Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces
The paper claims test-time gains from tuning only about 100,000 bias parameters, using majority-vote pseudolabels instead of labeled rewards.
With Qwen2.5-7B on MATH-500, the method reaches 76.67% accuracy while optimizing far fewer parameters than full-parameter TTRL. The same procedure is reported to improve results on MathVista, AI2D, LogicVista, and MMAU. The authors also say the learned steering vectors transfer to 4,500 held-out MATH problems, suggesting the effect is not confined to the optimization set. HF Daily Papers' note
With Qwen2.5-7B on MATH-500, the method reaches 76.67% accuracy while optimizing far fewer parameters than full-parameter TTRL. The same procedure is reported to improve results on MathVista, AI2D, LogicVista, and MMAU. The authors also say the learned steering vectors transfer to 4,500 held-out MATH problems, suggesting the effect is not confined to the optimization set. HF Daily Papers' note
score 5