Attractor Basin Dynamics in Language Models

Summarise with VIRAT
Authors: Andrew Smigaj
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

How do language models commit to interpretive states, and what happens when those states are wrong for the situation? This project studies attractor basin dynamics — stable geometric regions in residual stream space that predict model behavior before any output is generated. We identify basins through UMAP projection and clustering, track how they persist or release as context shifts, and measure where resistance to updating creates alignment failures. Current work establishes the methodology on polysemy and a safety-relevant probe in gpt-oss-20b. Future work extends to token-level basin transitions during reinterpretation, cognitive scaffolding effects on basin activation, and a MUD-based research environment generating naturalistic cognitive situations for AI agents while capturing their internal traces.

Fields of Study & Subjects

Explainable Artificial Intelligence (XAI) Topic Modeling Natural Language Processing Techniques

Publication details

Journal / source
OSF Preprints (OSF Preprints)
Publication date
29 March 2026
Language
EN
Document type
Preprint