EvoResearcher Enables Training-Free Self-Reflection with Early Stopping

EvoResearcher provides a training-free protocol for frozen LLMs to perform cost-bounded self-reflection using a generate-critique-revise loop. It introduces a CONFIRMED sentinel for early stopping, improving performance on GSM8K and MATH without any gradient updates.