LLM Reasoning Process Revealed Through Attention Layer Analysis
A new study on arXiv (2603.29735v2) investigates how large language models transition from token-level information to abstract relational structures during reasoning. By analyzing attention heads and layers in autoregressive reasoning, researchers found a consistent layer-wise division of labor: outer layers preserve and route input-related features, while middle layers reorganize them into transferable rule-level representations. This is supported by representation geometry showing middle-layer states occupy lower-dimensional manifolds and align across disjoint vocabularies with the same symbolic rules. Causal interventions confirm that removing middle-layer components causes larger downstream changes. The paper focuses on mathematical and symbolic reasoning tasks.
Key facts
- Study analyzes attention heads and layers in LLMs during reasoning
- Outer layers preserve input features; middle layers reorganize into rule-level representations
- Middle-layer states occupy lower-dimensional manifolds
- Middle layers show stronger alignment across disjoint vocabularies with same symbolic rules
- Removing middle-layer components causes larger downstream changes
- Tasks include mathematical and symbolic reasoning
- Paper is arXiv:2603.29735v2
- Research investigates internal stage where token information becomes abstract relational structure
Entities
Institutions
- arXiv