copying heads

**Copying heads** is the **attention heads that facilitate direct or indirect copying of tokens from prior context into output prediction pathways** - they are central to tasks that require exact string continuation and pattern reproduction. **What Is Copying heads?** - **Definition**: Heads route token identity information from source positions toward next-token logits. - **Use Cases**: Important in code, lists, names, and repeated-structure generation. - **Mechanism**: Often interacts with induction and residual stream composition components. - **Identification**: Detected via token-tracing experiments and copying-specific prompt tests. **Why Copying heads Matters** - **Behavior Insight**: Explains exact-match continuation strengths in language models. - **Safety Relevance**: Related to potential memorization and data leakage concerns. - **Performance**: Copying pathways can improve fidelity on structured tasks. - **Failure Modes**: Overactive copying can contribute to repetitive or context-locked outputs. - **Editing Potential**: Targetable mechanism for controlling copy bias in generation. **How It Is Used in Practice** - **Copy Benchmarks**: Use prompts requiring exact token carryover to measure head contribution. - **Causal Ablation**: Disable candidate heads and observe drop in exact-copy performance. - **Mitigation**: Apply targeted interventions if copying creates undesirable memorization behavior. Copying heads is **a central mechanistic pattern for context-token reuse in transformers** - copying heads provide a concrete bridge between attention dynamics and exact-sequence generation behavior.