Training-Free Task Vectors for LLM Behavioral Control
This work introduces Training-Free Task Vectors (TFTVs), a novel method to compute task-vector-like directions without requiring fine-tuning and evaluates TFTVs on large language model behavioral control tasks and shows that they consistently amplify, suppress, and compose target behaviors while preserving general know...
G. Perin, Lucas Boscaini, André Araújo et al.
· 0 citations