Abstract representational geometry supports inference in large language models

This paper investigates whether large language models (LLMs) perform complex reasoning through abstract structural inference—similar to how the human brain uses hippocampal cognitive maps—or if they rely primarily on surface-level statistical associations. The authors seek to determine if the "black box" of LLM reasoning contains identifiable, population-level geometric structures that support generalizable task adaptation.

To test this, the authors adapted a contextual reversal-learning paradigm into a text-based format. In this task, agents must infer a latent rule (context) from sparse feedback and apply it to novel stimuli. The researchers compared human performance with that of various LLMs (ranging from open-weight models like LLaMA to proprietary models like GPT-5.1). They analyzed the models' internal hidden states using metrics from systems neuroscience, specifically evaluating Cross-Condition Generalization Performance (CCGP) and Parallelism Scores (PS) to quantify whether the models organize task variables into orthogonal, disentangled manifolds.

The study reveals a striking convergence between biological and artificial intelligence. When LLMs successfully infer latent task structures, their internal states form abstract geometric structures that resemble the hippocampal manifolds observed in human reasoning. This geometry is not uniform; the models exhibit a functional hierarchy where lower layers encode stimulus identity, while higher layers form a specialized band enriched for abstract context geometry. Furthermore, the authors provide interventional evidence: training models on task sequences promotes geometric disentanglement, and applying geometric regularization to higher layers significantly boosts the models' capacity for generalizable inference.