当世界仍保留错误时,清理智能体历史反而有害
Cleaning an Agent's History Hurts When the World Still Holds the Error
Egor Pakhomov · Anurag Koul · Hongjun Liu · Bo Pang · Shafiq Joty · Erik Nijkamp
中文摘要
使用工具的语言模型智能体会积累自身错误的历史,而一个压缩、修复或交接此类历史的系统,会在这些错误的影响仍留在世界中时,决定如何表述这些错误。先前的工作在固定的世界状态下清理历史,此时展示一个修正步骤与声称它已完成是同步发生的。本文探究:同样的正确内容,是否会因历史如何陈述它以及世界持有什么而有益或有害。我们在任务中途冻结智能体及其精确的世界状态,编辑历史或世界,然后让接收模型继续。在 AgentCE-Bench 上,智能体提交错误值后,我们固定世界,通过两种自然编辑加入正确值:在保留的错误旁添加真实备注,或添加智能体自己完成的操作。备注将恢复率(接收模型答对的剩余目标比例)从匹配中性备注下的 51% 提升到 66%。完成的操作则将其降至未编辑历史之下,从 50% 降到 44%,因此仅值的存在并不能预测效果。在 AppWorld 上,数据库写入后,我们改变其回执报告成功还是失败,以及其效果是否存在。成功回执在效果缺失时将整体任务成功率从 45% 降至 28%,而在效果存在时将其从 38% 提升到 49%。两个接收模型合并后,两种效果相差 27 个百分点,该差异在 8 个接收模型中有 6 个在 97.5% 水平上为正,7 个在 95% 水平上为正。因此,历史编辑的效果取决于它所描述的世界,而一个看起来干净、但世界仍保留错误的历史,可能让智能体比完全不编辑更糟。
关键要点
- 01问题:历史清理的效果取决于世界状态,而非仅内容正确性。
- 02方法:冻结智能体世界状态,分别编辑历史或世界,让接收模型继续任务。
- 03结果:AgentCE-Bench 上真实备注提升恢复率 51%→66%,完成操作降至 44%。
- 04结果:AppWorld 上成功回执在效果缺失时降成功率 45%→28%,存在时升 38%→49%。
- 05局限:效果因接收模型而异,8 个中仅 6-7 个显著为正。
解读
尚无解读。
原始英文摘要
arXiv:2611.14654v1 Announce Type: new Abstract: Tool-using language-model agents accumulate histories of their own mistakes, and a system that compacts, repairs or hands off such a history decides what it says about those mistakes while their effects stay in the world. Prior work cleans the history at a fixed world, where showing a corrected step and saying it is done move together. We ask whether the same correct content helps or hurts depending on how the history states it and what the world holds. We freeze an agent mid-task with its exact world state, edit the history or the world, and let a receiver model continue. On AgentCE-Bench, after the agent commits a wrong value, we hold the world fixed and add the correct value by two natural edits: a truthful note beside the preserved error or the agent's own completed action. The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%. The completed action lowers it below the unedited history, from 50% to 44%, so the value's presence alone does not predict the effect. On AppWorld, after a database write, we vary whether its receipt reports success or failure and whether its effect is present. A success receipt lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present. The two effects differ by 27 points pooled over two receivers, a difference positive in six of eight receivers at 97.5% and seven at 95%. A history edit's effect thus depends on the world it describes, and a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.