arXiv cs.CL· Kazuki Nakayashiki·· 12 小时前AI 评分42
研究揭示 LLM 撰写的经验教训会把偶然性带入智能体下一次决策
Lessons from Luck: Do LLM-Written Lessons Carry Chance into an Agent's Next Decision?
AI 导读
一项预注册研究在 256 个合成世界中测试发现，当更可靠工具偶然失败后，模型据此写出的经验教训会让同一模型的新会话在 11.6% 的下一次运行中选用更差的工具，而使用无关记忆项或从成功结果写出的教训时该比例仅为 0.4%（交互效应 +0.112，95% 区间 [+0.097, +0.126]）。
正文
Abstract:Self-improving agents write lessons from the outcomes of their runs; when outcomes are noisy, a lesson can carry chance into later decisions. We measure one link of this chain in a registered study. In 256 synthetic worlds, each a recurring task with two interchangeable tools of recorded reliability, a scripted previous run used the more reliable tool (a good decision) or the less reliable one (a bad decision), with an assigned outcome: success or a transient failure. A model wrote a lesson from the run with ReasoningBank's memory-writing instruction; a fresh session of the same model, holding the authoritative log and the previous result, chose the next run's first tool with one memory item: that lesson, the lesson from the other outcome, an irrelevant item, or an unreasoned directive. The protocol was frozen, deposited and timestamped before any confirmatory call; 11 LLM endpoints gave 90,084 answers. After a chance failure of the better tool, the panel's lessons carried the chance (registered reading luck-carrying; interaction +0.112, 95% interval [+0.097, +0.126]): readers named the worse tool in 11.6% of next runs with the lesson written from that failure, against 0.4% with an irrelevant item and 0.4% with the lesson written from a success. The effect was +0.042 where the log stated that failures are independent and +0.181 where it did not (descriptive). After a lucky success of the worse tool, the interaction was bounded below 0.10 (+0.002), a bound, not evidence of no harm. In exploratory readings, 6 endpoints carried the chance (up to +0.352); 5 were bounded below 0.10. In secondary comparisons, a three-sentence decision-quality note in the writing instruction lowered the worse-tool rate after a chance failure to 7.5%, and one instruction for both outcomes with a no-change option to 3.1%; neither met the registered harm-reduction rule at every such endpoint.
| Comments: | 28 pages, 1 figure, 11 tables. A registered study: the protocol, code and stimuli were frozen, deposited (doi:https://doi.org/10.5281/zenodo.23253562) and timestamped in a Bitcoin block before the confirmatory run. The locked run records are public (doi:https://doi.org/10.5281/zenodo.23272809) |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Information Retrieval (cs.IR) |
| Cite as: | arXiv:2610.33029 [cs.LG] |
| (or arXiv:2610.33029v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33029 arXiv-issued DOI via DataCite |
Submission history
From: Kazuki Nakayashiki [view email]
[v1]
Mon, 12 Oct 2026 03:17:10 UTC (82 KB)
来源：arXiv cs.CL · arxiv.org