cs.AI, cs.LG

The Geometry of Thought: How Scale Restructures Reasoning In Large Language Models

arXiv:2601.13358v2 Announce Type: replace-cross
Abstract: Scale does not uniformly improve reasoning – it restructures it. Analyzing 25,000+ chain-of-thought trajectories across four domains (Law, Science, Code, Math) and two scales (8B, 70B parameter…