rome-decisive-token-position
IN premise — summaries/2026/08/24/meng-2022-rome-s0-abstract.md
Created 2026-08-25T02:58:14+00:00
The decisive processing step for factual recall in GPT occurs specifically at the last token of the subject name in the prompt, not the first token or an average across tokens.
Summary
When GPT retrieves a fact, the moment it actually commits to an answer is pinned to the very end of the subject's name in the question, not the beginning or a smear across the whole name. This matters because it tells you exactly where to look, monitor, or intervene if you want to understand or redirect how the model pulls a specific fact.