Self-Verification Dilemma: Experience-Driven Suppression of Overused Checking in LLM Reasoning

ABSTRACT

Recent advancements in agentic reasoning rely heavily on self-verification loops, where models constantly double-check their own logic. However, this often leads to excessive computational waste and "verification paralysis" on simple tasks. This paper introduces an experience-driven framework that allows LLMs to dynamically suppress their verification mechanisms based on task confidence and historical success rates, drastically improving inference efficiency.