Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can large language models (LLMs) notice unexpected behavior, such as heavier objects falling more slowly, and discover the law through experiments without being told what to investigate? To evaluate this ability, we introduce LongChang, an automated framework that creates simulated worlds with altered physics, where one or more laws differ from ordinary physics. The evaluated LLMs act as explorers, manipulating objects and building tools to investigate their own questions without research tasks or hints about these laws. Separate LLM agents generate and review scenes, review tool designs, and evaluate reports for correctness and experimental support using the altered physical laws and exploration records. These records guide scene evolution for further exploration, allowing the full cycle to run without human intervention. Among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%. The exploration records reveal uninvestigated anomalies and only partial identification of the altered physical laws.
Multi-agent reasoning systems adopt a generate-then-transfer paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, three frontier LLMs (Claude Opus 4.6, GPT-5.4, and GLM-5.2), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg.+7.3pp, max +22.4pp on HMMT2026; Claude Opus 4.6-High). Beyond these contributions, we discover a step-level scaling law: increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.
We study nonparametric regression using an over-parameterized two-layer neural network trained with algorithmic guarantees in this paper. We consider the setting where the training features are drawn uniformly from the unit sphere in $\RR^d$, and the target function lies in an interpolation space commonly studied in statistical learning theory. We demonstrate that training the neural network with a novel Preconditioned Gradient Descent (PGD) algorithm, equipped with early stopping, achieves a sharp regression rate of $\cO(n^{-\frac{2αs'}{2αs'+1}})$, where $α$ parametrizes the polynomial eigenvalue decay of the regular Neural Tangent Kernel (NTK), $λ_j\asymp j^{-2α}$, and $2α=d/(d-1)$ under the spherical uniform design considered here. The target function is assumed to belong to the interpolation space $\bth{\cH_K}^{s'}$ with $s'\ge3$. This rate is even sharper than the currently known nearly-optimal rate of $\cO(n^{-\frac{2αs'}{2αs'+1}})\log^2(1/δ)$~\citep{Li2024-edr-general-domain}, where $n$ is the size of the training data and $δ\in (0,1)$ is a small probability. Our analysis is based on two key technical contributions. First, we present a principled decomposition of the network output at each PGD step into a function in the reproducing kernel Hilbert space (RKHS) of a newly induced integral kernel, and a residual function with small $L^{\infty}$-norm. Second, leveraging this decomposition, we apply local Rademacher complexity theory to tightly control the complexity of the function class comprising all the neural network functions obtained in the PGD iterates. Our results show how PGD changes the effective kernel in the reference training dynamics, inducing an integral kernel with lower complexity than the regular NTK arising from vanilla GD.
Xuan He, Sining Zhoubian, Ruozhen Yang, Yuhao Wen, Zhineng Wang, Yuyang Hu, Ke Yang, Yuyu Zhang 等 11 位作者
Memory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multiple benchmarks and model backbones, we find many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value. Taking RAG as a reference, we dissect the net value of memory into the gains and losses over direct retrieval, termed memory benefit and memory tax. These gains and losses arise because memory process changes how the raw information is organized and accessed. We view such changes as the double-edged deviations from the raw history. They are originally designed and introduced to benefit the task, for example by reorganizing facts or distilling preferences, but some may also remove or obscure evidence, giving rise to memory tax. Since memory benefit reflects the expected gains of its design, we then focus more on the overlooked memory tax, study how it arises from the system and evaluation side, and generalize the concept to more application scenarios of memory. We hope the notion of memory tax can provide a new lens and shift more attention from standalone performance to the net value of memory, informing more reliable memory designs and evaluations. We will open-source the code and results to the community. Project Page: https://hexuan21.github.io/MemoryTax/
Zhen Yang, Wei Du, Jie Wang, Wenze Zhou, Xiangfeng Meng, Zhengyang Wang Anhui University, Suping Sun, Ziwei Du Anhui University 等 14 位作者
Recent progress in Large Language Model (LLM) based Table Question Answering (TableQA) has demonstrated strong performance on standard benchmarks. However, existing benchmarks mainly focus on well-structured tables and fail to reflect the irregular structures and complex reasoning commonly encountered in real-world scenarios. We propose CompTab, a benchmark designed to evaluate TableQA under complex reasoning and irregular table conditions. CompTab covers six representative types, including semantic ambiguity, multi-hop reasoning, transposed tables, merged cells, missing values, and outliers. It is constructed from real-world seed tables across multiple domains using controlled LLM based generation and human verification to ensure realism and diversity. In addition, to improve the generalization of LLMs under complex and irregular table settings, we propose a two-stage training framework that progressively aligns models with textual reasoning and executable decision signals, instantiated as CompTabLLM. Evaluations on 38 representative LLMs and CompTabLLM show clear limitations of existing LLMs under realistic conditions, while the proposed framework improves generalization. CompTab thus provides a challenging benchmark for advancing TableQA in real-world.
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
Minghan Zhang, Zhen Yang, Haodong Zou, Jie Chen Anhui University, Zhen Duan Anhui University, Shu Zhao Anhui University
Knowledge Base Question Answering (KBQA) leverages structured knowledge bases to offer superior interpretability and hallucination resistance, making it a critical technology for precise knowledge reasoning. However, the prevailing LLM-based generate-then-execute formulation of semantic parsing is limited by strict syntactic constraints, making it primarily prone to structural deviations that render queries unexecutable, while suffering from semantic deviations that yield incorrect execution results. To address these challenges, we propose the Execution as Verification (EVER) framework, reframing semantic parsing as an iterative, self-correcting reasoning process driven by execution feedback. First, motivated by the insight that query executability serves as a strong proxy for answer correctness, we introduce Fine-Grained Execution-Aware Planning. This mechanism decomposes complex semantic parsing into a sequence of stepwise reasoning processes oriented by executability verification, ensuring high query executability. We further design a Self-Guided Semantic Correction mechanism based on execution result verification, utilizing execution feedback to verify and calibrate semantic deviations, thereby ensuring the semantic correctness of executable queries. Experimental results on the WebQSP and CWQ datasets demonstrate that our method achieves significant improvements in both query executability and answer accuracy, achieving state-of-the-art performance, particularly in complex multi-hop scenarios. Our code is available at https://github.com/ahu-zmh/EVER.
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
Yuandong Wang Capital Normal University, Yao Cui, Yuxin Zhao, Zhen Yang, Yangfu Zhu, Zhenzhou Shao
Recent advances in Vision-Language Models (VLMs) have achieved impressive progress in multimodal mathematical reasoning. Yet, howmuch visual information truly contributes to reasoning remains unclear. Existing benchmarks report strong overall performance but seldom isolate the role of the image modality, leaving open whether VLMs genuinely leverage visual understanding or merely depend on linguistic priors. To address this, we present MathSight, a university-level multimodal mathematical reasoning benchmark designed to disentangle and quantify the effect of visual input. Each problem includes multiple visual variants—original, hand-drawn, photocaptured—and a text-only condition for controlled comparison. Experiments on state-of-the-art VLMs reveal a consistent trend: the contribution of visual information diminishes with increasing problem difficulty. Remarkably, Qwen3-VL without any image input surpasses both its multimodal variants and GPT-5, underscoring the need for benchmarks like MathSight to advance genuine vision-grounded reasoning in future models. The project page is available at https://cnu-bot-group.github.io/MathSight/.
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
Zhihao Xu, Fuzhen Yang, Liang Lin, Xiting Wang Renmin University of China
Recent advancements in Large Reasoning Models (LRMs) have showcased strong performance across various reasoning tasks by leveraging System-2 thinking capabilities. However, existing studies indicate that this reasoning ability alone does not reliably transfer to the general alignment domain. Inspired by cognitive science and how humans solve tasks, we argue that LRMs must be equipped with metacognitive knowledge to fully utilize their System-2 capabilities. In this paper, we propose Priority-Aware Metacognition (PAM), which guides the model to first identify the top-level human preference (e.g., harmlessness) as a means of understanding the alignment task’s nature, and then apply other kinds of metacognitive knowledge to better monitor and regulate the model’s thinking process. We implement PAM via a two-stage pipeline: a cold-start phase that collects structured metacognitive knowledge based on Flavell’s theoretical framework, and a preference-optimization phase that further reinforces such metacognition. Extensive experiments validate the effectiveness of PAM. Under the same training pipelines, PAM consistently yields higher performance, improving general domain alignment performance by ~10 points on the helpfulness and harmless benchmarks. Code is available at https://anonymous.4open.science/r/PAM-RM-02DF.
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
Zhen Yang, Ping Jian Beijing Institute of Technology, Zhongbin Guo, Zuming Zhang Nanyang Technological University, Chengzhi Li Beijing Institute of Technology, Yonghong Deng Beijing Institute of Technology, Xinyue Zhang, Wenpeng Lu
Over the past year, spatial intelligence has drawn increasing attention. Many prior works study it from the perspective of visual-spatial intelligence, where models have access to visuospatial information from visual inputs. However, in the absence of visual information, whether linguistic intelligence alone is sufficient to endow models with spatial intelligence, and how models perform relevant tasks with text-only inputs still remain unexplored. Therefore, in this paper, we focus on a fundamental and critical capability in spatial intelligence from a linguistic perspective: viewpoint rotation understanding (VRU). Specifically, LLMs and VLMs are asked to infer their final viewpoint and predict the corresponding observation in an environment given textual description of viewpoint rotation and observation over multiple steps. We find that both LLMs and VLMs perform poorly on our proposed dataset while human can easily achieve 100% accuracy, indicating a substantial gap between current model capabilities and the requirements of spatial intelligence. To uncover the underlying mechanisms, we conduct a layer-wise probing analysis and head-wise causal intervention. Our findings reveal that although models encode viewpoint information in the hidden states, they appear to struggle to bind the viewpoint position with corresponding observation, resulting in a hallucination in final layers. Finally, we selectively fine-tune the key attention heads identified by causal intervention to improve VRU performance. Experimental results demonstrate that such selective fine-tuning achieves improved VRU performance while avoiding catastrophic forgetting of generic abilities.
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
Zhen Yang The Hong Kong University of Science and Technology, Mingyang Zhang Alibaba Group, Feng Chen University of Adelaide, Ganggui Ding, Liang Hou Kuaishou Technology, Xin Tao Kuaishou, Ying-Cong Chen The Hong Kong University of Science and Technology
Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized—only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model’s KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks—e.g., +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0—while remaining highly efficient.
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
Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent 等 1137 位作者
In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.