Research map / Multi-agent systems and games / LLM agents
AgentFold: Long-Horizon Web Agents with Proactive Context Folding
Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, Pengjun Xie, Fei Huang and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AgentFold, a web agent that proactively folds its interaction context at multiple scales to avoid context saturation on long-horizon information-seeking tasks.
Problem
LLM web agents on long-horizon information-seeking tasks suffer from context saturation with raw ReAct histories, while fixed per-step full-history summarization risks irreversible loss of critical details.
Approach
Instead of accumulating raw histories or summarizing everything at each step, the agent proactively folds its context (a learned context-management behavior) so that it keeps compact, reversible-in-spirit representations of past interactions while preserving critical details.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use proactive context folding (AgentFold) instead of append-only ReAct context | web agent · context length at turn 100 | context tokens: over 84k tokens (92%) smaller | Improved |
| scale the turn limit up to 256 with folded context (ablation) | AgentFold-30B-A3B · BrowseComp | accuracy | Improved |
| increase turn limit beyond 64 for append-only context (ablation) | GLM-4.5-355B-A32B · BrowseComp | accuracy | No clear effect |
| train a 30B-A3B model with folding SFT | Qwen3-30B-A3B · BrowseComp | accuracy: 0.5 -> 36.2 | Improved |
| use AgentFold-30B-A3B in place of larger open models | DeepSeek-V3.1-671B-A37B · BrowseComp-ZH | accuracy: 49.2 -> 47.3 | Mixed |
Experimental setup
Baselines: ReAct-based agents fixed full-history summarization agents
Builds on: ReAct
Closest papers by meaning
- Scaling Long-Horizon Agent via Context Folding (ICML 2026 · similarity 0.88)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.87)
- AgentWebBench: Benchmarking Multi-Agent Coordination in Agentic Web (ICML 2026 · similarity 0.87)
- AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents (ICLR 2025 · similarity 0.87)
- InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
