Reflexion
Reflexion is an AI agent capability level that enables a model to perform self-evaluation and reflective correction of its own outputs. Through an internal review mechanism, it improves the quality of task completion.
Core Facts
Timeline (last 90 days)
ReAct, Plan-and-Execute, and Reflexion are three common agent control-flow patterns, differing in who decides the next step; they are essentially the harness's orchestration choices for loop structure, termination conditions, and feedback loops, not model capabilities.
The Reflexion pattern introduces a self-reflection step after an execution failure, writing the cause of failure as textual feedback for the next round of attempts; it is essentially maintaining a memory of experience at the harness layer.
ReAct, Reflexion, and Toolformer are the foundational papers for understanding agent design patterns.
Traditional frameworks such as ReAct and Reflexion encourage the Agent to explore and reflect as much as possible.
Ideas similar to generate-verify-regenerate also appear in academic methods such as Self-Refine and Reflexion.
Test-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
Reflexion-type agents have the model generate a self-evaluation report after each task completion and write the experience into memory for future reference.
Architectures such as Reflexion and Tree-of-Thought are designed to introduce systematic backtracking and verification capabilities in multi-step reasoning.
RACER's reflective loop design draws on self-examination mechanisms from frameworks such as ReAct and Reflexion.
Research from MemGPT, Reflexion, and others found that having agents explicitly generate experience summaries after task completion and write them back to memory can significantly improve cross-task transfer capability.
14 more timeline events
All Facts (20)
The Reflexion or Self-Refine mechanism was systematically proposed by Shinn et al. in a 2023 paper.
65%UnverifiedLLM-based Agent可以通过自然语言反思在极少次尝试中实现策略优化，这被学术界称为'反思型Agent'（Reflexion）架构
85%UnverifiedAI自我反思机制最早由Shinn等人在2023年提出的Reflexion框架中系统化阐述
85%Unverified2023年Noah Shinn等人提出的Reflexion框架首次系统化地论证了让LLM反思自身输出并迭代改进的有效性
85%Unverified斯坦福大学2023年发布了Reflexion框架，通过语言反馈而非梯度更新来实现Agent的学习
85%Unverified在Reflexion框架的编程任务中，加入自我反思循环后代码通过率从67%提升到了91%
80%UnverifiedReflexion机制已被证明能显著提升Agent在复杂任务中的成功率，部分实验显示成功率提升可达20%-30%
75%UnverifiedThe ReAct paradigm has spawned variants such as Reflexion, LATS, and Plan-and-Execute.
60%UnverifiedThe Reflexion mechanism allows an Agent to generate natural-language experience summaries after failures and reference them in subsequent decision-making.
60%UnverifiedThe Reflexion pattern introduces a self-reflection step after an execution failure, writing the cause of failure as textual feedback for the next round of attempts; it is essentially maintaining a memory of experience at the harness layer.
50%UnverifiedReAct, Plan-and-Execute, and Reflexion are three common agent control-flow patterns, differing in who decides the next step; they are essentially the harness's orchestration choices for loop structure, termination conditions, and feedback loops, not model capabilities.
50%UnverifiedReAct, Reflexion, and Toolformer are the foundational papers for understanding agent design patterns.
50%UnverifiedTraditional frameworks such as ReAct and Reflexion encourage the Agent to explore and reflect as much as possible.
50%UnverifiedIdeas similar to generate-verify-regenerate also appear in academic methods such as Self-Refine and Reflexion.
50%UnverifiedTest-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
50%UnverifiedReflexion-type agents have the model generate a self-evaluation report after each task completion and write the experience into memory for future reference.
50%UnverifiedArchitectures such as Reflexion and Tree-of-Thought are designed to introduce systematic backtracking and verification capabilities in multi-step reasoning.
50%UnverifiedRACER's reflective loop design draws on self-examination mechanisms from frameworks such as ReAct and Reflexion.
50%UnverifiedResearch from MemGPT, Reflexion, and others found that having agents explicitly generate experience summaries after task completion and write them back to memory can significantly improve cross-task transfer capability.
50%UnverifiedCurrent LLM-driven RSI practice relies on models' code generation and self-critique capabilities, with typical representatives including AlphaCode's iterative repair process and the reflection mechanism described in the Reflexion paper
50%