New preprint: Transformers build separate internal representations for each interacting agent
AnnaCiaunica · x · 2026-11-03
Fernando Rosas (Imperial College) released the preprint "Beliefs in feedback: Factored representations of interacting agents in transformers". It shows that transformers predicting multiple interacting agents spontaneously learn factored internal representations, tracking each agent separately. Training visualizations show the internal geometry converging to the theoretical belief-state simplex (R² from 0.914 to 1.000) and to each agent's emission-probability structure (R² up to 0.996), suggesting emergent decomposed belief representations of the interacting parties.
More from Research
- Graph Isomorphism Algorithm Improved to n^O(log n), Exponent Drops from 2 to 1 — fortnow · 2026-11-03
- Researchers Wrote a Paper Assessing Matt Pocock's /retro Skill for Coding Agents — mattpocockuk · 2026-11-03
- Anthropic Releases FlashProve Prototype for Provable LLM Inference — scaling01 · 2026-11-03
- Why AI cracking Navier-Stokes doesn't mean it can auto-discover better models — snikolov · 2026-11-03
- STEP: surgical point tracking from sparse endpoints cuts error from 31.5 to 16.8 pixels — ftm_guney · 2026-11-03
- Allen Institute's OpenScope transforms into Global Brain Observatory, sharing brain-wide recordings openly — KordingLab · 2026-11-03