AI Cosmos Team, Jinfeng Zhou, Yuxuan Chen, Fangming Chang, Mingyu Su, Yifan Zhu, Yuan Wang, Bo Zhang 等 35 位作者
Large language models (LLMs) can improve task performance during learning, yet these gains do not necessarily reflect learners' ability to thereafter apply the knowledge and skills independently. We present AI Cosmos, an agentic infrastructure for situated learning that organizes learning around two complementary principles: scenario-based participation and cumulative assets of practice. AI Cosmos organizes learners and role-specialized agents within learning scenarios that integrate task-relevant knowledge and, when needed, executable workbenches. An asset-memory loop preserves the processes and outcomes of practice as reusable assets with explicit provenance and dependencies. We evaluate scenario-based interaction through two preregistered randomized controlled studies examining the acquisition of legal reasoning and programming competence, holding the same underlying LLM across conditions. AI Cosmos achieves greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.
Thinking vision-language models reason for thousands of tokens, yet still miss small details in high-resolution images. Zooming methods pay for extra inference passes, and outcome rewards never tell the model where to look at. In Qwen3.5-4B, we find that the bottleneck is selection, not seeing: a few attention heads, which we call localization heads, pick the evidence out of thousands of visual tokens; ablating four of them costs 13.2 points more than ablating four random heads, and their focus fades as the reasoning grows. We therefore propose Localization-Head Alignment (LHA), which makes attention a training signal: whenever the reasoning states a visual fact, these heads learn to look at the focal region rather than the background, as an auxiliary loss in both SFT and RL training. On benchmarks of V*, HR-Bench 4K and TreeBench, LHA lifts Qwen3.5-4B from 74.1% to 76.6% accuracy, the best single-pass 4B result we evaluate, ahead of multi-pass zooming and released 7B models. Alignment alone adds 1.8 points over otherwise identical training while shortening reasoning by 20%, and the gains carry over to Qwen2.5-VL-3B.
Ruixin Fan, Weining Wang, Jiahui Sun, Tailai Chen, Jing Liu
Unified multimodal models (UMMs) jointly support visual understanding and image generation, yet information flow between these two capabilities remains largely asymmetric: understanding representations routinely condition generation, whereas generative representations are rarely reused for multimodal reasoning. In this work, we investigate whether intermediate representations from a UMM's native diffusion branch can enhance spatial reasoning without compromising general multimodal understanding. We propose DiFuse, which selectively incorporates intermediate diffusion representations into the understanding pathway of UMMs. Through systematic layer-wise analysis, we identify which diffusion representations are most informative and where they can be integrated with minimal interference to semantic processing. Guided by these findings, DiFuse spatially aligns diffusion features with visual tokens and injects them into selected multimodal decoder layers via residual fusion. Extensive experiments on UniWorld-V1 show that, DiFuse consistently improves spatial reasoning over a matched SFT baseline while maintaining broadly comparable performance on general multimodal benchmarks. Further analyses show that these gains depend critically on the choice of diffusion representations, injection positions, and the preservation of spatial correspondence between diffusion features and visual tokens. Results on OmniGen2 further demonstrate the transferability of DiFuse across unified multimodal architectures. Overall, our findings show that native diffusion representations can provide complementary spatial cues for multimodal reasoning when selectively integrated in a structure-preserving manner.
Erle Zhu, Genrui Zhang, Keyang Hu, Mingyu Su, Hongning Wang
Coding agents are moving from handling one request at a time to serving as persistent assistants to which users hand a stream of work: a build, a data-processing job and a bug fix may all be pending on the same machine at once. Such an agent must decide when to take on new tasks, how to delegate and overlap them, and how to stay within the machine's CPU and memory. Existing benchmarks cannot measure this, because they evaluate each task in isolation. We introduce AOS-Bench, which evaluates an agent system as a workload manager: verified tasks arrive over time and must be delivered under shared CPU and memory limits, and the same tasks run one at a time serve as a per-system reference. Across five model--harness pairs, isolated task success does not carry over evenly. On heavy workloads every system loses 13 to 23 points. On medium workloads, systems span 11 points in isolated success but 26 points in stream success, and two systems 0.6 points apart in isolation deliver 54\% and 68\% of the same tasks. Execution-policy interventions trace much of this gap to task intake: serial dispatch leaves 73\% of heavy-workload tasks unclaimed. Yet more elaborate management does not necessarily help. A fixed concurrency of two comes within four points of the best agent-managed policy on both medium and heavy workloads, and under every policy median CPU use stays below 30\% of the provisioned cores while tasks wait. These results identify workload management as a distinct capability for agents, and AOS-Bench as a way to measure it.
Reviewing medical records for clinical and insurance decisions must handle long, heterogeneous documents while producing consistent, traceable, guideline-compliant outcomes under strict latency and cost constraints. We propose GuideTree, which compiles textual guidelines into a fixed review tree of evidence-grounded verification primitives. GuideTree uses short per-document summaries only for routing each check to a minimal set of document types and candidates; final verification always reads full document text and returns structured evidence. The tree is induced offline via a cost-aware split-and-prune search and updated safely through regression-tested, versioned patches. Across 1,000 cases from four industrial review scenarios and four LLM backbones, GuideTree achieves 84.5–92.8 Macro-F1, outperforming the strongest non-expert baselines by 3.3–7.6 points and matching ExpertTree within 0.2–0.6 points (avg. 0.38). On chronic disease with Qwen3-235B-A22B-Instruct, GuideTree reduces average I/O volume to 74K input+output characters (-82% vs. long-context prompting) and average latency to 22s (-83% vs. long-context prompting), while reaching 99% decision consistency over K=5 reruns.
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 6: Industry Track) · 2026 · 正式发表
Zhicheng Ma, Xiang Liu, Zhaoxiang Liu, Ning Wang, Yi Shen, Kai Wang, Shuming Shi, Shiguo Lian
Large Language Models (LLMs) based on Mixture-of-Experts (MoE) are pivotal in industrial applications for their ability to scale performance efficiently. However, standard MoEs enforce uniform expert sizes, creating a rigidity that fails to align computational costs with varying token-level complexity. While heterogeneous expert architectures attempt to address this by diversifying expert sizes, they often suffer from significant system-level challenges, specifically unbalanced GPU utilization and inefficient parameter utilization, which hinder practical deployment.To bridge the gap between theoretical heterogeneity and robust industrial application, we propose Mixture of Heterogeneous Grouped Experts (MoHGE) which introduces a two-level routing mechanism to enable flexible, resource-aware expert combinations. To optimize inference efficiency, we propose a Group-Wise Auxiliary Loss, which dynamically steers tokens to the most parameter-efficient expert groups based on task difficulty.To address the critical deployment challenge of GPU load balancing, we introduce an All-size Group-decoupling Allocation strategy coupled with an Intra-Group Experts Auxiliary Loss. These mechanisms collectively ensure uniform computation distribution across GPUs.Extensive evaluations demonstrate that MoHGE matches the performance of MoE architectures while reducing the total parameters by approximately 20% and maintaining balanced GPU utilization. Our work establishes a scalable paradigm for resource-efficient MoE design, offering a practical solution for optimizing inference costs in real-world scenarios.
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 6: Industry Track) · 2026 · 正式发表
Sudharshan Balaji, Zhiyu Liu, Zhengyuan Jiang, Shuo Lei, Yimin Chen, Yang Xiao, Shone O. Almeida, Mathew Joseph Karivelil 等 10 位作者
While coronary imaging is widely used for anatomical assessment, CCTA reports play a distinct last-mile role in clinical care. Ratherthan serving as an intermediate signal, CCTA provides an assessment of coronary disease severity (known as the CAD-RADS score) toguide patient management. However, real-world clinical text exhibits substantial heterogeneity in terminology and structure, leadingto inconsistent interpretation by automated systems, even for clinically similar cases. Recent work leverages a direct application ofLLMs for automated CAD-RADS scoring, but is limited by small, non-public, and homogeneous clinical data. We introduce CCTA-RADS, the largest publicly available dataset of 940 real-world CCTA reports from a major cardiovascular center, each annotated with CAD-RADS scores. Our analysis reveals that direct approaches, including state-of-the-art LLMs (GPT-4o, GPT-o3) and fine-tuned BERT models underperform on diverse real-world clinical data. To address these limitations, we propose a two-stage pipeline that decouples structuring from classification: an LLM-based parser normalizes heterogeneous reports into structured format, followed by fine-tuned BERT classification. This approach substantially improves the F1-score by 6%-13% compared with direct methods. We deploy our system as an interactive web interface that allows clinicians to upload CCTA reports for automated CAD-RADS assessment with SHAP and LIME explainability visualizations.
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 6: Industry Track) · 2026 · 正式发表
Bosi Wen, Yilin Niu, Cunxiang Wang, Xiaoying Ling, Ying Zhang, Pei Ke University of Electronic Science and Technology of China, Hongning Wang Tsinghua University, Minlie Huang
Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.
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
Bosi Wen, Yilin Niu, Cunxiang Wang, Pei Ke University of Electronic Science and Technology of China, Xiaoying Ling, Ying Zhang, Aohan Zeng Tsinghua University, Tsinghua University, Hongning Wang Tsinghua University 等 9 位作者
Instruction-following is a fundamental ability of Large Language Models (LLMs), requiring their generated outputs to follow multiple constraints imposed in input instructions. Numerous studies have attempted to enhance this ability through preference optimization or reinforcement learning based on reward signals from LLM-as-a-Judge. However, existing evaluation models for instruction-following still possess many deficiencies, such as substantial costs and unreliable assessments. To this end, we propose IF-CRITIC, an LLM critic for fine-grained, efficient, and reliable instruction-following evaluation. We first develop a checklist generator to decompose instructions and generate constraint checklists. With the assistance of the checklists, we collect high-quality critique training data through a multi-stage critique filtering mechanism and employ a constraint-level preference optimization method to train IF-CRITIC. Extensive experiments show that the evaluation performance of IF-CRITIC can beat strong LLM-as-a-Judge baselines, including o4-mini and Gemini-3-Pro. With the reward signals provided by IF-CRITIC, LLMs can achieve substantial performance gains in instruction-following optimization under lowercomputational overhead compared to strong LLM critic baselines. Our code and model are available at https://github.com/thu-coai/IF-CRITIC.
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
Xiaocheng Zhang Meituan, Xi Wang, Yifei Lu Qiantang Credit, Jianing Wang, Zhuangzhuang Ye, Mengjiao Bao Meituan, Peng Yan Meituan, Xiaohong Su
With the surge of online misinformation, Large Language Models (LLMs) and Reasoning Large Language Models (RLMs) serving as Automatic Fact-Checking (AFC) systems have emerged as a prominent paradigm for reliable, explainable verification. However, our empirical study reveals that this paradigm faces a critical risk asymmetry challenge when deployed in real-world under resource-constrained environments. While Hotspot Perception Ability (HPA), the capacity to dynamically allocate reasoning resources based on social impact, is essential to mitigate this risk, existing benchmarks lack the social metadata and evaluation framework to meet this urgent evaluation needs, thereby hindering the advancement of these AFC systems. To bridge this gap, we introduce TrendFact, the first benchmark capable of evaluating HPA and three fact-checking tasks. It consists of 7,643 curated samples sourced from trending platforms and professional datasets, with an evidence library containing 366,634 entries. To enable HPA assessment, we propose two novel metrics: the Explanation Consistency Score (ECS) to evaluate the reliability of verification reasoning, and the Hotspot Claim Perception Index (HCPI) to quantify the overall HPA of AFC systems. Extensive experiments demonstrate that existing AFC systems exhibit limited performance on TrendFact. Furthermore, our proposed FactISR framework effectively enhances HPA and computational efficiency for RLM-driven 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
Dexuan Xu, Yuan Jiayin, Jianing Wang Peking University, Yanyuan Chen, Hanpin Wang Peking University, Yu Huang Peking University
Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities in interpreting single medical images. However, real-world clinical diagnosis is intrinsically a multi-view process, requiring the synthesis of information across volumetric slices, temporal sequences, and comparative modalities. Existing benchmarks fail to capture this complexity, limiting the assessment of models in realistic clinical workflows. To bridge this gap, we introduce MedMultiBench, the first large-scale benchmark specifically designed for medical multi-image understanding. Comprising 11,392 expert-curated samples, MedMultiBench evaluates MLLMs across four distinct dimensions: Joint Reasoning, Comparative Analysis, Comprehensive Perception, and In-Context Learning. We benchmark 13 state-of-the-art MLLMs, revealing that while current models excel in single-view tasks, they struggle significantly with multi-image contexts. Our experiments identify a performance degradation in open-source models when processing increased visual loads, whereas closed-source models demonstrate better scalability. MedMultiBench provides a robust framework to facilitate the development of MLLMs capable of holistic clinical reasoning.
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
Yucheng Zhou University of Macau, Junwei Sheng University of Macau, Qianning Wang, Jianbing Shen University of Macau
Supervised Fine-Tuning (SFT) is the predominant paradigm for aligning large language models (LLMs), yet it suffers from optimization instability and limited generalization. Recent work attributes this issue to pathological gradient scaling and proposes Dynamic Fine-Tuning (DFT) to correct it at the token level. However, DFT assumes all demonstrations are equally suitable learning targets, an assumption violated by the strong heterogeneity of large-scale instruction data, where demonstration-policy mismatch induces high-variance updates at the sample level. We introduce Compatibility-Aware Dynamic Fine-Tuning (CADFT), a principled extension of DFT that controls sample-level optimization variance. CADFT derives a dynamic, policy-dependent compatibility signal from model likelihoods to modulate supervised updates, suppressing high-variance gradients from incompatible demonstrations. We further propose a delayed, low-frequency compatibility-guided rewriting strategy to transform persistently incompatible demonstrations into learnable targets. We show that CADFT can be interpreted as a variance-controlled estimator that generalizes token-level stabilization in DFT to the sample level. Extensive experiments demonstrate improved stability, generalization, and cold-start reinforcement learning initialization, while remaining fully supervised and free of reward modeling.
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
Yuquan Wang Fudan University, Mi Zhang Fudan University, Yining Wang Fudan University, Geng Hong Fudan University, Mi Wen Shanghai University of Electric Power, Xiaoyu You East China University of Science and Technology, Min Yang Fudan University
Large Reasoning Models (LRMs) have demonstrated impressive performance in reasoning-intensive tasks, but they remain vulnerable to harmful content generation, particularly in the mid-to-late steps of their reasoning processes. Current defense methods, however, depend on costly fine-tuning and additional expert knowledge, which limits their scalability.In this work, we propose ReasoningGuard, an inference-time safeguard for LRMs.It injects timely safety aha moments during the reasoning process to guide the model towards harmless yet helpful reasoning.Our approach leverages the internal attention mechanisms of the LRM to accurately identify key points in the reasoning path, triggering safety-oriented reflections.To safeguard both the subsequent reasoning steps and the final answers, we implement a scaling sampling strategy during decoding to select the optimal reasoning path.With minimal additional inference cost, ReasoningGuard effectively mitigates four types of jailbreak attacks, including recent ones targeting the reasoning process of LRMs. Our approach outperforms nine existing safeguards, providing state-of-the-art defenses while avoiding common exaggerated safety issues.
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
Terry Jingchen Zhang Vector Institute, Gopal Dev EuroSafeAI, Max-Planck Institute and Jinesis Lab, University of Toronto & Vector Institute, Ning Wang ETHZ - ETH Zurich, Max Obreiter ETHZ - ETH Zurich, Punya Syon Pandey University of Toronto, Keenan Samway Max Planck Institute for Intelligent Systems, Max-Planck Institute, Wenyuan Jiang ETHZ - ETH Zurich, Yinya Huang ETH Zurich 等 11 位作者
Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.
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
Jiale Cheng, Yusen Liu, Xinyu Zhang, Yulin Fei The Hong Kong University of Science and Technology, Wenyi Hong Department of Computer Science and Technology, Tsinghua University, Ruiliang Lyu Zhipu AI, Weihan Wang, Zhe Su 等 14 位作者
Large language models (LLMs) conventionally represent text as sequences of discrete tokens, making long-context scaling largely a matter of processing more tokens more efficiently.We instead explore a complementary direction: increasing how much original context each token represents.To this end, we introduce Glyph, a framework that renders long texts into compact visual pages and processes them with a vision-language model (VLM), allowing a fixed context window to cover substantially more text.To make visual compression practical, Glyph combines continual pre-training on rendered long-text data, an LLM-driven genetic search to identify rendering configurations that balance compression and task performance, and post-training with supervised fine-tuning and reinforcement learning.Across multiple long-context benchmarks, Glyph achieves 3–4× token compression while maintaining performance comparable to strong text-only LLMs such as Qwen3-8B, with over 4× faster prefilling and decoding and 2× faster supervised fine-tuning.Under more aggressive compression, a VLM with a 128K context window can handle tasks that would otherwise require up to 1M input tokens.Our code and model are released at https://github.com/thu-coai/Glyph.
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
Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resource languages. We attribute this issue as a mismatch gap between language-agnostic semantic understanding ability and language dominant safety alignment biased toward high-resource languages. Based on above insights, we empirically identify the semantic bottleneck in LLMs: intermediate layers in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Then, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwen3 Instruct models (7B–32B). Besides, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model’s language-agnostic semantic space.
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
Data selection is a critical aspect of Reinforcement Learning with Verifiable Rewards (RLVR) for enhancing the reasoning capabilities of large language models (LLMs). Current data selection methods are largely heuristic-based, lacking theoretical guarantees and generalizability. This work proposes a theoretically-grounded approach using influence functions to estimate the contribution of each data point to the learning objective. To overcome the prohibitive computational cost of policy rollouts required for online influence estimation, we introduce an off-policy influence estimation method that efficiently approximates data influence using pre-collected offline trajectories. Furthermore, to manage the high-dimensional gradients of LLMs, we employ sparse random projection to reduce dimensionality and improve storage and computation efficiency. Leveraging these techniques, we develop Curriculum RL with Off-Policy Influence guidance (CROPI), a multi-stage RL framework that iteratively selects the most influential data for the current policy. Experiments on models up to 7B parameters demonstrate that CROPI significantly accelerates training. On a 1.5B model, it achieves a 2.66x step-level acceleration while using only 10% of the data per stage compared to full-dataset training. Our results highlight the substantial potential of influence-based data selection for efficient RLVR.
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
Ning Wang Jiangnan University, Zhanyang Liu, Taotao Zhou, Xinrui Zhang, Zongru Shao Silicon Austria Labs, Haojie Zhou Jiangnan University
Aligning Large Language Models (LLMs) with diverse and potentially conflicting human values necessitates navigating complex multi-objective landscapes. However, existing prompt-conditioned approaches face a critical training-inference discrepancy: they rely on ground-truth scores during training while requiring manual user-specification at inference. We introduce prediction of implicit preferences to bridge this gap while reducing user burden. To this end, we propose Self-Guided Alignment (SGA), a framework that transforms passive reward dependency into an intrinsic adaptive sensing capability. It employs a dual-head architecture to unify preference internalization with conditional generation, enabling the model to learn a latent mapping between raw prompts and preference profiles. Through adaptive preference sensing, the model autonomously predicts the latent preference score to self-guide the generation, thereby eliminating the need for manual specification at inference. Extensive experiments across diverse model scales demonstrate that SGA often outperforms state-of-the-art baselines, achieving superior multi-objective trade-offs and improved preference alignment. Code is available at https://github.com/python-yyds/SGA.
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
Andrew Zhuoer Feng, Cunxiang Wang, Yu Luo Huawei Technologies Ltd., Lin Fan zhipuAI, Irene Zhou, Zikang Wang, Xiaotao Gu Zhipu AI, Jie Tang Tsinghua University, Tsinghua University 等 10 位作者
Evaluating the writing capabilities of large language models (LLMs) remains a significant challenge due to the multidimensional nature of writing skills and the limitations of existing metrics. LLM’s performance in thousand-words level and open-ended writing is inadequately assessed by traditional reference-based metrics or modern LLM-as-a-judge methods. We propose Tree-of-Writing (ToW), to resolve the implicit inconsistency often found when LLM-as-a-judge aggregates all sub-features in text evaluation. ToW incorporates a tree-structured workflow by explicitly modeling the aggregation weights of sub-features. We also present HowToBench, a large-scale Chinese writing benchmark encompassing 12 genres and 1302 instructions across three task categories: contextual completion, outline-guided writing, and open-ended generation. ToW successfully mitigates the biases, achieving a 0.93 Pearson correlation with human judgments. Furthermore, we detect that both overlap-based text generation metrics and popular LLM-as-a-judge practices are vulnerable to textual disturbances, while ToW is robust to them. We also uncover a negative correlation between input length and content-related scores in the Guide task, showcasing that it cannot be simply improved by input-side information piling.
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
Rui Hu Beijing Jiaotong University, Delai Qiu, Yining Wang, Shengping Liu, Jitao Sang Beijing Jiaotong University
Omni-modal large language models (OLLMs) offer a promising end-to-end solution for slide-enhanced speech recognition due to their inherent multimodal capabilities. However, we found a fundamental issue faced by OLLMs: Visual Interference, where models show a bias towards visible text over auditory signals, causing them to hallucinate slide content that was never spoken. To address this, we propose Visually-Anchored Policy Optimization (VAPO), which aims to reshape models’ inference process to follow the human-like “Look-then-Listen” inference chain. Specifically, we design a temporally decoupled policy: the model first extracts visual priors in a think> block to serve as semantic anchors, then generates the transcription in an answer> block. The policy is optimized via multi-objective reinforcement learning. Furthermore, we introduce SlideASR-Bench, a comprehensive benchmark designed to address the scarcity of entity-rich data, comprising a large-scale synthetic corpus for training and a challenging real-world test set for evaluation. We conduct extensive evaluations demonstrating that VAPO effectively eliminates visual interference and achieves state-of-the-art performance on SlideASR-Bench and public datasets, significantly reducing entity recognition errors in specialized domains.
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
Fuyuan Liu, Dianyu Yu, He Ren, Nayu Liu Tianjin University, Xiaomian Kang Institute of automation, Chinese academy of science, Delai Qiu, Fa Zhang, Genpeng Zhen 等 13 位作者
Reading order detection is the foundation of document understanding.Most existing methods rely on uniform supervision, implicitly assuming a constant difficulty distribution across layout regions. In this work, we challenge this assumption by revealing a critical flaw: Positional Disparity, a phenomenon where models demonstrate mastery over the deterministic start and end regions but suffer a performance collapse in the complex intermediate sections.This degradation arises because standard training allows the massive volume of easy patterns to drown out the learning signals from difficult layouts.To address this, we propose FocalOrder, a framework driven by Focal Preference Optimization (FPO).Specifically, FocalOrder employs adaptive difficulty discovery with exponential moving average mechanism to dynamically pinpoint hard-to-learn transitions, while introducing a difficulty-calibrated pairwise ranking objective to enforce global logical consistency.Extensive experiments demonstrate that FocalOrder establishes new state-of-the-art results on OmniDocBench v1.0 and Comp-HRDoc.Our compact model not only outperforms competitive specialized baselines but also significantly surpasses large-scale general VLMs.These results demonstrate that aligning the optimization with intrinsic structural ambiguity of documents is critical for mastering complex document structures.
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
Large Reasoning Models (LRMs) have achieved remarkable success on reasoning-intensive tasks such as mathematics and programming. However, their enhanced reasoning capabilities do not necessarily translate to improved safety performance—and in some cases, may even degrade it. This raises an important research question: how should we enhance the safety of LRMs? In this paper, we present a comprehensive empirical study on how to enhance the safety of LRMs through Supervised Fine-Tuning (SFT). Our investigation begins with an unexpected observation: directly distilling safe responses from DeepSeek-R1 fails to significantly enhance safety. We analyze this phenomenon and identify five key risky patterns that contribute to it. We then demonstrate that explicitly addressing these issues during the data distillation process can lead to substantial safety improvements. Next, we explore whether a long and complex reasoning process is necessary for achieving safety. Interestingly, we find that simply using short or template-based reasoning process can attain comparable safety performance. These findings prompt a deeper reflection on the role of reasoning in ensuring safety. Finally, we conduct a comprehensive ablation study to reveal the impact of different training configurations. Overall, we hope our empirical study could provide a more holistic picture on enhancing the safety of LRMs.
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