Discovering Variable Binding Circuitry with Desiderata

07/07/2023
by   Xander Davies, et al.
0

Recent work has shown that computation in language models may be human-understandable, with successful efforts to localize and intervene on both single-unit features and input-output circuits. Here, we introduce an approach which extends causal mediation experiments to automatically identify model components responsible for performing a specific subtask by solely specifying a set of desiderata, or causal attributes of the model components executing that subtask. As a proof of concept, we apply our method to automatically discover shared variable binding circuitry in LLaMA-13B, which retrieves variable values for multiple arithmetic tasks. Our method successfully localizes variable binding to only 9 attention heads (of the 1.6k) and one MLP in the final token's residual stream.

READ FULL TEXT

Please sign up or login with your details

Forgot password? Click here to reset