fv-models-tested

IN premise — summaries/2026/08/24/todd-2023-function-vectors-s2-function-vector-causal-effects-cannot-be-recovered-from-the-.md

Created 2026-08-25T02:58:41+00:00

The function vector experiments were conducted on GPT-J (6B), GPT-NeoX (20B), and Llama 2 (7B/13B/70B).

Summary

This pins down exactly which models and sizes were actually put through the function vector testing: GPT-J at 6 billion parameters, GPT-NeoX at 20 billion, and Llama 2 at 7, 13, and 70 billion. Any conclusions the system draws from those experiments are only as broad as this set of models, so results can't be assumed to hold for architectures or scales outside it.