IC-723The entity tracking performance gap between Goat-7b and Llama-7b is primarily attributable to enhanced positional information in the value fetcher and position transmitter heads
Nikhil Prakash, Tamar Rott Shaham, Tal Haklay, Yonatan Belinkov, David Bau
Using cross-model activation patching (CMap), the authors patch the output of specific head groups from Goat-7b into Llama-7b. Patching the value fetcher heads alone recovers the full performance of Goat-7b (0.82) in Llama-7b (baseline 0.66). Patching the position transmitter heads also yields a substantial performance boost. This localizes the improvement to enhanced encoding of positional information and object value in the fine-tuned model's shared circuit components, rather than to entirely new circuitry.
Evidence
interventional
Key metric
Llama-7b baseline accuracy 0.66; patching Goat-7b value fetcher heads recovers full Goat-7b performance (0.82); patching position transmitter heads yields substantial boost; Goat-7b value fetcher shows 20% improvement over Llama-7b
Caveat
The authors speculate the enhanced encoding stems from both additional components in the circuit and improved encoding in shared components, but do not fully disentangle these two sources.