aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

ReHarness study uses task outcomes to revise agents’ old notesMachine translation

AI introduction

The MIT and Meta ReHarness study addresses agents’ reuse of outdated notes. After each task, it narrows or corrects retrieved notes using the outcome and only facts from each note’s original task. Across three benchmarks and three models, its average normalized gain was 0.240, compared with 0.168 for the best raw-trajectory baseline.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

– https://t.co/4rwDdFxaAc

标题:“用于智能体持续学习的循环式框架”

ReplyRohan Paul@rohanpaul_ai
Translation
麻省理工学院和 Meta 的一篇新论文发现,如果根据旧笔记下次被使用时发生的情况对其进行修正,智能体记忆的效果会更好。 在执行任务中学习的智能体会保存过去任务的笔记,并再次使用它们。笔记一旦写下,就很少改变,因此过于宽泛或过时的经验会不断误导智能体。 每项任务结束后,ReHarness 会将检索到的笔记和任务结果发送给一个修订智能体。它仅使用各条笔记原始任务中的事实,缩小或修正笔记的表述。 在 3 项基准测试和 3 个前沿模型上,它的归一化增益平均为 0.240;这一指标表示相对于无记忆智能体,实际获得的改进占可能改进的比例。表现最好的原始轨迹基线得分为 0.168;在使用 Claude 的 CL-Bench 测试中,重放完整原始历史所用的 token 数量是其 3.4 倍。 如果你的智能体从经验中学习,就让每次结果修改影响该结果的笔记,而不只是添加新笔记。
Original

New MIT and Meta paper finds that agent memory works better when old notes get corrected using what happened the next time they were used. Agents that learn on the job save notes from past tasks and reuse them. Once written, those notes rarely change, so an overly broad or outdated lesson keeps misleading the agent. After each task, ReHarness sends the retrieved notes and the result to a revision agent. It narrows or corrects each note, using only facts from that note's original task. Across 3 benchmarks and 3 frontier models, it averaged 0.240 in normalized gain, the share of possible improvement over a memoryless agent. The best raw-trajectory baseline scored 0.168, and replaying full raw history used 3.4× as many tokens on CL-Bench with Claude. If your agent learns from experience, let each outcome edit the notes that shaped it, not only add new notes.

View this post on X ↗

Source: Rohan Paul · x.com

Research
Found an error? Send a correction