Reasoning emerges from constrained inference manifolds in large language models
This work provides a geometric and informational framework for understanding reasoning in LLMs, offering a complementary approach to benchmark-centric evaluation.
The study investigates reasoning in large language models as an intrinsic dynamical process, finding that effective reasoning requires a constrained structural regime with adequate representational expressivity, spontaneous manifold compression, and preservation of non-degenerate information volume. They introduce a label-free diagnostic computed from internal dynamics.
Reasoning in large language models is predominantly evaluated through labeled benchmarks, conflating task performance with the quality of internal inference. Here we study reasoning as an intrinsic dynamical process by examining the evolution of internal representations during inference. We find that inference-time dynamics consistently self-organize into low-dimensional manifolds embedded within high-dimensional representation spaces. we find that such geometric compression, although pervasive, is not sufficient for stable or reliable reasoning. Instead, effective reasoning dynamics emerge within a constrained structural regime characterized by three conditions: adequate representational expressivity, spontaneous manifold compression, and preservation of non-degenerate information volume within the compressed subspace. Models outside this regime exhibit characteristic pathological inference dynamics. Based on these insights, we introduce a unified, label-free diagnostic computed solely from internal dynamics. These findings suggest that reasoning in LLMs is fundamentally governed by geometric and informational constraints, offering a complementary framework to benchmark-centric assessment.