Meta 新论文提出 RSD 方法，让同一模型在任务结束后回看每一步、知道结果如何，从而给每个动作打分并从分数中学习，无需人工标注对错或奖励。ScienceWorld 上 Qwen3-14B 成功率从 12.9% 升至 20.6%，ALFWorld 上 Qwen3-8B 达 76.6%，超过所有对比方法，包括使用奖励或 ground truth 的方法。
New Meta paper finds that AI agents can learn from their own past attempts at tasks, even ones nobody marked right or wrong, by looking back at each step once they know how it ended.
Most agent runs come with no correct answer and no reward, which standard training relies on. RSD has the same model reread each finished run, knowing how it ended and what worked on similar tasks. It scores every action, and the agent learns from those scores.
On ScienceWorld, Qwen3-14B's success rate rose from 12.9% to 20.6%. On ALFWorld, Qwen3-8B hit 76.6%, beating every other method tested, even ones given rewards or ground truth.
Keep your agent's unlabeled runs, failures included, and train it on a hindsight review of them. Rebuild the review each round, since frozen notes left ALFWorld success at 59.7% by round 3, versus 67.9% with fresh ones.
来源：Rohan Paul · x.com