Label-free steering compresses test-time RL into bias-only subspaces
This research shows that significant test-time adaptation can be achieved by optimizing a tiny bias-only subspace using entirely label-free rewards, which could lead to more efficient and data-efficient AI systems.