IC-721The 72-head entity tracking circuit identified in Llama-7b achieves high faithfulness in Vicuna-7b and Goat-7b without any modification to the circuit graph

Nikhil Prakash, Tamar Rott Shaham, Tal Haklay, Yonatan Belinkov, David Bau

SourceFine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking

Using path patching, the authors identify a sparse circuit of 72 attention heads in four groups (value fetcher, position transmitter, position detector, structure reader) in Llama-7b that reproduces its full entity tracking accuracy. When this identical circuit is mean-ablated in Vicuna-7b and Goat-7b, it recovers 97% and 89% of their respective full-model performance, respectively. Random circuits of the same size achieve near-zero accuracy, confirming the circuit is not a generic set of heads. This demonstrates that fine-tuning preserves the base model's entity tracking circuit rather than replacing it.

Evidence

interventional

Key metric

faithfulness: Llama-7b 1.00, Vicuna-7b 0.97, Goat-7b 0.89; circuit accuracy: Llama-7b 0.66, Vicuna-7b 0.65, Goat-7b 0.73; full-model accuracy: Llama-7b 0.66, Vicuna-7b 0.67, Goat-7b 0.82; random circuit accuracy 0.00-0.01

Caveat

Achieving the full performance of Goat-7b requires incorporation of additional components beyond the Llama-7b circuit; the Goat-7b circuit is larger (175 heads) and approximately a superset of the Llama-7b circuit.

Model

LLaMA Llama 7B, Vicuna Vicuna-7B, GOAT-7B

Methods

Path Patching [primary], Mean Ablation / direct effect mean ablation / Mean token ablation [primary]

Related findings

IC-722, IC-723

Extraction

automatic-extraction