Unverified50% confidenceFactExact time
Reflective Recovery is a self-supervised method that converts the model's failed reasoning attempts into recovery training data.
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSelf-reflection is achieved by having the model evaluate the quality of its own output and iteratively improve it; the Reflexion framework demonstrates that an agent can learn from failed attempts and adjust its strategy in the next iteration.68% similarUnverified分布漂移是行为克隆的核心缺陷，因训练数据全部来自正确操作轨迹，模型未见过犯错后如何恢复的场景67% similarUnverifiedREVERSAL-BENCH provides a ground-truth reset oracle to determine whether a given state can truly be recovered.65% similarUnverifiedReAct范式通过强制要求每次工具调用后观察真实环境反馈，将推理链锚定在可验证的现实结果上，抑制错误累积64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931200API
curl https://kongchang.com/api/v1/knowledge/claims/931200MCP
get_claim(id=931200)