Over-refusal is a critical side effect of safety alignment in large language models (LLMs), causing harmless but superficially risky queries to be mistakenly rejected. Despite the effectiveness of existing mitigation methods, why safety-aligned LLMs may still reject harmless queries under correct safety supervision remains insufficiently understood. To address this gap, we analyze over-refusal through the lens of representation and optimization. We make two observations: 1) the representations of harmful and over-refusal samples are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers; and 2) harmful-sample gradients exhibit high intra-class similarity and concentrate in a low-rank subspace. Moreover, they are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts. This imbalanced gradient coupling can bias over-refusal samples toward refusal, while the reduced apparent representational separation in deeper layers may make it more difficult for the model to maintain a clear boundary between harmful and over-refusal samples. To this end, we propose GOAL (Gradient Orthogonalization with contrAstive Learning), a safety-alignment framework for mitigating over-refusal. Specifically, GOAL consists of two components: 1) to prevent imbalanced gradient coupling from biasing the optimization of over-refusal samples toward refusal, we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients; and 2) to explicitly enhance the separation between harmful and over-refusal representations, we introduce a contrastive learning objective. Extensive experiments across multiple models and benchmarks demonstrate that GOAL effectively mitigates over-refusal while preserving safety performance and general capabilities.
Zhen Yu, Yang Liu, Jianfeng Feng, Xiahai Zhuang, Qingchao Chen
Foundation models hold promise for generalizable analysis of structural brain magnetic resonance imaging (MRI) scans across diverse populations and imaging conditions. However, existing models typically rely on centralized pretraining using data pooled across cohorts, which is constrained by privacy and data-governance requirements. Moreover, larger cohorts can dominate optimization, underutilizing anatomical information from smaller cohorts and limiting generalization. Here we present BrainFedFM, a structural brain MRI foundation model federatively pretrained on 164,707 three-dimensional scans drawn from diverse real-world data distributions and organized across 42 federated sites. BrainFedFM uses dual-priority federated pretraining, in which spatial-priority masking emphasizes informative anatomical regions during local training, while site-priority aggregation adaptively weights contributions during global aggregation. Across 20 downstream datasets spanning 17 classification, regression and segmentation tasks, BrainFedFM achieved the state-of-the-art performance across seven foundation models, including four centralized models, while showing consistent advantages in classification and regression tasks and strong robustness under underrepresented populations and challenging imaging conditions. These findings demonstrate the generalizability of BrainFedFM and highlight federated pretraining as a practical strategy for developing neuroimaging foundation models from distributed data without pooling raw images.
Skills, as reusable external knowledge, have demonstrated substantial potential to enhance the ability of Large Language Models (LLMs) to solve complex tasks. However, existing skill-augmented generation methods typically rely on separate models for skill retrieval and skill utilization, resulting in disconnected optimization of these two processes. To address this gap, we propose Skill-Uni, a unified reinforcement learning framework that jointly optimizes skill retrieval and utilization. Specifically, we adopt a generative skill retrieval paradigm that enables a single model to both learn how to retrieve relevant skills and leverage them for task solving. To effectively train this unified framework, we introduce a two-stage training strategy consisting of warm-up instruction tuning and GRPO, allowing the model to progressively improve its skill retrieval and utilization capabilities. Extensive experiments on five skill-related benchmarks demonstrate the effectiveness of Skill-Uni, yielding average performance gains of 66.59% and 34.82% over the base Qwen3-4B and LLaMA3.1-8B models, respectively. Besides, Skill-Uni is compatible with various skill retrieval engines, demonstrating its flexibility and generalizability. Additional in-depth analyses also provide insights into the effectiveness of Skill-Uni.
Hanwen Cui, Yuanyuan Li, Qing Guo, Zhida Zhang, Chao Chen, Xi Chen, Jie Cao
World Action Models (WAMs) combine future-video prediction with robot action learning, yet improvements in task success alone do not reveal how predictive supervision manifests in execution behavior or whether its benefits persist under task and environmental shifts. We conduct controlled future-video prediction loss ablations within FastWAM and OpenWAM, keeping model architecture, initialization, robot training data, and training steps fixed within each model family. Under clean conditions on seen tasks, predictive supervision increases success rates by 24.38 and 72.91 percentage points, respectively. Trajectory clustering further reveals higher success rates within several coarse trajectory clusters, including cases where cluster frequencies change little, indicating that the gains are not solely reflected in changes in the prevalence of trajectory patterns. However, these benefits do not consistently persist under distribution shifts: improvements on unseen tasks vary across models and tasks, while both models exhibit low success rates under the evaluated environmental randomizations, without a consistent advantage from predictive supervision. We further introduce future-video prediction as an auxiliary training objective for $π_{0.5}$, where a lightweight Transformer predictor used only during training increases success from 82.94% to 89.19%. These findings support future-video prediction as effective auxiliary supervision for action learning, while distinguishing improved execution on familiar tasks from reliable transfer and showing that such gains need not rely on a large pretrained video diffusion transformer (DiT) in the evaluated setting.
Zian Liu, Xiaoxi Meng, Viet Vo, Siqi Ma, Chao Chen, Dongxi Liu, Jun Zhang, Shigang Liu 等 10 位作者
Binary similarity detection is a critical technique that has been applied in many real-world scenarios where source code is not available, e.g., bug search, malware analysis, and code plagiarism detection. Existing works are ineffective in detecting similar binaries in cases where different compiling optimizations, compilers, source code versions, or obfuscation are deployed. We observe that all the cases do not change a binary's key code behaviors although they significantly modify its syntax and structure. With this key observation, we extract a set of key instructions from a binary to capture its key code behaviors. By detecting the similarity between two binaries' key instructions, we can address well the ineffectiveness limitation of existing works. Specifically, we translate each extracted key instruction into a self-defined key expression, generating a key-semantics graph based on the binary's control flow. Each node in the key-semantics graph denotes a key instruction, and the node attribute is the key expression. To quantify the similarity between two given key-semantics graphs, we first serialize each graph into a sequence of key expressions by topological sort. Then, we tokenize and concatenate key expressions to generate token lists. We calculate the locality-sensitive hash value for all token lists and quantify their similarity. %We implement a prototype, called SemDiff, consisting of two modules: graph generation and graph diffing. The first module generates a pair of key-semantics graphs and the second module diffs the graphs. Our evaluation results show that overall, SemDiff outperforms state-of-the-art tools when detecting the similarity of binaries generated from different optimization levels, compilers, and obfuscations. SemDiff is also effective for library version search and finding similar vulnerabilities in firmware.
Long-horizon embodied navigation may be interrupted by a higher-priority instruction before the current task sequence is complete. The agent must finish its current atomic task, execute the inserted task, and resume the remaining sequence from the updated environment state, where its viewpoint and available navigation candidates may have changed. To evaluate this capability, we introduce PIFR benchmark, a novel task for runtime task preemption and resumption in embodied navigation. PIFR includes single, repeated, and nested interruption settings,with stage-specific metrics for inserted-task completion, resumption of the original task, and completion of the full sequence. Evaluations show that recovery performance varies across interruption conditions, while aggregate success can obscure differences in interruption handling and resumption. We further propose a lightweight candidate-ranking baseline to improve target re-grounding after interruption. Together, PIFR and the baseline provide a systematic framework for studying interruption handling and task recovery in long-horizon embodied navigation.
Single-document corpus poisoning under black-box access poses a severe threat to Retrieval-Augmented Generation (RAG). However, existing attacks predominantly rely on explicit factual refutations or imperative instruction overrides. Modern aligned LLMs often arbitrate such conflicts conservatively, rejecting overt contradictions in favor of parametric memory or clean retrieved context while triggering input guardrails. We propose MONOLOGUE, a single-document poisoning framework based on Logic Hijacking without explicit refutation. Rather than disputing the ground truth, our method concedes it as an authoritative historical baseline before introducing the update, weaving a grounded record-correction narrative (e.g., dated institutional reattributions) that presents the counterfactual target as the most recent update. To prevent suspicious live writes, payloads are refined entirely offline via local multi-dimensional gated optimization, coupling automated text quality audits with surrogate behavioral verification. Evaluations across three benchmarks, three retrieval backends, and four open-weight generator regimes show that our method achieves the strongest overall attack efficacy among the evaluated single-document poisoning attacks.
Most agent-memory benchmarks frame long-term memory as retrieval: can an agent recover the relevant record from a long and evolving history? Scientific agents face a stricter problem: a correctly remembered fact may still be invalid to use. The same reported test result can support confirmation when the outcome and stopping rule were fixed in advance, yet only exploration when the outcome was selected from twenty inspected results. We call this scientific use validity: whether remembered evidence is eligible to influence a decision under the state governing its interpretation. We introduce SciValidityBench, an executable benchmark factory with 64 TaskSpecs spanning five validity modules and six scientific domains. TaskSpecs couple generated experimental histories with scientific-state definitions, programmatic decision oracles, controlled interventions, and automatic scoring. The SciStateForge compiler-reader framework turns distributed histories into explicit decision-governing state. Counterfactual pairs and targeted corruption-repair controls then test whether models use those dependencies and localize failures. Across four model families, compiled scientific state improves valid decision-making by 9.2-18.8 percentage points over a generic Event Graph. Retrieval alone does not reproduce this cross-family gain, while content-identical flat and typed views show that the improvement comes from exposing governing state rather than a particular graph serialization. SciValidityBench thus shifts memory evaluation from finding the right fact to deciding whether that fact should control the decision.
Tuochao Chen, Bandhav Veluri Sesame AI, Hongyu Gong FAIR at Meta, Shyamnath Gollakota Department of Computer Science, University of Washington
Dialogue models falter in noisy, multi-speaker environments, often producing irrelevant responses and awkward turn-taking. We present AV-Dialog, the first multimodal dialog framework that uses both audio and visual cues to track the target speaker, predict turn-taking, and generate coherent responses. By combining acoustic tokenization with multi-task, multi-stage training on monadic, synthetic, and real audio-visual dialogue datasets, AV-Dialog achieves robust streaming transcription, semantically grounded turn-boundary detection and accurate responses, resulting in a natural conversational flow. Experiments show that AV-Dialog outperforms audio-only models under interference, reducing transcription errors, improving turn-taking prediction, and enhancing human-rated dialogue quality. These results highlight the power of seeing as well as hearing for speaker-aware interaction, paving the way for spoken dialogue agents that perform robustly in real-world, noisy environments.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Speculative decoding has emerged as a promising paradigm for accelerating large language model inference by leveraging a lightweight draft model to generate multiple candidate tokens. However, existing methods often incur substantial training overhead to mitigate information misalignment between autoregressive draft model training and decoding. To address this challenge, we propose EDSD, an Entropy-Driven Speculative Decoding framework that uses entropy as a unified, interpretable signal for both draft model training and architectural design. EDSD drives the draft model to progressively align with the target model in an easy-to-hard manner while establishing token-level alignment as a dominant design principle. Extensive experiments on seven LLMs demonstrate that EDSD improves training efficiency by 24.8%, increases the average acceptance length by 4.0%, and achieves a 4.1% speedup compared to state-of-the-art methods. Furthermore, EDSD improves robustness to system prompt variations by more than 5x. Our findings establish entropy-driven alignment as an effective and principled foundation for efficient speculative decoding.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Michelle Chao Chen, Moritz Miller, Bernhard Schölkopf ELLIS Institute and Max Planck Institute for Intelligent Systems, Max-Planck Institute, Siyuan Guo Prior Labs
Learning structural information from observational data is central to producing new knowledge outside the training corpus. This holds for mechanistic understanding in scientific discovery as well as flexible test-time compositional generation. We thus study how language models learn abstract structures and utilize the learnt structural information at test-time. To ensure a controlled setup, we design a natural language dataset based on linguistic structural transformations. We empirically show that the emergence of learning structural information correlates with complex reasoning tasks, and that the ability to perform test-time compositional generation remains limited.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表