Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces
Open paper
Benchmark
Model
Rank
Results
multimodal-reasoning-on-mathvista
Qwen2.5-VL-7B-Instruct + Label-Free Bias-Only TTRL
#82
Accuracy: 65.4
mathematical-reasoning-on-math-500
Qwen2.5-Math-7B + Label-Free Bias-Only TTRL
#93
Accuracy: 79.5
Rank counts only results with a code link.