Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B

Part of

LLM Self-Distillation For Enhanced Reasoning