circuit discovery

**Circuit discovery** is the **process of identifying interacting model components that jointly implement a specific behavior in a language model** - it aims to map behavior from outputs back to causal internal computation. **What Is Circuit discovery?** - **Definition**: Treats groups of heads, neurons, and residual pathways as functional subcircuits. - **Target Behaviors**: Common targets include induction, factual retrieval, and arithmetic-style reasoning. - **Method Stack**: Uses activation patching, ablation, attribution, and feature analysis together. - **Output Form**: Produces mechanistic hypotheses that can be tested with interventions. **Why Circuit discovery Matters** - **Causal Understanding**: Moves beyond correlation to identify which components are necessary. - **Safety Utility**: Helps locate pathways linked to harmful outputs or policy failures. - **Model Editing**: Enables targeted interventions instead of broad retraining. - **Debug Speed**: Narrows failure investigation to small internal regions. - **Research Progress**: Builds reusable knowledge about transformer computation patterns. **How It Is Used in Practice** - **Behavior Spec**: Define narrow behavior tests before searching for candidate circuits. - **Intervention Tests**: Validate circuit necessity with controlled patching and ablation experiments. - **Replication**: Check discovered circuits across prompts, seeds, and nearby checkpoints. Circuit discovery is **a core workflow for mechanistic transformer analysis** - circuit discovery is most useful when hypotheses are validated with explicit causal interventions.