Dominic GriecoA working record
Back to research

research / Published

Feedback is not proof.

Reading Reflexion and thinking about the gap between a useful memory and a verified result.

By Dominic Grieco.

Reading: Reflexion: language agents with verbal reinforcement learning

Reflexion is a framework by Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. This note considers the gap between a useful reflection and a verified result.

What the paper reports

The NeurIPS 2023 paper describes agents that keep written reflections on feedback for later trials, rather than updating model weights. The authors report improvements in their evaluated tasks.

Interpretation

Store the test result alongside the explanation. A model can produce a plausible explanation without establishing the result.

Limits and open questions

The reported results do not establish that any agent memory system will improve, or that a reflection is a reliable account of what happened.

Sources and context

Paper
Reflexion: language agents with verbal reinforcement learning
Paper authors
Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao
Read at