IC-722Entity tracking in Llama-7b is implemented by detecting and transmitting the positional information of the correct entity, with distinct head groups for position detection, transmission, and value fetching
Using desiderata-based component masking (DCM), the authors show that 71.43% of group B heads and 70.0% of group C heads encode the positional information of the correct object, while all group A heads encode the object value. Activation patching confirms that group A (value fetcher) retrieves the object at the position identified by groups B and C (position detector and transmitter). This position-based scheme is the same in Vicuna-7b and Goat-7b, confirming that fine-tuning does not change the mechanism, only enhances it.
Evidence
interventional
Key metric
71.43% of group B heads and 70.0% of group C heads encode positional information (DCM); group A heads primarily encode object value; position detector and transmitter accuracy similar across Llama-7b, Vicuna-7b, Goat-7b
Caveat
The exact nature of the positional information is not fully characterized; the authors note that additional segments at the start and boxes before the correct segment interfere with the positional encoding, suggesting relative position from the start token.