Bias-only steering vectors from label-free test-time RL (GRPO, majority-vote reward). One repo per benchmark/base model.
-
Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces
Paper • 2609.18587 • Published • 4 -
OLAResearchX/label-free-bias-steering-ai2d-qwen2.5-vl-7b
Reinforcement Learning • Updated -
OLAResearchX/label-free-bias-steering-logicvista-qwen2.5-vl-7b
Reinforcement Learning • Updated -
OLAResearchX/label-free-bias-steering-mathvista-qwen2.5-vl-7b
Reinforcement Learning • Updated