reflexion

Why we included this project

This repo is the official NeurIPS 2023 release of Reflexion, and it is the best place to start if you want to study or adapt one of the more influential agent-improvement ideas in recent LLM research. The core idea is simple: a language agent reviews its own reasoning trace after a failed attempt, writes a short verbal reflection on what went wrong, and uses that text to do better next time. It is a lightweight alternative to gradient-based reinforcement learning because it works entirely through plain text prompts. The code covers question-answering over HotPotQA sets and decision-making in the AlfWorld text environment, with notebooks that let you compare agent types such as ReAct and chain-of-thought under different reflection strategies. Treat it as research scaffolding rather than a production service; the value is in running the notebooks, reproducing the reported results, and stealing the self-reflection loop as a template for your own agent experiments.