Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capabilities of Large Language Models. When applied to RLVR, Multiple-Choice Questions (MCQs) offer a scalable source of verifiable data but risk inducing reward hacking, where models shortcut reasoning via random guessing or simple elimination. Current approaches often mitigate this by converting MCQs to open-ended formats, thereby discarding the contrastive signal provided by expert-designed distractors. In this work, we systematically investigate the impact of option design on RLVR. Our analysis highlights two primary insights: (1) Mismatches in option counts between training and testing degrade performance. (2) Strong distractors effectively mitigate random guessing, enabling effective RLVR training even with 2-way questions. Motivated by these findings, we propose Iterative Distractor Curation (IDC), a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning. Experiments on various benchmarks demonstrate that our method effectively enhances distractor quality and yields significant gains in RLVR training compared to the original data.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities. However, the enhancement of reasoning abilities and the exposure of internal reasoning processes introduce new safety vulnerabilities. A critical question arises: when reasoning becomes intertwined with harmfulness, will LRMs become more vulnerable to jailbreaks in reasoning mode? To investigate this, we introduce HauntAttack, a novel and general-purpose black-box adversarial attack framework that systematically embeds harmful instructions into reasoning questions. Specifically, we modify key reasoning conditions in existing questions with harmful instructions, thereby constructing a reasoning pathway that guides the model step by step toward unsafe outputs. We evaluate HauntAttack on 11 LRMs and observe an average attack success rate of over 70%, achieving up to 13 percentage points of absolute improvement over the strongest prior baseline. Our further analysis reveals that even advanced safety-aligned models remain highly susceptible to reasoning-based attacks, offering insights into the urgent challenge of balancing reasoning capability and safety in future model development.
EchoMLLM: Incentivizing Echocardiographic Video Understanding with Keyframe Grounding and Report Generation
PDF ↗Echocardiography analysis demands a dual capability: rigorous quantitative keyframe localization for evidence verification and comprehensive qualitative synthesis for diagnostic reporting. However, current Multi-Modal Large Language Models (MLLMs) struggle to meet these clinical requirements due to a misalignment with diagnostic workflows, a scarcity of video instruction data, and the critical challenge of cyclic temporal ambiguity—where the repetitive nature of cardiac cycles renders standard single-frame supervision ill-posed. To bridge this gap, we introduce EchoMLLM, a unified framework designed for real-world echocardiography video understanding. First, we align model capabilities with clinical needs by defining two fine-grained tasks: cycle- and pathology-conditioned keyframe grounding and video report generation. To facilitate this, we curate EchoMM-120k, a large-scale instruction dataset specifically constructed to support temporal localization and professional reporting. Furthermore, to resolve the cyclic ambiguity, we propose a multi-stage training paradigm incorporating a novel cycle-aware Reinforcement Learning (RL) strategy. By prioritizing logical consistency over rigid index matching, our approach moves beyond rote memorization to elicit invariant reasoning. Extensive experiments demonstrate that EchoMLLM reduces temporal grounding errors by up to 76% and improves report generation quality by 65% over its backbone, achieving state-of-the-art performance against both generalist and medical baselines.
Towards IP Intelligence: Benchmarking Large Language Models on Intellectual Property Knowledge and Practice
PDF ↗Intellectual Property (IP) is a highly specialized domain that integrates technical and legal knowledge, making it inherently complex and knowledge-intensive. Recent advancements in LLMs have demonstrated their potential to handle IP tasks, enabling more efficient analysis, understanding, and generation of IP-related content. However, existing datasets and benchmarks focus narrowly on patents or cover limited aspects of the IP field, lacking alignment with real-world scenarios. To bridge this gap, we introduce **IPBench**, the first comprehensive IP task taxonomy and a large-scale bilingual benchmark encompassing **8 IP mechanisms and 20 distinct tasks**, designed to evaluate LLMs in real-world IP practice. We benchmark **19 main LLMs**, ranging from general purpose to domain-specific, including chat-oriented and reasoning-focused models, under zero-shot, few-shot, and chain-of-thought settings. Our results show that even the top-performing model, DeepSeek-V3, achieves only 75.8% accuracy, indicating significant room for improvement. Notably, open-source IP and law-oriented models lag behind closed-source general-purpose models. To foster future research, we publicly release IPBench, and will expand it with additional tasks to better reflect real-world complexities and support model advancements in the IP domain. We provide the data, code in the supplementary materials.
Recent advancements in Large Language Models (LLMs) have empowered home assistants with natural language interaction capabilities. However, current assistants overlook the progressive omission that occurs in human dialogue as shared context accumulates, leading to more elliptical expressions for efficient communication. Thus, current assistants still struggle to interpret such elliptical expressions accurately, which limits their effectiveness in real-world applications. In practical smart home scenarios, assistants face two major challenges caused by elliptical commands: (1) referential ambiguity caused by different environmental expectations among multiple users; and (2) intention ambiguity resulting from user preferences that evolve over time or change with the environment. To address these challenges, we introduce PEC-Home, the first simulated home dataset specifically designed for interpreting progressively elliptical commands in smart homes. Extensive experiments on various LLMs, including GPT-4o, show that existing home assistants struggle to execute user-intended operations based solely on elliptical commands. Even when equipped with tools for storing and retrieving user dialogue history, execution accuracy remains below that achieved with complete commands. Our code and dataset will be publicly available.
Autoregressive decoding in Large Language Models (LLMs) generates one token per step, causing high inference latency. Speculative decoding (SD) mitigates this through a guess-and-verify strategy, but existing training-free variants face trade-offs: retrieval-based drafts break when no exact match exists, while logits-based drafts lack structural guidance. We propose \textbf{RACER} (\textbf{R}etrieval-\textbf{A}ugmented \textbf{C}ont\textbf{e}xtual \textbf{R}apid Speculative Decoding), a lightweight and training-free method that integrates retrieved exact patterns with logit-driven future cues. This unification supplies both reliable anchors and flexible extrapolation, yielding richer speculative drafts. Experiments on Spec-Bench, HumanEval, and MGSM-ZH demonstrate that RACER consistently accelerates inference, achieving more than 2\times speedup over autoregressive decoding, and outperforms prior training-free methods, offering a scalable, plug-and-play solution for efficient LLM decoding. Our source code is available at https://github.com/hkr04/RACER.
Automated Essay Scoring (AES) faces significant challenges in cross-prompt settings, where models must generalize to unseen writing prompts. To address this limitation, we propose MAPLE, a meta-learning framework that leverages prototypical networks to learn transferable representations across different writing prompts. Across three diverse datasets (ELLIPSE and ASAP (English), and LAILA (Arabic)), MAPLE achieves state-of-the-art performance on ELLIPSE and LAILA, outperforming strong baselines by 8.5 and 3 points in QWK, respectively. On ASAP, where prompts exhibit heterogeneous score ranges, MAPLE yields improvements on several traits, highlighting the strengths of our approach in unified scoring settings. Overall, our results demonstrate the potential of meta-learning for building robust cross-prompt AES systems.
HIPO: A Hierarchical Prompt Optimization Framework with Task Awareness and Fine-Grained Debugging
PDF ↗Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks. However, their performance often hinges on carefully designed prompts, whose creation requires substantial human effort. While numerous automatic prompt optimization techniques have been proposed, existing methods typically apply the same prompt across all samples within a dataset, ignoring variation in sample difficulty. To address these limitations, we propose HIPO, a HIerarchical Prompt Optimization framework that shifts the paradigm from dataset-level to sample-level optimization. Our framework first employs a lightweight router model, trained offline, to predict the difficulty of each sample at test time. Based on this prediction, HIPO dynamically selects a prompt from a five-tiered hierarchy, tailoring complexity to sample difficulty. Furthermore, two refinement stages—Task Description Prompt Refine and Attribution-Based Prompt Refine—enhance generalizability and fine-grained optimization. Extensive experiments on 27 tasks demonstrate that HIPO outperforms all baselines, achieving state-of-the-art performance on 25% more tasks than the strongest baseline. Cost analysis further demonstrates substantial efficiency gains, reducing API calls, token consumption, and overall cost by 1.2× to 80×. Our implementation is publicly available at https://github.com/LuQiCode/HIPO.
SPD-Faith Bench: Diagnosing and Improving Faithfulness in Chain-of-Thought for Multimodal Large Language Models
PDF ↗Chain-of-Thought reasoning is widely used to improve the interpretability of multimodal large language models (MLLMs), yet the faithfulness of the generated reasoning traces remains unclear. Prior work has mainly focused on perceptual hallucinations, leaving reasoning level unfaithfulness underexplored. To isolate faithfulness from linguistic priors, we introduce SPD-Faith Bench, a diagnostic benchmark based on fine-grained image difference reasoning that enforces explicit visual comparison. Evaluations on state-of-the-art MLLMs reveal two systematic failure modes, perceptual blindness and perception-reasoning dissociation. We trace these failures to decaying visual attention and representation shifts in the residual stream. Guided by this analysis, we propose SAGE, a train-free visual evidence-calibrated framework that improves visual routing and aligns reasoning with perception. Our results highlight the importance of explicitly evaluating faithfulness beyond response correctness. Our benchmark and codes are available at https://anonymous.4open.science/r/SPD-Faith/.
Large Language Models (LLMs) are often augmented with Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) prompting, yet static “always-on” use is computationally wasteful. Existing adaptive methods typically optimize a single axis, overlooking that evidence need and reasoning depth are only partially correlated. We present , a dual-axis routing framework that separates retrieval necessity from reasoning necessity under a user-defined cost–quality trade-off. Offline, profiles four pipelines (Direct, RAG, CoT, RAG+CoT) and derives supervision by selecting the utility-maximizing strategy that trades answer quality against token usage and latency. Online, a compact dual-head router, conditioned on cost weights, uses lightweight probes—retrieval-score dispersion (NQC) and single-pass draft negative log-likelihood (NLL)—to decide whether to invoke RAG and/or CoT without sampling or model internals. Across six QA benchmarks, reduces token usage by up to 86% and latency by up to 84% while improving answer quality over strong baselines.
Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.
Scene Graph Generation (SGG) structures visual scenes as graphs of objects and their relations. While Multimodal Large Language Models (MLLMs) have advanced end-to-end SGG, current methods are hindered by both a lack of task-specific structured reasoning and the challenges of sparse, long-tailed relation distributions, resulting in incomplete scene graphs characterized by low recall and biased predictions. To address these issues, we introduce SGG-R^{\rm 3}, a structured reasoning framework that integrates task-specific Chain-of-Thought (CoT)-guided supervised fine-tuning (SFT) and reinforcement learning (RL) with group sequence policy optimization (GSPO), designed to engage in three sequential stages to achieve end-to-end unbiased scene graph generation. During the SFT phase, we propose a relation augmentation strategy by leveraging an MLLM and refined via embedding similarity filtering to alleviate relation sparsity. Subsequently, a stage-aligned reward scheme optimizes the procedural reasoning during RL. Specifically, we propose a novel dual-granularity reward which integrates fine-grained and coarse-grained relation rewards, simultaneously mitigating the long-tail issue via frequency-based adaptive weighting of predicates and improving relation coverage through semantic clustering. Experiments on two benchmarks show that SGG-R^{\rm 3} achieves superior performance compared to existing methods, demonstrating the effectiveness and generalization of the framework.
While the massive scale of modern LLMs enables remarkable performance, their static, input-agnostic computational graph incurs substantial resource wastage and high latency during inference. Existing dynamic schemes, such as early-exit and layer-drop reduce FLOPs but break batch processing or introduce KV-cache inconsistency. We propose Deputy, a dynamic low-rank substitution framework that employs a lightweight decision module at each layer to dynamically determine the execution branch for different tokens: Attention layers choose between full and low-rank computation to mitigate the KV cache issue, while FFN layers additionally support skipping to further reduce computation. We fine-tune the LLM with LoRA and then derive an additional low-rank matrix C via a least-squares fit \textbf{BC} \approx \textbf{W}_{\text{pre}}, where B is the shared LoRA matrix, so that only one extra low-rank matrix is introduced, effectively reducing memory overhead. Moreover, a hybrid KV cache strategy stores KV values generated by the low-rank branch, achieving a 38% reduction in cache storage. Experiments on Llama models demonstrate that Deputy reduces computation by approximately 40% compared to the original dense model while outperforming existing baseline methods.
Rethinking LLM Watermark Detection in Black-Box Settings: A Non-Intrusive Third-Party Framework
PDF ↗While watermarking serves as a critical mechanism for LLM provenance, existing secret-key schemes tightly couple detection with injection, requiring access to keys or provider-side scheme-specific detectors for verification. This dependency creates a fundamental barrier for real-world governance, as independent auditing becomes impossible without compromising model security or relying on the opaque claims of service providers. To resolve this dilemma, we introduce TTP-Detect, a pioneering black-box framework designed for non-intrusive, third-party watermark verification. By decoupling detection from injection, TTP-Detect reframes verification as a relative hypothesis testing problem. It employs a proxy model to amplify watermark-relevant signals and a suite of complementary relative measurements to assess the alignment of the query text with watermarked distributions. Extensive experiments across representative watermarking schemes, datasets and models demonstrate that TTP-Detect achieves superior detection performance and robustness against diverse attacks.
Large Language Models (LLMs) excel at diverse tasks, but remain vulnerable to malicious inputs such as jailbreak attacks. Current one-size-fits-all safety guardrails built from static datasets ignore each model’s unique safety profile and often force trade-offs between safety and utility. To address this gap, we propose LS-Guard, a framework for learning model-specific guardrails tailored to each LLM’s vulnerabilities. LS-Guard operates in two stages: First, it dynamically profiles a given LLM by probing it with malicious prompts to elicit the model’s responses, which are then dynamically labeled to reveal model-specific failure modes. Second, it uses this data to train a safety classifier with a collaborative multi-LoRA architecture. An orthogonality-constrained multi-task loss enables a central expert to learn general safety features while each subject-specific expert encodes the distinctive vulnerability patterns of one LLM. During inference, LS-Guard activates the central expert together with its model-specific expert to perform content moderation, yielding reliable safety decisions. Extensive experiments on multiple real-world LLMs demonstrate that LS-Guard significantly outperforms strong baseline guardrails, achieving superior robustness, adaptability, and generalization.
Deep Research agents powered by Large Language Models (LLMs) have exhibited extraordinary potential in automated paper writing tasks. However, existing systems rely heavily on literature retrieval and synthesis through internet and local knowledge bases, often resulting research lacking insight and creativity in social science. To address this issue, we propose "Memory-Augmented Social Simulation (MASS)”, an innovative paradigm that leverages highly realistic and research-oriented social simulations to the creativity and empirical founding of LLMs-generated research. Specifically, MASS integrates three core components—dynamic goal-path planning with multi-level social norm restraint to guide the simulation, a multi-disciplinary behavior dataset for agent memory cold-start, and a structured forgetting mechanism inspired by the Ebbinghaus curve. Together, these ensure simulation authenticity and provide a robust empirical foundation for generating innovative scholarly papers. Experimental results demonstrate the effectiveness of our method, showing a 6.81% improvement in generation overall quality over foundation LLMs and 17.19% gain in Insight over strong baselines. Dataset and codes will be released.
CodeBind: Decoupled Representation Learning for Multimodal Alignment with Unified Compositional Codebook
PDF ↗Multimodal representation alignment is pivotal for large language models and robotics. Traditional methods are often hindered by cross-modal information discrepancies and data scarcity, leading to suboptimal alignment spaces that overlook modality-unique features. We propose CodeBind, a framework that optimizes multimodal representation spaces through a modality-shared-specific codebook design. By incrementally aligning target and bridging modalities, CodeBind bypasses the need for fully paired data. Unlike traditional hard alignment, CodeBind decomposes features into shared components for semantic consistency and specific components for modality-unique details. This design utilizes a compositional vector quantization scheme, where a shared codebook bridges modality gaps and modality-specific codebooks mitigate representation bias by preventing dominant modalities from overshadowing others. Validated across nine modalities (text, image, video, audio, depth, thermal, tactile, 3D point cloud, EEG), CodeBind achieves state-of-the-art performance in multimodal classification and retrieval tasks. Project page: https://visual-ai.github.io/codebind
Large language models (LLMs) are increasingly used as conversational partners for learning, yet the interactional dynamics supporting users’ learning and engagement are understudied. We analyze the linguistic and interactional features from both LLM and participant chats across 397 human–LLM conversations about socio-political issues to identify the mechanisms and conditions under which LLM explanations shape changes in political knowledge and confidence. Mediation analyses reveal that LLM explanatory richness partially supports confidence by fostering users’ reflective insight, whereas its effect on knowledge gain operates entirely through users’ cognitive engagement. Moderation analyses show that these effects are highly conditional and vary by political efficacy. Confidence gains depend on how high-efficacy users experience and resolve uncertainty. Knowledge gains depend on high-efficacy users’ ability to leverage extended interaction, with longer conversations benefiting primarily reflective users. In summary, we find that learning from LLMs is an interactional achievement, not a uniform outcome of better explanations. The findings underscore the importance of aligning LLM explanatory behavior with users’ engagement states to support effective learning in designing Human-AI interactive systems.
Reinforcement fine-tuning improves the reasoning ability of large language models, but it can also encourage them to answer unanswerable queries by guessing or hallucinating missing information. Existing abstention methods either train models to produce generic refusals or encourage follow-up clarifications without verifying whether those clarifications identify the key missing information. We study queries that are clear in meaning but cannot be reliably resolved from the given information, and argue that a reliable model should not only abstain, but also explain what is missing. We propose a clarification-aware RLVR reward that, while rewarding correct answers on answerable queries, jointly optimizes explicit abstention and semantically aligned post-refusal clarification on unanswerable queries. Using this reward, we train Abstain-R1, a 3B model that improves abstention and clarification on unanswerable queries while preserving strong performance on answerable ones. Experiments on Abstain-Test, Abstain-QA, and SelfAware show that Abstain-R1 substantially improves over its base model and achieves unanswerable-query behavior competitive with larger systems including DeepSeek-R1, suggesting that calibrated abstention and clarification can be learned through verifiable rewards rather than emerging from scale alone.
Unleashing the Native Recommendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers
PDF ↗Leveraging the vast open-world knowledge and understanding capabilities of Large Language Models (LLMs) to develop general-purpose, semantically-aware recommender systems has emerged as a pivotal research direction in generative recommendation. However, existing methods face bottlenecks in constructing item identifiers. Text-based methods introduce LLMs’ vast output space, leading to hallucination, while methods based on Semantic IDs (SIDs) encounter a semantic gap between SIDs and LLMs’ native vocabulary, requiring costly vocabulary expansion and alignment training. To address this, this paper introduces Term IDs (TIDs), defined as a set of semantically rich and standardized textual keywords, to serve as robust item identifiers. We propose GRAM, a novel framework centered on TIDs, employs Context-aware Term Generation to convert item’s metadata into standardized TIDs and utilizes Integrative Instruction Fine-tuning to collaboratively optimize term internalization and sequential recommendation. Additionally, Elastic Identifier Grounding is designed for robust item mapping. Extensive experiments on real-world datasets demonstrate that GRAM significantly outperforms baselines across multiple scenarios, pointing a promising direction for generalizable and high-performance generative recommendation systems.