Reflexion: Language Agents with Verbal Reinforcement Learning
What this work does
Stores verbal self-reflections from evaluated trials in memory and conditions later attempts on them without gradient-based policy updates.
Connection to discovery
An essential memory-only baseline for distinguishing harness adaptation from test-time weight training.
Evidence
Inspected the actor/evaluator/reflection loop and ALFWorld, QA and programming experiments; some figure glyphs in the PDF are corrupted, while body text is readable.
Scope and limitations
The term verbal reinforcement learning does not imply parameter RL. An incorrect evaluator or misleading reflection can reinforce errors.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.