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.
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.
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
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
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
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