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

Most AI agents get one shot. They take a question, run a search or plan, give an answer, and move on. If the answer is wrong, that failure is lost. The agent starts fresh next time with no memory of what went wrong. Humans do not work this way. We fail, think about why, and try again with a better plan. From December 2025 to March 2026, three independent research teams at AI2, EPFL, and Tsinghua University arrived at the same idea. Give the agent multiple tries. Make it reflect on each failure. Feed that reflection into the next attempt. They call it Meta-Reinforcement Learning with Self-Reflection. Why single-shot agents fall short Standard RL-trained agents treat each attempt as independent. They cannot carry lessons from one try to the next. Three problems come together here. Sparse rewards make it hard to learn. The agent only gets a signal at the end (right or wrong), so it cannot tell which steps were good and which were bad. Independent tries mean the agent repeats the same mist