Search papers, labs, and topics across Lattice.
The paper introduces EvoResearcher, a training-free protocol that enables self-reflection and cost-bounded early stopping for large language models (LLMs) during inference. By employing a generate-self-critique-revise loop, the method allows a frozen LLM to self-verify its outputs without additional training, using four meta-reward components to guide the process. Validation on multiple reasoning benchmarks shows that EvoResearcher can maintain accuracy while significantly reducing the number of generations needed for responses, achieving early termination in 82-88% of cases.
Cost-bounded self-verification allows LLMs to self-correct efficiently, cutting down on response generations while maintaining accuracy.
Reinforcement-learning training of reasoning LLMs (e.g., GRPO) is expensive and requires a controllable environment, committing every contribution to a full training pipeline. We present EvoResearcher, a training-free, inference-time protocol that adds cost-bounded self-reflection to a single frozen LLM backbone. The protocol iterates generate -> self-critique -> revise until a maximum depth D is reached or the critique returns the CONFIRMED sentinel, an implicit early stop that lets the backbone self-verify its answer under a strict compute budget. Four self-reflective meta-reward components (correctness, efficiency, reflection depth, tool-call diversity) act as design principles instantiated as prompt-level mechanisms, so their benefits accrue with zero gradient updates. We validate the protocol on Big-Bench Hard (100 questions) and establish cross-domain behavior on GSM8K (500) and MATH (500) on the same frozen backbone, with cross-model replication on Qwen2.5-72B. All experiments use pure-reasoning benchmarks; the tool-call diversity component is validated in prompt-level form, and the environment-level and multi-agent extensions are design blueprints left to future work. On clean BBH the protocol does not raise accuracy beyond the 95% Wilson interval; its value is cost-bounded self-verification, with the CONFIRMED early stop terminating 82-88% of items at equal accuracy (about 2.1 generations per question).
