HeadEdit Calibrates Model Behavior via the Unembedding Matrix
October 2, 2026
HeadEdit is a gradient-free method that calibrates model behavior by extracting a low-rank behavioral subspace from the frozen unembedding matrix. It implements adaptive steering without parameter updates or manually specified target tokens, correcting errors like unnecessary tool calls or false compliance.
HOW THIS AFFECTS YOU
●
builderYou can steer model behavior and correct refusals or tool-use errors without expensive fine-tuning.
●
researcherThis demonstrates that behavioral errors can be addressed via the unembedding matrix rather than just through computation.