Yariv Barsheshat

Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs

Robert Graham, Edward Stevinson, Yariv Barsheshat

Finetuning language models on small, curated datasets is standard practice for adapting them to specific policies or domains. We show that finetuning on narrow, factually-defensible, moderation-passing data can cause broad ideological shifts across unrelated domains, while preser…Read more