Directional Freezing in Language Models: Residual-Trajectory Metrics for Detecting Overconfident Agreement Dynamics
Under agreement pressure, a model can undergo geometric freezing: it stays active and confident while its residual or hidden-state update trajectory loses orthogonal novelty and grows increasingly confined to a persistent direction.
Authors
- Jonas Jakob Gebendorfer
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-06-13
- DOI
- https://doi.org/10.5281/zenodo.20677589
- Primary Topic
- Speech and dialogue systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00