Fast & Faithful Function Vectors for Steering LLMs
arXiv 2606.05079·medium signal
Pham, Segeler, and Wiegand improve function vectors (FVs) — task representations elicited during in-context learning that can steer LLM behavior — making them both faster to compute and more faithful to the intended task. FVs offer a lightweight alternative to fine-tuning for controlling model output. Useful for builders wanting cheap, targeted behavioral steering.