Reproduce “Training-Free Task Vectors for LLM Behavioral Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gabriel J. Perin, Lucas Boscaini, André Araujo, Nina S. T. Hirata
TFTV turns contrastive examples of a behavior into a persistent rank-one weight edit without first fine-tuning a behavior-specific checkpoint. It derives a steering direction from judged completion activations and combines it with weight-matrix SVD statistics and expected module inputs. On Llama-3.1-8B-Instruct, its three-trait edit scored 0.03 evil versus 17.05 with inference steering; utility remains specific to the evaluated setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.