Angelos Poulis

Takeaway: truth directions in LLMs seem robust mostly in a limited range of pure-factual tasks for specific prompt formats, but break down when truth assessment requires tracking intermediate results. 📄Testing the Limits of Truth Directions in LLMs: arxiv.org/pdf/2604.03754

Geometrically, we observe that as task difficulty increases, activations of true and false statements become indistinguishable.

Does an LLM have an internal representation of truth? Yes... but it is more limited than previously assumed. E.g., counting how many (out of 3) cities are in the same country can significantly degrade truth representations. New preprint with @mcrovella and Evimaria Terzi🧵