The Geometry of Reasoning Failure: Detecting Scale Boundaries in LLM Latent Trajectories

Andrey E Ustyuzhanin, Nikita Kazeev, Maxim Borisyak

Submission Type: I want my submission to be considered for both oral and poster presentation.

Keywords: LLM, latent space, Scale boundary detection, Latent trajectory geometry, Linear probing of abstraction, directional consistency, Topological compression

TL;DR: We detect when an LLM’s scientific reasoning hits a scale boundary by monitoring geometric breakdowns in its hidden-state trajectory, enabling timely, automatic shifts in abstraction before it gets stuck in local loops.

Confirmation Of Submission Requirements: I submit an abstract. It uses the template provided on the submission page and is no longer than 2 pages.

Submission Number: 427