Spilios Dimakopoulos

TwinLens: an interpretability digital twin for LLMs

TwinLens builds a transcoder-based digital twin of a language model, predicts the effect of feature interventions, and checks each prediction against the real model. Across 8,064 interventions on GPT-2 small (56 prompts), the linear twin that attribution graphs describe ranks real effects poorly (Spearman rho 0.20). Th...