EvoResearcher Enables Training-Free Self-Reflection and Early Stopping for LLMs

Training large language models (LLMs) for complex reasoning tasks, especially with reinforcement learning, is computationally intensive and requires controlled environments. This process commits every improvement to a full training pipeline, which can be inefficient. A new protocol, EvoResearcher, offers an alternative by introducing training-free, inference-time self-reflection to a single, frozen LLM backbone.EvoResearcher operates through an iterative process: the LLM generates a response, then self-critiques it, and subsequently revises it. This cycle continues until a predefined maximum depth is reached or the critique mechanism signals "CONFIRMED," indicating self-verification and triggering an early stop. This implicit early stopping mechanism allows the LLM to self-verify its answer within a strict compute budget.The protocol incorporates four self-reflective meta-reward components—correctness, efficiency, reflection depth, and tool-call diversity—instantiated as prompt-level mechanisms. These design principles accrue benefits without requiring any gradient updates. Validated on benchmarks like Big-Bench Hard, GSM8K, and MATH, EvoResearcher maintains accuracy comparable to full reasoning while terminating 82-88% of items early, averaging about 2.1 generations per question. This significantly reduces inference costs for complex reasoning tasks.