Study Finds Shorter Reasoning Can Maintain Accuracy in LLMs

A recent investigation published on arXiv (2608.03401) examines the impact of shortening reasoning duration in large language models (LLMs) on their accuracy. Researchers conducted paired tests on the same questions at equivalent reasoning intervals, utilizing 198 GPQA Diamond and 500 MMLU-Pro queries. They implemented a numeric/concision prompt that sets a token limit for Qwen3-14B and the trained configurations of gpt-oss-20b and -120b. The Qwen prompt led to a 12-17% reduction in reasoning traces, with minimal and varied accuracy shifts at the same token limits. Notably, a concise/early-answer directive improved MMLU-Pro accuracy by 3.8 percentage points at 512 tokens. The study indicates that shorter reasoning may enhance accuracy, challenging the notion that longer reasoning is always superior.