[Meta-RL] We told an AI agent 'you can fail 3 times.' Accuracy went up 19%.

Three independent research teams (AI2, EPFL, Tsinghua) developed Meta-Reinforcement Learning with Self-Reflection, a pattern that gives AI agents multiple attempts per problem and has them reflect on each failure before retrying. Results showed accuracy gains up to 19.3% on QA benchmarks, with reflection-only memory outperforming full trajectory history, and the approach generalizing beyond training-time attempt counts without requiring weight updates at inference.