon outperforms baselines and actively improves translation quality in 11 languages. We further assess translation quality across 203 languages and varying annotation complexity, finding consistent improvement attributed to additional fine-tuning. Finally, across 27 languages and three downstream tasks, we report substantial gains in cross-lingual transfer over comparable work, up to +40.2 F1 on NER. Overall, our results demonstrate that XML-tagged label projection provides effective and efficient label transfer without compromising translation quality.

Wenhao Wang, Yanyan Li, Long Jiao, Jiawei Yuan

Recent advances in large language models (LLMs) provide robots with contextual reasoning abilities to comprehend human instructions. Yet, current LLM-enabled robots typically depend on cloud-based models or high-performance computing infrastructure, which limit their deployment on robots under unreliable internet environments or with constrained computational resources, such as UAVs and small ground vehicles. Thus, deploying fine-tuned small language models (SLMs) that support onboard deployment offers a promising alternative. This paper introduces Ro-SLM, a framework that enables reliable SLM-driven robot operation by distilling LLMs’ knowledge and reasoning. Ro-SLM starts from dataset synthesis by leveraging LLMs to generate diverse task instructions, produce corresponding ground truth code with minimal human assistance, and augment instructions into real-world application scenarios. Ro-SLM is then fine-tuned with the dataset, in which LLM serves as a reward function to guide the training. Extensive experiments on UAV operation tasks demonstrate that Ro-SLM improves the performance of SLM from being incapable of supporting robotic task planning and code generation to achieving performance that approaches LLM.

Dan Shi, Renren Jin, Zhuowen Han, Yuqi Ren, Xinwei Wu, Zhigen Li, Deyi Xiong

Despite recent advances in safety alignment, large language models (LLMs) remain highly susceptible to adversarial attacks, while the internal mechanisms behind such vulnerabilities are still poorly understood. Existing gradient-based attribution methods offer valuable interpretability for analyzing information storage and processing in LLMs. However, they are inapplicable to adversarial attacks, which typically occur in open-ended generation settings without fixed ground-truth outputs. To address these challenges, we propose a novel similarity-based gradient attribution method to identify key neurons sensitive to adversarial behaviors in open-ended generation tasks. The detected neurons, termed targeted neurons, play a critical role in safety training. Building on this neuron-level perspective, we uncover two key neuronal patterns: (i) universal neurons that are consistently exploited across multiple attack strategies, and (ii) interference neurons that hinder safety improvements when fine-tuned indiscriminately, providing mechanistic insights into the interpretability of adversarial vulnerabilities. Inspired by these findings, we propose a neuron-level defense strategy, Targeted Neuron Tuning (TNT), which selectively fine-tunes the identified targeted neurons for specific attacks. Experimental evaluations across multiple LLM architectures and scales demonstrate that TNT substantially improves model robustness against a wide range of jailbreak attacks, achieving safe rates exceeding 90% and even approaching 100%, while preserving general task performance, enabling precise and robust safety interventions. Warning: This paper contains example data that may be harmful.

Xi Chen, Wei Xue, Yike Guo

Role-playing has garnered rising attention as it provides a strong foundation for human-machine interaction and facilitates sociological research. However, current work is confined to textual modalities, neglecting speech, which plays a predominant role in daily life, thus limiting genuine role-playing. To bridge this gap, we conceptualize and benchmark speech role-playing through ActorMindBench, and we present a corresponding reasoning framework, called ActorMind. Specifically, (1) Speech Role-Playing enables models to deliver spontaneous responses with personalized verbal traits based on their role, the scene, and spoken dialogue. (2) ActorMindBench is a hierarchical benchmark comprises Utterance-Level content with 7,653 utterances, Scene-Level content with 313 scenes, and Role-Level content with 6 roles. (3) ActorMind is an off-the-shelf, multi-agent, chain-of-though style reasoning framework that emulates how human actors perform in theaters. Concretely, ActorMind first reads its assigned role description via Eye Agent, then comprehends emotional cues within contextual spoken dialogues through Ear Agent. Subsequently, Brain Agent generates a descriptive emotional state, and finally, Mouth Agent delivers the scripts infused with corresponding emotion state. Experimental results demonstrate the effectiveness of ActorMind in enhancing speech role-playing. The project page is available at https://github.com/OzymandiasChen/ActorMind.

Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei

Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are calculated on the current task, thus they do not guarantee broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training focus each objective should receive. To address this, we propose RECAP—a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts in an online manner using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks based on Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.

Amir Saeidi, Venkatesh Mishra, Souradeep Mukhopadhyay, Gaowen Liu, Ali Payani, Jayanth Srinivasa, Chitta Baral

Large Language Models are being increasingly deployed as the decision-making core of autonomous agents capable of effecting change in external environments. Yet, in conversational benchmarks, which simulate real-world customer-centric issue resolution scenarios, these agents frequently fail due to the cascading effects of incorrect decision-making. These challenges are particularly pronounced for open-source LLMs with smaller parameter sizes, limited context windows, and constrained inference budgets, which contribute to increased error accumulation in agentic settings. To tackle these challenges, we present the **Failure-Aware Meta-Agentic (FAMA)** framework. FAMA operates in two stages: first, it analyzes failure trajectories from baseline agents to identify the most prevalent errors; second, it employs an orchestration mechanism that activates a minimal subset of specialized agents tailored to address these failures by injecting a targeted context for the tool-use agent before the decision-making step. Experiments across open-source LLMs demonstrate performance gains up to **27%** across evaluation modes over standard baselines. These results highlight that targeted curation of context through specialized agents to address common failures is a valuable design principle for building reliable, multi-turn tool-use LLM agents that simulate real-world conversational scenarios.

Zhenheng Tang, Qihua Pan, Jingya Shen, Xiang Liu, Qian Wang, Bo Li, Xiaowen Chu

The value alignment of Large Language Models (LLMs) is critical because value is the foundation of LLM decision-making and behavior. Some recent work show that LLMs have similar value rankings. However, little is known about how susceptible LLM value rankings are to external influence and how different values are correlated with each other. In this work, we investigate the plasticity of LLM value systems by examining how their value rankings are influenced by different prompting strategies and exploring the intrinsic relationships between values. To this end, we design 6 different value transformation prompting methods including direct instruction, rubrics, in-context learning, scenario, persuasion, and persona, and benchmark the effectiveness of these methods on 3 different families and totally 8 LLMs. Our main findings include that the value rankings in large LLMs are much more susceptible to external influence than small LLMs, and there are intrinsic correlations between certain values (e.g., Privacy and Respect). Besides, through detailed correlation analysis, we find that the value correlations are more similar between large LLMs of different families than small LLMs of the same family. We also identify that scenario method is the strongest persuader and can help entrench the value rankings.

Wenxuan Wang, Yuk-Kit Chan, Zixuan Ling, Shi Juluan, Youliang Yuan, Jen-tse Huang, Yifei Zhang, Wenxiang Jiao, Zhaopeng Tu, Michael R. Lyu

Large Language Models (LLMs) like ChatGPT are foundational in various applications due to their extensive knowledge from pre-training and fine-tuning. Despite this, they are prone to generating factual and commonsense errors, raising concerns in critical areas like healthcare, journalism, and education to mislead users. Current methods for evaluating LLMs’ veracity are limited by the need for extensive human labor, test data contamination, or limited scope, hindering efficient and effective exposure of errors. To address these challenges, we propose HalluHunter, a novel, fully automated framework for systematically uncovering factual inaccuracies in LLMs. HalluHunter employs a knowledge-graph-based approach, extracting fact triplets to generate diverse question types for single- and multi-hop reasoning using rule-based Natural Language Processing (NLP) techniques. Its iterative process starts with random triplet selection for question generation, followed by adaptive selection in subsequent iterations, targeting triplets where LLMs frequently err based on their performance analysis. Our extensive tests on nine prominent LLMs reveal that HalluHunter can trigger factual errors in up to 55% of questions in these models. Moreover, we demonstrate that HalluHunter’s test cases, particularly in adaptive selection, could further expose the weaknesses in benchmarking the factuality in LLMs meanwhile maintaining the coverage of questions. All code, data, and results will be released for future research.

Rian Touchent, Nathan Godey, Éric Villemonte de la Clergerie

We annotate PubMed Central paragraphs for document type, domain, and educational quality using a two-stage pipeline: Llama-3.1-70B labels 400K paragraphs, then a fine-tuned XLM-RoBERTa propagates annotations to the full corpus. This paragraph-level approach captures content diversity within scientific articles that document-level labels miss. The resulting Biomed-Enriched corpus contains 2M clinical case paragraphs, providing a publicly available alternative to restricted clinical datasets. For decoders, continual pretraining experiments enable targeted improvements, with clinical upsampling boosting performance by 4 points on MMLU ProfMed and educational filtering improving MedQA and MedMCQA by ~1 point. Combinations of these techniques led to faster convergence, reaching the same performance with a third of training tokens. For encoders, our best recipe matches BioClinical-ModernBERT on 11 tasks (77.3% vs 77.1% F1) while using 2.5x fewer tokens and only public data.

Chinonso Cynthia Osuji, Simon Mille, Mark Andrade, Jane Adkins, Ornait O’Connell, Elaine Uí Dhonnchadha, Bláithín Heffernan, Fírinne Nic an tSaoir, Anya Belz, Thiago Castro Ferreira 等

Generating coherent, semantically accurate text from large structured inputs remains a persistent challenge in data-to-text generation, as single-step LLM mappings from data-to-text limit control over discourse structuring and amplify hallucinations and omissions as input size grows. We introduce a new dataset of extended DBpedia triple sets (up to 199 triples per input), and a modular multi-agent framework: specialised LLM agents handle content ordering, text structuring, and surface realisation under the supervision of an orchestrator and guardrail control loop. The system generates multi-paragraph outputs in English and Irish (low-resource). We compare a three-worker multi-agent configuration against a single-worker multi-task variant and a strong end-to-end baseline. Quality is assessed via human evaluation and LLM-as-a-judge (with truncation-based sanity checks). Results show slightly superior coherence for the multi-agent approach in both languages, with statistically significant inter-rater correlation over all criteria for English and no statistically significant correlation for Irish. Human-LLM alignment is very weak overall, thus exposing key limits in scalable NLG evaluation.

Shilin Tang, Zunyi Yin, Xuefeng Liang, Guanghui Shi, Song Tong, Chen Guangyu

LLMs have shown promise in mental health counseling, but existing models are limited to surface-level empathy or predefined therapeutic procedures and lack the ability to actively explore the root causes of psychological distress. Inspired by case conceptualization, we formalize counseling as the online reconstruction of a client’s underlying causal graph through multi-turn dialogue. To this end, we propose TRACE, a two-phase reinforcement learning framework. It implements a causal-graph-driven reward scheme across two phases: an exploration phase that rewards the causal graph reconstruction following a surface-to-deep path, and an intervention phase that rewards targeted restructuring of irrational beliefs. Extensive experiments show that TRACE outperforms existing models, enabling causal-chain-aware psychological intervention beyond surface-level empathy.

Yaya SY, Dioula Doucouré, Christophe Cerisara, Irina Illina

Speakers of under-represented languages face both a language barrier, as most online knowledge is in a few dominant languages, and a modality barrier, since information is largely text-based while many languages are primarily oral. We address this for French-Wolof by training the first bilingual speech-text Matryoshka embedding model, enabling efficient retrieval of French text from Wolof speech queries without relying on a costly ASR-translation pipelines. We introduce large-scale data curation pipelines and new benchmarks, compare modeling strategies, and show that modality fusion within a frozen text Matryoshka model performs best. Although trained only for retrieval, the model generalizes well to other tasks, such as speech intent detection, indicating the learning of general semantic representations. Finally, we analyze cost-accuracy trade-offs across Matryoshka dimensions and ranks, showing that information is concentrated only in a few components, suggesting potential for efficiency improvements.

Tianning Chai, Chancharik Mitra, Brandon Huang, Gautam Rajendrakumar Gare, Zhiqiu Lin, Assaf Arbelle, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Deva Ramanan 等

Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is crucial for improving their real-world behavior. A common approach is to use reward models that enable reinforcement-learning post-training. However, traditional reward modeling requires finetuning on large preference datasets, limiting adaptability to new preferences. We introduce Activation Reward Models (Activation RMs)—the first mechanistic interpretability approach that steers LLM activations to align with few-shot preference data without finetuning. Our method combines activation denoising and output token likelihood scoring, achieving state-of-the-art performance on standard reward modeling benchmarks, surpassing zero-shot, few-shot, and voting-based baselines. We further demonstrate that Activation RMs mitigate reward hacking behaviors and remain robust to noisy exemplars and spurious reward signals. To evaluate this, we propose PreferenceHack, a novel few-shot benchmark testing reward models on reward hacking in a paired preference format, where Activation RMs achieve state-of-the-art performance, surpassing GPT-4o.

Jialu Gao, Hua Li, Tracy Ortman, Yeyun Gong, Jian Jiao

Large language models (LLMs) are increasingly deployed in monetization-driven systems such as search engines, advertising platforms, and e-commerce services, where decision making is shaped by complex interactions among user intent, advertiser objectives, and platform constraints. Despite rapid progress, existing benchmarks primarily focus on shopping-centric scenarios and user-facing data, capturing only a limited subset of real-world monetization pipelines and overlooking intermediate decision stages and robustness considerations. In this work, we introduce MonBench, a high-quality multi-task benchmark designed to evaluate LLMs in realistic monetization contexts. The benchmark is constructed from large-scale production data collected from multiple search engines, including both intermediate candidate pools and user-visible outcomes, better reflecting the distributional characteristics of real monetization systems. MonBench covers key capability dimensions such as intent understanding, commercial matching, and user behavior modeling, and adopts a unified multiple-choice formulation to enable systematic comparison across models. We further propose a comprehensive evaluation protocol that measures both performance and robustness. We evaluate a diverse set of state-of-the-art LLMs and conduct detailed task-level analyses. Our results reveal monetization-specific behaviors, including gaps between relevance optimization and broader decision-making capabilities, as well as differences in robustness across model families. These findings provide new insights into the strengths and limitations of current LLMs and highlight the need for richer domain-specific supervision in monetization-oriented applications.

Louis Estève, Christophe Servan, Thomas Lavergne, Agata Savary

Diversity has been gaining interest in the NLP community in recent years. At the same time, state-of-the-art transformer models such as ModernBERT use very large pre-training datasets, which are driven by size rather than by diversity. This summons to investigate theimpact of diversity on pre-training. We do so in this study, with the express intent of reducing pre-training dataset size, while retaining atleast comparable performance. We compare diversity-driven sampling algorithms, and we use the best one to pre-train several ModernBERT models on French with a fixed compute budget. We fine-tune and evaluate them on a variety of French benchmarks. We compare them with models pre-trained on randomly sampled data of commensurate size, with the same compute budget. We find that both random and diversity-driven sampling may reduce the pre-training dataset by up to 94% and the pre-training time by up to 73% while maintaining performance. Moreover, in some tasks, the inherent quality of models, estimated via head-only fine-tuning, is up to 10 points higher with diversity sampling than with random sampling.

Koki Ryu, Hitomi Yanaka

Personalized image aesthetics assessment (PIAA) is an important research problem with practical real-world applications. While methods based on vision-language models (VLMs) are promising candidates for PIAA, it remains unclear whether they internally encode rich, multi-level aesthetic attributes required for effective personalization. In this paper, we first analyze the internal representations of VLMs to examine the presence and distribution of such aesthetic attributes, and then leverage them for lightweight, individual-level personalization without model fine-tuning. Our analysis reveals that VLMs encode diverse aesthetic attributes that propagate into the language decoder layers. Building on these representations, we demonstrate that simple linear models can achieve effective personalized image aesthetics assessment. We further analyze how aesthetic information is transferred across layers in different VLM architectures and across image domains. Our findings provide insights into how VLMs can be utilized for modeling subjective, individual aesthetic preferences.

Jiawei Li, Akshayaa Magesh, Venugopal Veeravalli

While Large Language Models (LLMs) have emerged as powerful foundational models to solve a variety of tasks, they have also been shown to be prone to hallucinations, i.e., generating responses that sound confident but are actually incorrect or even nonsensical. Existing hallucination detectors propose a wide range of empirical scoring rules, but their performance varies across models and datasets, and it is hard to determine which ones to rely on in practice or to treat as a reliable detector. In this work, we formulate the problem of detecting hallucinations as a hypothesis testing problem and draw parallels with the problem of out-of-distribution detection in machine learning models. We then propose a multiple-testing-inspired method that systematically aggregates multiple evaluation scores via conformal p-values, enabling calibrated detection with controlled false alarm rate. Extensive experiments across diverse models and datasets validate the robustness of our approach against state-of-the-art methods.

Francielle Vargas, Jackson Trager, Diego Alves, Matteo Guida, Surendrabikram Thapa, Berk Atıl, Daryna Dementieva, Andrew J Smart, Ameeta Agrawal

Existing hate speech detection models are often opaque and rely on surface-level lexical cues, which makes them vulnerable to spurious correlations and limits robustness, interpretability and cultural contextualization. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance while enhancing both faithful and plausible explanations. Although explanations become more concise, sufficiency decreases, indicating more compact and informative rationales. Fairness remains stable, suggesting that improvements in explanation quality do not introduce significant bias trade-offs.