Context Compression Hurts Long-Horizon Agents at a Few Key Points, PAIR Pinpoints Why
dair_ai · x · 2026-10-06
核心发现
- 上下文压缩是长程 Agent 的重大瓶颈:研究发现在几个特定节点上,压缩会显著损害长时程 Agent 的表现。
- PAIR 评估方法:不再比较差异巨大的整段运行,而是让 Agent 从同一状态分别以「有压缩/无压缩」重放,实现可控对比。典型压缩会多增加几步操作,成功率的大幅下滑来自少数几次压缩事件。
- 两类有害压缩模式:
- 丢弃了 Agent 尚未解决的任务条件——例如在一个 Venmo 任务中,摘要丢掉了「仅限同事」的过滤条件,却把全部 36 笔付款的总额当作答案输出;
- 把 Agent 已读过的 API 规格压缩成模糊描述,导致 Agent 重新打开文档、重新登录,约多花五步。
- PAIR 的处理:诊断出压缩丢弃了哪些信息,并针对性重写对应的压缩内容。
对做 Agent 记忆与上下文工程的人来说,这是「压缩不是无害摘要」的实证警示。
More from coding & agent
- 70% of web traffic now comes from agents, says Dynamic Web Paige — DynamicWebPaige · 2026-10-06
- Dev tries CodeRabbit and Greptile, calls both 'complete trash' for PR review — KyeGomezB · 2026-10-06
- Kabir returns to Stripe, doubling down on agentic payments infrastructure — jeff_weinstein · 2026-10-06
- Trained agentic context management: 8K-context small model matches GPT-5.4 at 1M on OOLONG — xennygrimmato_ · 2026-10-06
- Granola now auto-captures screen shares in meetings, not just what's said — soleio · 2026-10-06
- One Agent's Silent Retries Were Eating 60% of His API Spend — Artistic-Earth8997 · 2026-10-06