AI 速览：提出首个 token 级自适应在线策略自蒸馏方法 MTA-OPSD，让每个 token 在训练中从多教师池中选择偏好的教师。
On-policy Self-distillation (OPSD) has proven effective for post-training large language models by constructing a teacher conditioned on privileged references. However, existing approaches keep the teacher fixed either throughout training by consistently exposing it to the full reference, or within each response by adaptively varying partial reference exposure only across responses. In this paper, we argue that neither of them is sufficient. Instead, every token needs its own teacher, as the teacher preference can vary token by token within the same response. Through a teacher-intervention pilot study, we validate that the teacher preference differs across tokens, and reveal that a preferred teacher can be identified for each token through majority voting and KL-based selection. Motivated by these findings, we propose Multi-teacher Adaptive On-policy Self-Distillation (MTA-OPSD), the first token-level adaptive OPSD method. Our method constructs a multi-teacher pool consisting of candidate teachers, and enables every token to adaptively select its preferred teacher during training. Experiments with Qwen3 models on mathematical reasoning tasks show that MTA-OPSD consistently outperforms existing OPSD approaches with an average gain of up to $+5.9$ over the base model, highlighting the importance of adapting teacher supervision at the token level. The code is available at https://github.com/xingzhejun/MTA-OPSD.
Hongqi Chen, Chengwang Liao, Ziwei Mei, Zhentao Shi
LASSO is one of the most popular regularization methods for high-dimensional regression models. While cointegration is common in practical applications of predictive regressions with many regressors, the convex penalty of LASSO overly suppresses the coefficients of cointegrated regressors and therefore fails to deliver consistent parameter estimation. We employ folded concave penalties (FCP), including the well-known SCAD and MCP, to address this issue. Using FCP, we establish two asymptotic results: (i) a local oracle property for the penalized regression, and (ii) a global oracle property for regressions with a penalty-level constraint. In particular, the latter guarantees a global solution that consistently estimates the true coefficients and enjoys oracle prediction performance. Simulation studies and an application to predicting the euro area unemployment rate show that FCP improves variable selection and out-of-sample prediction relative to LASSO.
Analogical reasoning is widely considered a fundamental component of human intelligence. Few benchmarks are designed to explicitly test far analogical reasoning - reasoning between domains that are superficially dissimilar yet share deep relational or structural similarities. In this work, we introduce an analogical reasoning benchmark called Abstraction and Reasoning Corpus of ANAlogy (ARCANA). The tasks of ARCANA require performing far analogical mappings from real-world objects or physical phenomena to abstract grid-based transformations, different from the well-known ARC-AGI benchmark which primarily emphasized innate human priors. We validate ARCANA through a human study, demonstrating that humans solve these tasks reliably. In contrast, evaluations across a range of large language models show a distinct performance gap compared to humans, revealing systematic failure modes that remain challenging for current models. To address this gap, we propose ConceptSIR, a reasoning framework based on concept slippage, introspection, and reflection. Experiments demonstrate that ConceptSIR enables standard LLMs to solve several challenging ARCANA tasks that were unsolved by direct prompting in our evaluation. Overall, our results suggest that far analogical reasoning remains an open problem for contemporary LLMs. All the materials are available at sites.google.com/view/arcanacogtest. The complete codebase, including benchmark, and ConceptSIR, is publicly available at https://github.com/density-11/arcana-test. All ARCANA data and task instances can be accessed and explored at https://arcanacogtest.org/.
Zhefan Rao, Meng Chu, Xuan Billy Zhang, Jingfan Chen, Guohao Cai, Runyi Li, Shiyi Du, Xu Huang 等 92 位作者
World Models (WMs) can support Multimodal Agents (MAs) through simulated experience, while Agent trajectories can inform World Model refinement. Evaluating this interaction requires separate evidence about action selection and state evolution. We introduce Dual Exam, which tests Agents under fixed environment response mechanisms and World Models under fixed action interventions across physical, digital, social, and scientific domains. The Agent track compares 10 configurations on 1,088 items in 15 reporting categories. The World Model track contains 400 cases across eight scenarios; 15 configurations have full coverage and one covers six scenarios. GPT 6 Astra leads the case-weighted Agent Index at 31.3, and GPT-Image-2 leads the full-coverage WM Overall at 21.7. These scores use different tasks and metrics and are not directly comparable. A 150-case pilot of one coupled pairing records zero task successes and motivates component-level diagnosis. Results reveal uneven task performance and failures to preserve task-relevant state. Their value for predicting coupled success or sustained reciprocal learning remains to be established.
Wan Jiang, Jingqi Cheng, Hui Zhang, Xu Zhao, Xiaojing Chen, Shengeng Tang, Jingjing Wu, Richang Hong
Web-enabled digital health services are expanding access to physiological monitoring and mental health screening. However, electroencephalography (EEG)-based major depressive disorder (MDD) detection remains challenging across heterogeneous data sources, where individual variability and acquisition differences may affect disease-related representations. In this work, we investigate the use of healthy normative references for EEG-based MDD detection. Specifically, structural and distributional reference models are independently trained using healthy control (HC) recordings to characterize departures from healthy EEG patterns. We develop two diagnostic approaches that incorporate normative deviations at different representation levels. Scalar-level Normative Deviation-guided Diagnosis (S-NDD) summarizes structural and distributional deviations into scalar scores and integrates them with deep EEG representations. Feature-level Normative Deviation-guided Diagnosis (F-NDD) retains dimension-wise residuals and combines them with deep EEG representations through attention-based aggregation and auxiliary supervision. We also employ target HC-assisted score calibration for cross-source evaluation without updating the trained models. Experiments on two public EEG datasets examine both approaches under subject-independent within-source and bidirectional cross-source settings. The results demonstrate the effectiveness of normative deviation-guided diagnosis across different evaluation settings.
Robust anatomical segmentation of chest X-rays (CXRs) remains challenging due to the scarcity of comprehensive annotations and the substantial variability of real-world acquisition conditions. We propose AnyCXR, a unified framework that enables generalizable multi-organ segmentation across arbitrary CXR projection angles using only synthetic supervision. The method combines a Multi-stage Domain Randomization (MSDR) engine, which generates over 100,000 anatomically faithful and highly diverse synthetic radiographs from 3D CT volumes, with a Conditional Joint Annotation Regularization (CAR) learning strategy that leverages partial and imperfect labels by enforcing anatomical consistency in a latent space. Trained entirely on synthetic data, AnyCXR achieves strong zero-shot generalization on multiple real-world datasets, providing accurate delineation of 54 anatomical structures in PA, lateral, and oblique views. The resulting segmentation maps support downstream clinical tasks, including automated cardiothoracic ratio estimation, spine curvature assessment, and disease classification, where the incorporation of anatomical priors improves diagnostic performance. These results demonstrate that AnyCXR establishes a scalable and reliable foundation for anatomy-aware CXR analysis and offers a practical pathway toward reducing annotation burdens while improving robustness across diverse imaging conditions.
Open-Set Test-Time Adaptation (OSTTA) operates on non-stationary target streams where covariate and semantic shifts coexist. Existing TTA methods often obtain adaptation signals by updating model parameters or maintaining mutable memories, which can be costly and vulnerable to unknown samples. We present TALOR, a lightweight, backward-free rectification module that uses the source linear head as a geometric anchor. The head-induced basis decomposes each normalized feature into high-gain principal coordinates and low-gain tail coordinates. TALOR estimates a soft-routed affine correction in the principal subspace, combining a regime-level principal bias with a tail-conditioned slope that captures sample-level tail-to-principal coupling. Using weighted ridge regression, it subtracts this drift estimate from the principal coordinates, reconstructs the feature and feeds it into the source head. On the ImageNet-C benchmark with six csOOD datasets, TALOR outperforms the next-best UniEnt by 2.2% H-score while running 2.5x faster and using only 12% of its GPU memory; as a post-correction plug-in, it further boosts COME to 68.4% H-score with only 2% additional memory overhead.
Multi-turn jailbreaks turn dialogue safety into a trace property: a violation may become identifiable only when user intent and assistant behavior are considered across multiple turns. Existing guard models largely classify individual messages or exchanges and return risk scores or labels, without executable safety semantics or a formally checkable decision trail. We present SpecGuard, a training-free, specification-driven framework for monitoring multi-turn LLM interactions. SpecGuard uses a fixed atomic-proposition contract to map natural-language observations into causally scoped, evidence-backed facts, and expresses forbidden interaction patterns as Linear Temporal Logic specifications. Each specification is compiled into a symbolic finite automaton that incrementally consumes the resulting dialogue trace. The same monitoring mechanism supports buffered runtime checking and retrospective auditing, while accepting runs produce replayable witnesses that identify the triggered specification, supporting facts, and relevant dialogue turns. Experiments on multi-turn safety benchmarks show that SpecGuard consistently outperforms representative open-source guard models while maintaining controlled false-positive rates. These results demonstrate a practical path from heuristic risk classification toward formally specified, verifiable, and auditable LLM safety monitoring, without learned detector parameters or access to the protected model.
Automated fake news detection is vital on social media, where fake news is easily rewritten to fool detectors. However, existing approaches typically rely on superficial writing cues or single-style text rewrites, resulting in poor robustness under diverse cosmetic modifications. We construct MANTIS-Bench, a multi-axis presentation-attack benchmark that pairs original news with rewritten views across writing style, emotional expression, and dissemination framing. We further propose MANTIS, an instance-adaptive adversarial framework enabling dynamic worst-case training tailored to each news article with zero inference-time overhead. This reframes fake news defense from heuristic format smoothing into targeted, fact-invariant credibility assessment. Experiments on MANTIS-Bench show that MANTIS reduces the Union Attack Success Rate (UASR) by 3.55 percentage points relative to SheepDog on GossipCop and consistently outperforms prior single-axis defenses on Macro-F1 while maintaining competitive clean accuracy, demonstrating the effectiveness of multi-axis and instance-adaptive worst-case training.
Machine learning models are now widely used to predict environmental variable dynamics in domains like hydrology, climate science, agriculture, and ecology. These models are usually compared on fixed benchmark datasets using a single average accuracy score. As models improve, their scores become very close when averaged over large regions and long time periods, so the benchmark can no longer tell them apart, and a single number hides \emph{where} and \emph{when} a model fails. Three failures are common across environmental domains: models get worse as the climate shifts (non-stationarity), models fail at unmonitored sites (spatial extrapolation), and models miss rare but important regimes such as floods and droughts. We present \textbf{EvoBench}, a benchmark framework that shows where a model fails, tracks failures over time, and evolves to focus on model capability gaps. We apply EvoBench to daily streamflow prediction as a case study (\textbf{CAMELS-Evolve}, about 9{,}000 U.S. basins, 1985--2024) and compare an operational physical model (NWM), a conceptual reference (HBV), and several data-driven models. No model wins on every testing scenarios (e.g., in terms of average accuracy, performance at low flows and unmonitored basins). Also we observe that all data-driven models under-predict floods. Additionally, we show that targeted fix based on the reported failures could mitigate weaknesses the benchmark identified in certain scenarios. EvoBench turns model evaluation from a saturated ranking into a moving map of model limitations.
Omnimodal large language models (Omni-LLMs) are increasingly expected to assist users while observing continuous audiovisual streams. However, existing evaluations provide limited insight into how reliably these models connect stream evidence and user needs to timely, sustained interventions. We introduce \textbf{Omni-InteractScope}, a benchmark of 1,479 human-annotated, cross-reviewed interactions with 1,790 temporal targets across eight task groups. It links input modalities, event triggers, intervention types, and user needs, covering event-timed tool invocation and query-free assistance alongside notifications, warnings, state tracking, and continuous responses. Our experimental results on the proposed benchmark reveal three critical limitations of current Omni-LLMs: (1) \textbf{Timely tool invocation remains unreliable}: all 12 evaluated models produce an on-time first response for at most 40\% of tool-invocation targets, and joint timing-and-tool accuracy reaches at most 32.0\%. (2) \textbf{Query-free reminders remain poorly covered}: all evaluated models achieve lower timely reference-target recall on query-free reminders than on explicit questions, measured on their respective target sets. Reminder recall never exceeds 23.0\%. (3) \textbf{End-to-end guidance coverage remains limited}: nine models cover none of the 26 guidance sequences in full, and the remaining three cover only one to three sequences each. Targeted controls further show that tool-invocation scores alone do not establish event grounding and that reminder coverage depends on response history and frequency. These findings motivate developing and evaluating streaming assistants for timely action, initiative under general user needs, and coverage of entire response sequences.
Jie Li, Bin Guo, Georgina Curto, Zikang Liu, Mengqi Chen, yu li
One-person companies (OPCs) are emerging enterprises operated by individual entrepreneurs through digital platforms and LLM-based AI agents. These highly skilled entrepreneurs are characterized by being digital nomads, with limited reliance on local ecosystems. Governments in China use distinct regional support policies to attract mobile OPCs, turning local support into a cross-regional coordination problem. OPCs compare regions to maximize their own utility, but imperfect policy perceptions may cause them to reject otherwise suitable locations. Existing policy-learning methods optimize resources under fixed perceptions, while consensus-formation methods aggregate views without feeding cognitive changes back into policy learning. This disconnect produces mismatched location choices and inefficient regional policy adjustment. We introduce PolicyMediator, a cognition-constrained leader-multi-follower Stackelberg framework with bilateral mediation. The mediator explains verified policies to OPCs to build awareness and trust, and reports their needs and unresolved concerns to governments for policy learning. Its self-evolving skills retain, revise, or prune strategies using verification, allocation, capacity, and objective trajectories rather than mediation success alone. Across six-region simulations in China with five OPC types, five scenarios, and three OPC-simulator backbones, PolicyMediator ranks first in all 15 OPC-backbone--scenario combinations. Over the eight-baseline mean, it improves the aggregate government objective, combining economic performance, employment benefit, and policy-access equality, by 42.4% and average OPC utility by 40.9%.
Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random. We trace this to the gate that scales each retrieved row. A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code. GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest. Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models.
Qi Chen, Yuwei Wu, Quan Chen, Kunpeng Gao, Jin Chen, Che Sun, Zhi Gao, Yunde Jia
Vision-language-action (VLA) models have recently achieved remarkable progress by leveraging large-scale robot datasets and foundation models for general-purpose robot manipulation. However, their ability to rapidly adapt to novel manipulation tasks beyond the training distribution remains limited. Existing approaches often require task-specific robot demonstrations and additional training, which are labor- and compute-intensive, limiting the scalability of VLA adaptation to open-world scenarios. To reduce this dependence on target-task robot data, we propose CapPro, a framework that enables a VLA policy to adapt to previously unseen manipulation tasks using only a single human demonstration video. CapPro consists of two components. First, a phase-aware human video representation module constructs a compact video representation through learned visual token selection and learns task-relevant query representations under phase-caption supervision. Second, a video-conditioned adaptation module jointly incorporates the compact video and query-derived representations into the action-conditioning context of the pretrained VLA policy. % This design introduces task-specific guidance while largely preserving the original VLA action-generation backbone. In this way, the human demonstration provides task-specific execution guidance rather than an embodiment-aligned trajectory to imitate, without requiring target-task robot demonstrations or target-task fine-tuning. Experiments across source and unseen target tasks show that CapPro improves average performance across unseen target groups while maintaining comparable performance on the source tasks. These results demonstrate the effectiveness of human-video guidance for cross-task VLA adaptation.
This paper studies the hierarchical joint source-channel coding with information leakage constraint in the first-phase reconstruction and distortion constraints. The receiver's access to the data varies and is evaluated by the quality of the side information. Due to the consideration of channel capacity limitation or the efficiency of the system performance, the encoder may send some additional information in Phase 1 that can only be decoded in Phase 2 with higher-quality side information. While this can optimize the overall performance, the additional information causes excessive information leakage. We provide general inner and outer bounds for the conditions such that a given distortion-leakage pair $(D_1,D_2,L)$ is achievable, together with a capacity-achieving condition.
Tongyao Zhu, Qian Liu TikTok, Chang Ma, Jinghan Zhang Hong Kong University of Science and Technology, Longxu Dou Sea AI Lab, Junxian He Hong Kong University of Science and Technology, Shiqi Chen The Hong Kong University of Science and Technology, Northwestern University and City University of Hong Kong
Low-resource languages challenge multilingual LLMs due to limited high-quality training data, leading to weaker performance on complex reasoning and knowledge tasks. To address this, we propose improving training data quality through data synthesis, moving beyond simple resource scaling. First, we introduce SynTrans, which translates high-quality, knowledge-rich English data into low-resource languages during pre-training to inject world knowledge, though at the cost of semantic fluency. To overcome low-quality data issues while maintaining fluency, we also propose SynRank. SynRank leverages synthetic data as positive samples to train a classifier that ranks and filters noisy real-world data, enabling the extraction of high-quality subsets without expensive human cleaning. Experiments show SynRank matches handcrafted rule-based filtering by human experts and significantly improves knowledge-intensive task performance at the same filtering rate. Remarkably, higher filtering rates even improve performance with less data, demonstrating the efficiency and effectiveness of our method, surpassing expert filtering. Lastly, we introduce DA-QwenScore, a training-free metric that evaluates corpus quality by normalizing model loss with diversity measures, further enhancing evaluation efficiency. Our insights into knowledge injection could advance low-resource multilingual LLM development.
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
Yu-An Liu, Ruqing Zhang Institute of Computing Technology, Chinese Academy of Sciences, Hongru Song , Chinese Academy of Sciences, Jiafeng Guo Institute of Computing Technolgy, Chinese Academy of Sciences, Yixing Fan, Xueqi Cheng Institute of Computing Technology, Chinese Academy
While neural ranking models (NRMs) have achieved state-of-the-art performance in information retrieval, they remain highly vulnerable to imperceptible adversarial perturbations. Existing defenses are predominantly data-centric, exemplified by adversarial training, which requires constructing large collections of adversarial examples. By treating NRMs as black boxes and indiscriminately optimizing all model parameters, these methods incur substantial computational cost and often degrade performance on clean data due to overfitting. In this paper, we advocate that adversarial vulnerability is not uniformly distributed across model parameters, but instead originates from specific internal units. We propose a paradigm shift toward a model-centric defense that addresses vulnerability at its architectural source, without requiring costly retraining or adversarial data generation. Specifically, we introduce Search in the Model, a novel training-free framework that performs fine-grained identification and rectification of vulnerable neurons directly within the model. By formulating neuron identification as a ranking problem, we develop a maximum marginal vulnerability criterion to precisely locate the top-K neurons most responsible for model vulnerability, and apply targeted neuronal inverse perturbation to correct them. Extensive experiments on MS MARCO and TREC 19 show that our approach outperforms state-of-the-art baselines in both defense efficiency and robustness to seen and unseen attacks, while preserving strong performance on clean data.
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
Wei Shao Institute of Computing Technology, Chinese Academy of Sciences, Yihang Wang ByteDance Inc. and University of Chinese Academy of Sciences, Gao yu Zhu, Ziqiang Cheng, Lei Yu, Jiafeng Guo Institute of Computing Technolgy, Chinese Academy of Sciences, Xueqi Cheng Institute of Computing Technology, Chinese Academy
Existing detoxification methods for large language models mainly focus on post-training stage or inference time, while few tackle the source of toxicity, namely, the dataset itself. Such training-based or controllable decoding approaches cannot completely suppress the model’s inherent toxicity, whereas detoxifying the pretraining dataset can fundamentally reduce the toxicity that the model learns during training. Hence, we attempt to detoxify directly on raw corpora with SoCD (Soft Contrastive Decoding), which guides an LLM to localize and rewrite toxic spans in raw data while preserving semantics, in our proposed HSPD (Hierarchical Semantic-Preserving Detoxification) pipeline, yielding a detoxified corpus that can drop-in replace the original for fine-tuning or other training. On GPT2-XL, HSPD attains state-of-the-art detoxification, reducing Toxicity Probability (TP) from 0.42 to 0.18 and Expected Maximum Toxicity (EMT) from 0.43 to 0.20. We further validate consistent best-in-class results on LLaMA2-7B, OPT-6.7B, and Falcon-7B. These findings show that semantics-preserving, corpus-level rewriting with HSPD effectively suppresses downstream toxicity while retaining data utility and allowing seamless source-level mitigation, thereby reducing the cost of later model behavior adjustment.
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
Biao Wu, Yutong Xie Mohamed bin Zayed University of Artificial Intelligence, Zeyu Zhang Australian National University, Vu Minh Hieu Phan, Qi Chen Adelaide University, Ling Chen University of Technology Sydney, Qi Wu The University of Adelaide
Vision-and-language pretraining (VLP) in medicine leverages contrastive learning on image–text pairs, often enhanced with masked modeling. However, existing methods face two challenges: difficulty reconstructing key pathological features due to limited data, and reliance on either paired or image-only datasets without combining both. To address this, we propose MMCLIP (Masked Medical Contrastive Language–Image Pre-training), which introduces two modules: AttMIM: Masks image features highly correlated with text to improve reconstruction of fine medical details. EntMLM: Masks key medical entities in text and reconstructs them using visual cues. Furthermore, MMCLIP incorporates unpaired data through disease-kind prompts, achieving state-of-the-art performance in zero-shot and fine-tuning across five benchmarks.
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
Zihao Wei Institute of Computing Technology, Chinese Academy of Sciences, Liang Pang Institute of Computing Technology, Chinese Academy of Sciences, Jiahao Liu Meituan, Wenjie Shi Meituan, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Jingang Wang Meituan 等 11 位作者
Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., lt;/think gt;). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.
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
Lingrui Mei, Shenghua Liu Institute of Computing Technology, Chinese Academy of Sciences, Yiwei Wang University of California, Merced, Yuyao Ge, Baolong Bi, Jiayu Yao, Jun Wan UBS Group AG, Ziling Yin 等 10 位作者
Long contexts break transformers: attention scores dilute across thousands of tokens, critical information gets lost in the middle, and the model cannot adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory—transient parameters updated on the current context—but existing approaches employ uniform write policies that waste computation on low-value regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, asking: given limited computational budget, which parts of the context should be consolidated into working memory? We propose GDWM (\textbf{G}ated \textbf{D}ifferentiable \textbf{W}orking \textbf{M}emory), a framework that introduces a Write Controller to gate the memory consolidation process. Our controller estimates Contextual Utility—an information-theoretic measure quantifying how much each region depends on long-range context—and allocates gradient steps accordingly, subject to a coverage constraint that ensures global representation. Theoretically, we prove that our chunk-restricted sampling strategy reduces gradient variance by eliminating inter-chunk variance via the Law of Total Variance. Experiments on ZeroSCROLLS and LongBench v2 benchmarks demonstrate that GDWM achieves comparable or superior performance with 4 ×fewer gradient steps compared to uniform baselines—excelling on sparse-information tasks (+6–13% on Qasper, +5–13% on GovReport for smaller models) while revealing principled trade-offs on dense-coverage tasks, establishing a new efficiency-performance Pareto frontier for test-time adaptation.
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
Lingrui Mei, Shenghua Liu Institute of Computing Technology, Chinese Academy of Sciences, Yiwei Wang University of California, Merced, Baolong Bi, Ruibin Yuan, Xueqi Cheng Institute of Computing Technology, Chinese Academy
As Large Language Models (LLMs) grow increasingly powerful, ensuring their safety and alignment with human values remains a critical challenge. Current alignment approaches predominantly rely on refusal alignment, such as training models to refuse harmful prompts or implementing filters at various stages to block certain responses. These methods are designed toward a binary outcome: either denying to answer the question entirely or answering with full access to the model’s parametric knowledge. The binary nature of current alignment approaches presents significant limitations. These methods often fail to balance safety and utility, resulting in either overly cautious responses or overlooking subtle harmful content. They also prevent users from accessing benign information when it’s mixed with harmful content. For instance, a model might refuse to provide basic, public information about a medication’s composition due to misuse concerns. Furthermore, these approaches struggle with context-dependent sensitivity, potentially over-censoring harmless content or missing nuanced harmful outputs. Ideally, LLMs should offer informative responses while avoiding the disclosure of harmful and sensitive information. To address these challenges, we introduce HiddenGuard, a novel framework for fine-grained safe generation in LLMs. Our method incorporates PRISM (rePresentation Router for In-Stream Moderation), a specialized moudule that operates alongside the LLM architecture. By leveraging intermediate hidden states, HiddenGuard enables real-time, token-level harmfulness detection and redaction, without loss in capability. This approach captures deeper semantic information, allowing for more nuanced and context-aware content control compared to traditional filtering techniques. Consequently, the model can generate informative responses while selectively redacting or replacing sensitive information, rather than refusing to answer outright. We also contribute a comprehensive dataset with token-level fine-grained annotations of potentially harmful information across diverse contexts. Our experiments demonstrate that HiddenGuard achieves over 90% in F1 score for detecting and redacting harmful content while preserving the overall utility and informativeness of the model’s responses.
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
Zenghao Duan, Zhiyi Yin , Chinese Academy of Sciences, Zhichao Shi UCAS, Liang Pang Institute of Computing Technology, Chinese Academy of Sciences, Shaoling Jing, Zihe Huang, Jiayi Wu, Yu Yan 等 11 位作者
Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks. Prior mechanistic studies characterize toxic regions as “toxic vectors” or “layer-wise subspaces”, yet our analysis identifies critical limitations: i) Removed toxic vectors can be reconstructed via linear combinations of non-toxic vectors, demanding targeting of entire toxic subspace; ii) Contrastive objective over limited samples inject noise into layer-wise subspaces, hindering stable extraction. These highlight the challenge of identifying robust toxic subspace and removing them. Therefore, we propose GLOSS (GLobal tOxic Subspace Suppression), a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters. Experiments on LLMs (e.g., Qwen3) show GLOSS achieves SOTA detoxification while preserving general capabilities without requiring large-scale retraining.
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
Renqi Chen, Zeyin Tao Alibaba Group, Jianming Guo, Jingzhe Zhu, Yiheng Peng, Qingqing Sun, Tianyi Zhang Alipay, Shuai Chen Ant Group
E-commerce risk management requires aggregating diverse, deeply embedded web data through multi-step, stateful interactions, which traditional scraping methods and most existing Graphical User Interface (GUI) agents cannot handle. These agents are typically limited to single-step tasks and lack the ability to manage dynamic, interactive content critical for effective risk assessment. To address this challenge, we introduce RISK, a novel framework designed to build and deploy GUI agents for this domain. RISK integrates three components: (1) RISK-Data, a dataset of 8,492 single-step and 2,386 multi-step interaction trajectories, collected through a high-fidelity browser framework and a meticulous data curation process; (2) RISK-Bench, a benchmark with 802 single-step and 320 multi-step trajectories across three difficulty levels for standardized evaluation; and (3) RISK-R1, a R1-style reinforcement fine-tuning framework considering four aspects: (i) Output Format Constraint, (ii) Single-step and (iii) Multi-step Level Reward, and (iv) Task Level Reweight. Experiments show that RISK-R1 achieves a 6.8% improvement in offline single-step and an 8.8% improvement in offline multi-step, using only 7.2% of the parameters of the SOTA baseline. Moreover, it attains a top task success rate of 70.5% in online evaluation. RISK provides a scalable, domain-specific solution for automating complex web interactions in e-commerce risk management. The code is available at https://github.com/RenqiChen/RISK-GUI.
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
Meiqi Chen, Fandong Meng WeChat AI, Tencent Inc., Jie Zhou
Complex reasoning problems often involve implicit spatial and geometric relationships that are not explicitly encoded in text. While recent reasoning models perform well across many domains, purely text-based reasoning struggles to capture structural constraints in complex settings. In this paper, we introduce FIGR, which integrates executable visual construction into multi-turn reasoning via end-to-end reinforcement learning. Rather than relying solely on textual chains of thought, FIGR externalizes intermediate hypotheses by generating executable code that constructs diagrams within the reasoning loop. An adaptive reward mechanism selectively regulates when visual construction is invoked, enabling more consistent reasoning over latent global properties that are difficult to infer from text alone. Experiments on seven challenging mathematical benchmarks demonstrate that FIGR outperforms strong text-only chain-of-thought baselines, improving the base model by 13.12% on AIME 2025 and 11.00% on BeyondAIME. These results highlight the effectiveness of precise, controllable figure construction of FIGR in enhancing complex reasoning ability.
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
Yiming Ni University of Washington, Zhi-Qi Cheng University of Washington Tacoma and University of Washington Graduate School, Jiayu Li University of Washington, Wei Cheng University of Washington
Sign languages are expressive visual languages used by Deaf and Hard-of-Hearing (DHH) communities. Despite substantial progress in sign-language recognition, translation, and production, advances remain constrained by fragmented datasets, inconsistent annotations, and limited linguistic coverage. Existing benchmarks often fail to reflect real-world communication needs, and systematic analyses of these limitations remain limited. In this survey, we present a comprehensive index of sign-language datasets, covering 120 resources across 35 sign languages. We analyze key challenges such as modality imbalance, annotation granularity, and signer bias, and outline considerations for future dataset design. We also introduce a 24-field Sign-Language Datasheet and release a public GitHub repository to support standardized documentation and reproducible evaluation. Overall, our work provides a unified and practical foundation for developing inclusive, robust, and scalable sign-language technologies in real-world applications.
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
Hexiang Tan, Wanli Yang, Junwei Zhang, Xin Chen Meituan, Rui Tang, Du Su Institute of Computing Technology, Chinese Academy of Sciences, Jingang Wang Meituan, Yuanzhuo Wang Chinese Academy of Sciences 等 10 位作者
Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM’s responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM’s output layer to derive base-calibrated confidence for PoLLM’s responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90% compared to the best unsupervised baselines.
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
Xin Liu , Chinese Academy of Sciences, Yu-An Liu, Ruqing Zhang Institute of Computing Technology, Chinese Academy of Sciences, Yixing Fan, Lixin Su Baidu, Jiafeng Guo Institute of Computing Technolgy, Chinese Academy of Sciences, Xueqi Cheng Institute of Computing Technology, Chinese Academy
Retrieval-augmented generation (RAG) has emerged as a promising paradigm for mitigating hallucinations in large language models (LLMs).However, the intrinsic heterogeneity between the retriever and the generator often leads to a mismatch between retrieved evidence and generation needs, hindering effective coordination.We argue that competition between discriminative retrieval and generative modeling can more effectively expose their mutual weaknesses and induce deeper interaction. Motivated by this insight, we propose CARL (Co-opetition AdveRsarial Learning), a framework that formulates retriever–generator training in RAG as a minimax game. In this game, the retriever is optimized to retrieve both useful and adversarially useless documents to challenge the generator, while the generator learns to identify useful evidence and remain robust to misleading retrievals to produce accurate answers.Experiments on seven benchmark datasets demonstrate that CARL consistently improves RAG performance, validating the effectiveness of adversarial co-opetition in enhancing retriever–generator synergy.
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
Zhaodan Zhang, Jin Zhang, Jiafeng Guo Institute of Computing Technolgy, Chinese Academy of Sciences, Xueqi Cheng Institute of Computing Technology, Chinese Academy
Stance detection aims to identify the attitude expressed in text towards a given target, with applications in public opinion analysis and misinformation mitigation. Despite recent advances in large language models (LLMs), two key challenges remain: (1) spurious correlations between superficial features and stance labels, and (2) lack of cognitive modeling that simulates the transition from intuitive perception to deliberate reasoning. To address these issues, we propose Cognitive-Driven Stance Detection (CDSD), inspired by Kahneman’s Dual-Process Theory. CDSD integrates fast intuitive judgment (System 1) and analytical reasoning (System 2), enhanced by three key modules: attention-based cognitive alignment to compare system focus, uncertainty-aware belief update using Bayesian inference, and self-doubt-triggered counterfactual reasoning for re-evaluation under low consistency or high uncertainty. Experimental results on SEM16, P-Stance, and VAST show that CDSD outperforms state-of-the-art methods across multiple LLMs. Notably, CDSD exhibits strong robustness against textual perturbations such as emotional word removal and rhetorical restructuring. By integrating cognitive theory with NLP, our work provides a promising path toward more reliable and interpretable stance detection systems.
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
Yihang Wang ByteDance Inc. and University of Chinese Academy of Sciences, Bin Wu, Yueyang Su Chinese Academy of Sciences, Tianfu Zhang ByteDance Inc., Yiqi Du ByteDance Inc., Lei Yu, Jiafeng Guo Institute of Computing Technolgy, Chinese Academy of Sciences, Xueqi Cheng Institute of Computing Technology, Chinese Academy
Large embedding models have become the backbone of modern retrieval systems, offering strong semantic representations at the cost of substantial storage and computation. While recent work explores quantizing embeddings into discrete document identifiers for generative retrieval, most existing approaches rely on Euclidean quantization, which is poorly aligned with the angular geometry induced by contrastive embedding training and often requires long identifier sequences to preserve semantic fidelity. In this work, we propose Hyperspherical Householder Quantization (HHQ), a geometry-aware distillation method that compresses large embeddings into short discrete representations via iterative Householder transformations on the unit hypersphere. By explicitly preserving cosine similarity at each step, HHQ distills semantic structure into compact identifiers that remain faithful to the original embedding space. To support reliable generation of these identifiers, we introduce constrained supervised fine-tuning and tree-aware dynamic masking to enforce structural validity during training and inference. Experiments on NQ and MS MARCO show that HHQ achieves competitive or superior retrieval performance using only five tokens per document, substantially reducing decoding cost while retaining strong semantic retrieval accuracy.
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