Superficial Reflection or Genuine Thought? A Fine-Grained Cognitive Analysis of Large Reasoning Models
This research creates a detailed taxonomy to analyze how Large Reasoning Models (LRMs) think by mapping their reasoning steps to seventeen fine-grained human cognitive processes. Using an annotation framework called CAPO, researchers analyzed over 277,000 reasoning steps, revealing that LRM 'double-checks' are often superficial. The findings suggest improving model performance requires focusing on high-quality information organization and comprehensive multi-step reflection.