Tag
This paper investigates reasoning in LLMs as an intrinsic dynamical process, finding that inference-time representations self-organize into low-dimensional manifolds. It proposes a label-free diagnostic based on internal dynamics to assess reasoning quality, suggesting that effective reasoning is governed by geometric and informational constraints.
Neural networks appear to speak English on the surface, but internally organize information in geometric space (curves, loops, surfaces, manifolds). Understanding "neural geometry" may be the key to understanding, debugging, and controlling models.