Encodes creative works — books, films, music, and artworks.
What It Does
Identity.Work neurons activate on named creative works: books ('Pride and Prejudice'), films ('Casablanca'), musical compositions ('Beethoven's Ninth'), artworks ('the Mona Lisa'), games ('Minecraft'), and other creative artifacts that have distinct identities as works. They encode the work as a specific named entity separate from its creator and content.
How It Behaves
Work neurons show a late-layer concentration, reflecting the high contextual demands of work-entity recognition — the same title can refer to different works ('The Ring' is a film, a Wagner opera, and a Tolkien artifact), and disambiguation requires full contextual integration. Work neurons co-activate with Person neurons (linking works to their creators) and Time neurons (publication or release dates). They show higher activation in models with strong literary and cultural training data.
Research Example
In OLMo 3 7B, Identity.Work neurons activate differently on 'Hamlet' as a character name vs. 'Hamlet' as the name of the play — the same word, in context, triggers either a Person-adjacent or Work-specific firing signature. Confusing characters with works is a source of specific hallucination types where the model attributes dialogue to the work rather than the character within it.