🤖 AI EngineeringID: reflexion-self-correction

Reflexion Self-Correction

Q: Detail the tripartite architecture of Reflexion: Actor, Evaluator, and Self-Reflection modules and their execution flow?

【Reflexion Self-Correction 核心考点 1】Detail the tripartite architecture of Reflexion: Actor, Evaluator, and Self-Reflection modules and their execution flow?。技术解析：Definition: Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/AI_Engineering/card/reflexion-self-correction

Q: Explain the FIFO sliding window eviction and vector-based semantic retrieval across Reflexion memory buffers?

【Reflexion Self-Correction 核心考点 2】Explain the FIFO sliding window eviction and vector-based semantic retrieval across Reflexion memory buffers?。技术解析：Definition: Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/AI_Engineering/card/reflexion-self-correction

Q: Why does unit-test grounded Reflexion boost HumanEval/LeetCode pass@1 from ~60% to over 90% across 3 trials?

【Reflexion Self-Correction 核心考点 3】Why does unit-test grounded Reflexion boost HumanEval/LeetCode pass@1 from ~60% to over 90% across 3 trials?。技术解析：Definition: Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/AI_Engineering/card/reflexion-self-correction

Q: Design prompt constraints to prevent models from authoring vague, ungrounded reflections that fail to guide self-correction?

【Reflexion Self-Correction 核心考点 4】Design prompt constraints to prevent models from authoring vague, ungrounded reflections that fail to guide self-correction?。技术解析：Definition: Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/AI_Engineering/card/reflexion-self-correction

Q: Compare Reflexion vs traditional Reinforcement Learning (PPO/DPO) in sample efficiency, inference cost, and cold-start feasibility?

【Reflexion Self-Correction 核心考点 5】Compare Reflexion vs traditional Reinforcement Learning (PPO/DPO) in sample efficiency, inference cost, and cold-start feasibility?。技术解析：Definition: Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self... 完整推导与工程落地见：https://talentme.airsota.com/resources/tech/roadmap/mindmap/AI_Engineering/card/reflexion-self-correction

Reflexion 自我反思与反思记忆

🎯Core Definition

Reflexion is an advanced agentic architecture that reinforces task performance through verbal self-reflection rather than traditional reinforcement learning weight updates; upon task execution failure evaluated by an objective judge (e.g., failed unit tests), the Self-Reflection module analyzes the entire trajectory, authoring a natural language critique detailing the failure root cause and strategic corrections; this reflection memory is appended to working memory, guiding the agent in subsequent trials to actively avoid past mistakes.

💡Use Cases

Autonomous code generation and test-driven self-repair (HumanEval, SWE-bench), multi-step scientific deduction, and strategic decision simulations.

⚡Key Problems Solved

Standard ReAct agents trapped in a bad state loop repeatedly repeat identical missteps across retries; Reflexion externalizes past mistakes into explicit textual memory buffers, enabling trial-over-trial error recovery and behavioral self-correction.

🎯5 High-Frequency Exam Points

Detail the tripartite architecture of Reflexion: Actor, Evaluator, and Self-Reflection modules and their execution flow?

Explain the FIFO sliding window eviction and vector-based semantic retrieval across Reflexion memory buffers?

Why does unit-test grounded Reflexion boost HumanEval/LeetCode pass@1 from ~60% to over 90% across 3 trials?

Design prompt constraints to prevent models from authoring vague, ungrounded reflections that fail to guide self-correction?

Compare Reflexion vs traditional Reinforcement Learning (PPO/DPO) in sample efficiency, inference cost, and cold-start feasibility?

📖 In-depth Guide:📄 agent-design-patterns →

Updated 2026-08-14

🎯

Test Your Knowledge: Practice Questions for "Reflexion Self-Correction"

Single choice pitfall questions with instant feedback and mistake tracking.

🚀 Start Card Practice ➔

← Previous CardReAct Thought-Action-Observation Next Card →Plan-and-Execute Pattern

🔗 More AI Engineering Knowledge Cards

Vector Distance Metrics & L2 Normalization Scalar Quantization (SQ8/SQ4)Product Quantization (PQ)ADC Asymmetric Distance Computation