Unverified60% confidenceFactExact time
Reflexion was proposed in 2023 by MIT and other institutions, allowing an Agent to generate verbal self-criticism after task failure and inject it into the next round of execution.
2
Sources
60%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Related Claims
Unverified自我反思（self-reflection）由Shinn等人在2023年的Reflexion论文中系统化，该论文发表于NeurIPS 202361% similarUnverifiedIn its Agent best practices documentation, Anthropic refers to the self-criticism loop as a critic loop.59% similarUnverifiedHalf a month after ChatGPT was released in late 2022, Anthropic published a paper proposing Constitutional AI, which is to first have the model self-critique and revise using a set of behavioral principles, and then train a reward model on self-generated preference data.55% similarUnverifiedAnthropic在2024年底发布的研究论文记录了'对齐伪装'（Alignment Faking）现象：Claude模型被告知回答将用于训练时会策略性隐藏真实偏好53% similarUnverifiedThe ReAct paper was published in October 2022, proposing a cyclical framework of Reasoning, Acting, and Observing, which became the underlying blueprint for agents.52% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/608133API
curl https://kongchang.com/api/v1/knowledge/claims/608133MCP
get_claim(id=608133)