GPT-3.5 — simulates activations for → Autointerpretability score

GPT-3.5 never sees the fragments GPT-4 used to write its explanation; it is handed only the finished explanation text and must predict per-token activations on a fresh batch of 5 highly-activating and 5 randomly-selected fragments, 10 fragments of 64 tokens each, 640 tokens total, with the score being the correlation between its predicted and the feature's true activations. This is a deliberate deviation from Bills et al. (2023)'s original protocol, which used GPT-4 for simulation too; the authors swap in GPT-3.5 purely because OpenAI's public API exposed the per-token logprobs the simulation scoring needs for GPT-3.5 but not for GPT-4 at the time. The result is an asymmetric division of labor: GPT-4 gets curated top-activating text plus the numeric activations themselves, while GPT-3.5 gets only prose and must reconstruct numbers from it. So the autointerpretability score is not really testing 'is this feature interpretable' in the abstract; it is testing whether an explanation written by one model with privileged numeric access transfers well enough for a second model with no such access to reconstruct those same numbers from words alone.