1 paper · 1 filter
Yukiya Horiba, Hakaze Cho, Takahiro Yakoh
Task Vectors (TVs) provide a compact way to steer large language models, but they are typically treated as atomized directions merely extracted from hidden states, leaving their in…