AI
Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 14:56:08 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 13:18:24 UTC.
Briefings
Label-free test-time reinforcement learning (TTRL) achieves high accuracy by optimizing only ~100K bias parameters instead of full model weights. On MATH-500, Qwen2.5-7B reaches 76.67% accuracy using majority-vote pseudolabels as rewards, optimizing 76,000x fewer parameters than full-parameter TTRL. This method improves reasoning across vision-language and audio tasks, with learned steering vectors transferring to 4,500 held-out MATH problems. The study suggests that bias subspaces with greater accessible gradient energy exhibit stronger downstream trainability.
Key points
- The study introduces label-free bias-only TTRL, which optimizes only ~100K bias parameters while keeping the pretrained backbone frozen.
- On MATH-500, Qwen2.5-7B reaches 76.67% accuracy using majority-vote pseudolabels as rewards, optimizing 76,000x fewer parameters than full-parameter TTRL.
Context and limitations
Results are limited to MATH-500 and MATHVista, AI2D, LogicVista, and MMAU tasks; transfer to 4,500 held-out MATH problems is confirmed but not exhaustive.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 14:56:08 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeia3i6rgjgz3jhfddoonln2yvfcjgeg4lhxkpiqd2gc5hzjo37mp2u