Reflexion: Language Agents with Verbal Reinforcement Learning

Converted scalar RL rewards into verbal reflective feedback stored in episodic memory, enabling agents to self-improve across trial attempts without weight updates.

Directly guides architectural decisions, alignment strategy, and serving infrastructure for task-autonomous-agents, task-code-generation.

Assumptions

• Empirical distribution regularity holds and target domain adheres to pretraining linguistic/visual support

Limitations

• Resource scaling, inference memory requirements, and alignment robustness vary with model size and hardware topology