Valen Tagliabue, Leonard Dung, Cameron Berg · 2026 · arXiv (preprint, ongoing work)
Extracts a pain direction across 25 open models, distinct from fear and generic negative valence, that rises when harm is aimed at the model and not when the user is suffering. Steered models pay costs to press a relief button and press again less when relief is real. The authors note that steering may evoke a character in pain rather than the model being in pain, and leave that question to future work.