How Much Thinking is Enough? Quantifying and Understanding Redundancy in LLM Reasoning

In three linesStudy quantifying reasoning redundancy in LLMs: 61-93% of thinking steps can be truncated without affecting correct answers. Analysis across 4 frontier models and 2 math benchmarks (MATH-500). Redundancy is structural, stemming from length-agnostic outcome rewards, not model-specific artifact.

Summary generated by Claude — human-verified