Reasoning LLMs Waste Compute, Degrade Hard-Problem Accuracy

Researchers from multiple institutions publish an arXiv paper, 'Thinking Harder, Not Smarter', showing that chain-of-thought reasoning LLMs often waste compute and can reduce accuracy when given more tokens. The study analyzes models including OpenAI's o1 series and finds excessive reasoning on easy problems and diminishing or negative returns on hard tasks. The results challenge naive test-time compute scaling and stress the need for adaptive compute strategies.