Training-Free Task Vectors for LLM Behavioral Control
The paper proposes task-vector-style model edits without doing fine-tuning first.
Its method turns activation steering vectors into rank-one weight-space edits using forward-pass statistics. The authors say those edits support adding, subtracting, and composing behaviors, including learning-like amplification and forgetting-like suppression. In their LLM behavior-control tests, the edits reportedly controlled target traits while preserving general knowledge and problem-solving ability. ArXiv · AI/CL/LG's note
Its method turns activation steering vectors into rank-one weight-space edits using forward-pass statistics. The authors say those edits support adding, subtracting, and composing behaviors, including learning-like amplification and forgetting-like suppression. In their LLM behavior-control tests, the edits reportedly controlled target traits while preserving general knowledge and problem-solving ability. ArXiv · AI/CL/LG's note
score 6