Fluid Representations in Reasoning Models
Source: https://arxiv.org/html/2602.04843v1 Author: Various (arXiv) Date: 2026-02-12
Summary
Paper studying how reasoning models develop “fluid representations” — abstract action patterns that focus on structure rather than surface features. Key finding: well-trained reasoning models develop representations that generalize across different surface forms of the same problem.
Key Claims
- Fluid representations: rather than storing “rule: when you see X do Y,” the model learns “when structure A occurs, apply transformation B” — independent of surface symbols.
- Evidence: same reasoning model handles algebra problems written as equations, word problems, or visual diagrams with similar performance.
- Contrast: poorly-trained models are “brittle” — high performance on training distribution, low performance on different surface forms.
- Mechanism: fluidity emerges from training on diverse surface forms of the same underlying structure.
- Implication for training: diversity in problem presentation matters as much as problem difficulty.
Concepts
- Test-Time Compute — fluid representations enable genuine test-time reasoning
- Mechanistic Interpretability — fluid vs. brittle representations are interpretability findings
- Synthetic Data — diverse surface forms require synthetic data generation