Scaling Long-Horizon Agent via Context Folding

Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, Jiecao Chen
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:116894-116910, 2026.

Abstract

Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. Existing agent frameworks usually rely on manually defined context engineering pipelines, such as multi-agent or post-hoc summary. We introduce Context Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we propose FoldGRPO, an end-to-end reinforcement learning framework with specific process rewards to encourage effective task decomposition and context management. On long-horizon tasks, our agent matches the performance of baselines while using an active context up to 10x smaller, and significantly outperforms models constrained to the same context size. Code is available at https://github.com/sunnweiwei/FoldAgent.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-sun26x, title = {Scaling Long-Horizon Agent via Context Folding}, author = {Sun, Weiwei and Lu, Miao and Ling, Zhan and Liu, Kang and Yao, Xuesong and Yang, Yiming and Chen, Jiecao}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {116894--116910}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/sun26x/sun26x.pdf}, url = {https://proceedings.mlr.press/v306/sun26x.html}, abstract = {Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. Existing agent frameworks usually rely on manually defined context engineering pipelines, such as multi-agent or post-hoc summary. We introduce Context Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we propose FoldGRPO, an end-to-end reinforcement learning framework with specific process rewards to encourage effective task decomposition and context management. On long-horizon tasks, our agent matches the performance of baselines while using an active context up to 10x smaller, and significantly outperforms models constrained to the same context size. Code is available at https://github.com/sunnweiwei/FoldAgent.} }
Endnote
%0 Conference Paper %T Scaling Long-Horizon Agent via Context Folding %A Weiwei Sun %A Miao Lu %A Zhan Ling %A Kang Liu %A Xuesong Yao %A Yiming Yang %A Jiecao Chen %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-sun26x %I PMLR %P 116894--116910 %U https://proceedings.mlr.press/v306/sun26x.html %V 306 %X Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. Existing agent frameworks usually rely on manually defined context engineering pipelines, such as multi-agent or post-hoc summary. We introduce Context Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we propose FoldGRPO, an end-to-end reinforcement learning framework with specific process rewards to encourage effective task decomposition and context management. On long-horizon tasks, our agent matches the performance of baselines while using an active context up to 10x smaller, and significantly outperforms models constrained to the same context size. Code is available at https://github.com/sunnweiwei/FoldAgent.
APA
Sun, W., Lu, M., Ling, Z., Liu, K., Yao, X., Yang, Y. & Chen, J.. (2026). Scaling Long-Horizon Agent via Context Folding. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:116894-116910 Available from https://proceedings.mlr.press/v306/sun26x.html.

Related Material