Transformer Attention Heads: Positional vs Symbolic Dynamics
A study on decoder-only Transformers (GPT-J) reveals that successful multi-hop reasoning requires the emergence of pure attention heads—either positional or symbolic. Two structurally equivalent tasks (number and letter reasoning) impose different mechanistic demands despite their equivalence.
Key facts
- Study uses GPT-J model
- Two tasks: number (positional) and letter (symbolic)
- Pure heads emerge during successful learning
- Tasks are structurally equivalent but require different head types
- Number task needs both positional and symbolic heads
- Letter task requires only symbolic heads
- Research aims to understand safe deployment of LLMs
- Published on arXiv (2605.31558)
Entities
Institutions
- arXiv