arXiv cs.AI· Chenle Chen, Yangbo Wei, Zhijie Zheng, Zhenyu Su, Qifan Wang, Junhong Qian, Chao Yao, Chen Wu, Lei He·· 12 小时前AI 评分34
RECAP：为长程智能体实现风险受控的上下文自适应遗忘
When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents
AI 导读
研究提出 RECAP 框架，通过匹配上下文干预、执行关系预测与独立策略认证，决定长程智能体何时保留、压缩或删除失败历史。在 Hunyuan-A13B 上，它编辑了 43.24% 的保留检查点，同时匹配 KEEP 的 91.89% 续执行成功率，并将无条件平均续执行提示词成本降低 4,619.8 tokens（2.13%）。五个骨干模型上均产生正向配对成功节省估计，但可靠性-效率权衡因骨干而异。
正文 · 原文
Abstract:Long-horizon agents learn from failure, but the value of a failure record can change as execution progresses. We study when retaining, compressing, or removing the same history block improves subsequent execution relative to full-history KEEP. We introduce RECAP, a framework that combines matched context interventions, execution-relational prediction, and independent policy certification. Matched environment states and rollout seeds measure success and continuation prompt-cost differences; prefix-only relations between a failure block and the current execution propose promising edits. Independent task bundles then select a policy subject to harmful-intervention and non-positive-benefit risk constraints under a fixed measurement protocol. A five-seed longitudinal case holds both the failure block and its summary fixed: compression fails at an early state and becomes a cost-saving intervention at a later state. Across five backbones, RECAP produces positive paired-success saving estimates with backbone-specific reliability-efficiency tradeoffs. On Hunyuan-A13B, it edits 43.24% of retained checkpoints while matching KEEP's observed 91.89% continuation success and reducing unconditional mean continuation prompt cost by 4,619.8 tokens (2.13%). The framework provides an execution-level reference for deciding when to forget: retain history by its value to what comes next.
| Comments: | 32 pages, 5 figures, 14 tables |
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.22897 [cs.AI] |
| (or arXiv:2610.22897v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.22897 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Chenle Chen [view email]
[v1]
Fri, 16 Oct 2026 04:01:00 UTC (2,438 KB)
来源：arXiv cs.AI · arxiv.org