Large language models (LLMs) demonstrate remarkable reasoning capabilities, yet their performance often deteriorates sharply in long-horizon tasks, exhibiting systematic breakdown beyond certain scales. Conventional explanations primarily attribute this phenomenon to task complexity, such as combinatorial search explosion or long-term credit assignment challenges. In this work, we argue that these explanations are incomplete: even in linear, unbranched tasks without semantic ambiguity, autoregressive execution is subject to an intrinsic stability limit.
We propose that the fundamental constraint on long-horizon reasoning arises from process-level instability in autoregressive generation rather than solely from search or task complexity, reframing long-horizon reasoning as a problem of structural governance. We derive Theorem~A, showing that decision advantage in single-path autoregressive reasoning decays exponentially with execution length, imposing a fundamental bound on maintainable reasoning chains. This result implies a structural consequence: stable long-horizon reasoning requires discrete segmentation, naturally inducing graph-like execution structures such as directed acyclic graphs (DAGs).
Empirical studies in both synthetic environments and real TextWorld tasks reveal observable performance cliffs consistent with theoretical predictions. Our findings provide a dynamical perspective on long-horizon reasoning failure and suggest new limitations on maintaining long-term coherence under purely autoregressive architectures. Furthermore, we highlight that short-horizon evaluation protocols may obscure structural instability, indicating a potential shift from scaling toward structured governance in future reasoning systems.
The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
University of Cambridge(剑桥大学);Institute for AI, University of Stuttgart(斯图加特大学人工智能研究所);Max Planck Institute for Intelligent Systems(智能系统马克斯·普朗克研究所);ELLIS Institute Tübingen(图宾根ELLIS研究所);University of Southampton(南安普顿大学);Tübingen AI Center(图宾根人工智能中心)
Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution
School of Computer Science and Engineering, Northeastern University(东北大学计算机科学与工程学院);Department of Computer Science and Technology, Tsinghua University(清华大学计算机科学与技术系);China Mobile Online Services Co., Ltd.(中移在线服务有限公司)
SAGE: Scene Graph-Aware Guidance and Execution for Long-Horizon Manipulation Tasks
School of Artificial Intelligence, Shanghai Jiao Tong University(上海交通大学人工智能学院);ShanghaiTech University(上海科技大学);University of Sussex(Sussex大学);Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所)
Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning
Tsinghua University(清华大学);Institute of Microelectronics, Chinese Academy of Sciences(中国科学院微电子研究所);The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Nanyang Technological University, Singapore(新加坡南洋理工大学)
Autoregressive long-horizon prediction of plasma edge dynamics
Computational Sciences and Engineering Division, Oak Ridge National Laboratory(橡树岭国家实验室计算科学与工程部);Fusion Energy Division, Oak Ridge National Laboratory(橡树岭国家实验室等离子体能源部);Computer Science and Mathematics Division, Oak Ridge National Laboratory(橡树岭国家实验室计算机科学与数学部)
HERO: History-Enriched Rollout Training for Long-Horizon Autoregressive Neural Operators
School of Computer Science and Engineering, UESTC(电子科技大学计算机科学与工程学院);Electronics and Information Convergence Engineering, KHU(明知大学电子信息融合工程系);Department of Mathematical Sciences, UOL(利物浦大学数学科学系);School of Computer Science and Technology, XDU(西安电子科技大学计算机科学与技术学院);School of Mechanical and Electrical Engineering, UESTC(电子科技大学机械与电气工程学院);College of Computer and Information Engineering, XAUAT(西安建筑科技大学计算机与信息工程学院)
Diffusion-corrected Autoregressive Fourier Neural Operator for Droplet Evolution Prediction
New Jersey Institute of Technology(新泽西理工学院);University of Georgia(佐治亚大学);University at Buffalo(纽约州立大学布法罗分校);Hanyang University(汉阳大学);Northeastern University(东北大学)
Chain-of-Goals Hierarchical Policy for Long-Horizon Offline Goal-Conditioned RL
Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea(首尔国立大学电气与计算机工程系);Department of Automotive Engineering, Ajou University, Gyeonggi-do, South Korea(全州大学汽车工程系)
Autoregression-Free Neural Operators for Time-Dependent PDEs
School of Computer Science and Engineering, University of Electronic Science and Technology of China(电子科技大学计算机科学与工程学院);School of Robotics and Advanced Manufacture, Harbin Institute of Technology(哈尔滨工业大学机器人与先进制造学院);School of Mathematical Sciences, Capital Normal University(首都师范大学数学学院);College of Information and Control Engineering, Xi’an University of Architecture and Technology(西安建筑科技大学信息与控制工程学院);Laboratory of Intelligent Recognition and Image Processing, School of Computer Science and Engineering, Beihang University(北京航空航天大学智能识别与图像处理实验室);School of Computer Science and Technology, Tongji University(同济大学计算机科学与技术学院)