whisperers · prefix dictionary
whisperers searches for short strings of tokens, prefixes, that make a language model do a task when they are placed in front of an input. Most read as token soup, yet with one in front of hot -> the model answers cold. This page lists every prefix we have searched: the model it was found on, what the search was aimed at, how the checkpoint was picked, and how it did on held-out words the search never saw.
Headline
On base models, plain GCG on the answers matches or beats prefixes aimed at attention heads — by writing the instruction in words. Ban those words and heads prefixes come out ahead for antonyms, but not for synonyms or English→Spanish. On filter heads the paper’s own intervention still wins (one run each).
Epistemic status: every number comes from a single run unless an entry says otherwise. Claims are about base models only; chat, reasoning and other post-trained models are work in progress.