Reflexion: agents that learn from mistakes without gradient updates
In one sentence MIT and Northeastern propose Reflexion: agents that self-reflect in natural language after each failure, accumulating insights in episodic memory without modifying weights.
When we make a mistake, we don't rewrite our brain — we reflect on it, understand what went wrong, and do better next time. Reflexion brings this same capability to AI agents.
After each failed attempt, the agent generates a verbal reflection in natural language: "I was wrong because I assumed X, but actually Y." This reflection is saved in episodic memory and used as context in subsequent attempts.
The result is an agent that improves trial after trial without any fine-tuning: on HotpotQA it goes from 33% to 51% success, on AlfWorld from 42% to 97%. It's reinforcement learning in natural language, without a gradient in sight.
Companies
MIT, Northeastern University, Princeton University
Tools
Reflexion, GPT-4, GPT-3.5
Tags
Sources