Training-Free Task Vectors Turn Activation Steering Into Rank-One Weight Edits Using Only Forward Passes

arXiv 2609.09054·low signalXBlueskyLinkedInCopy link

Task vectors normally require computing the difference between a fine-tuned model and its pretrained initialization, which makes discovering useful directions expensive. TFTVs map activation steering vectors to rank-one weight-space edits using only forward-pass statistics while preserving the arithmetic properties that support learning via addition, forgetting via subtraction, and composition of multiple edits. On LLM behavioral control tasks the method consistently amplifies, suppresses and composes target behaviors while preserving general knowledge and problem-solving ability, achieving stronger trait control than editing and steering baselines with better or competitive utility preservation. Code is at tftv-llm.github.io.