Abstract
Test-time reinforcement learning (TTRL) enables models to improve their reasoning without relying on labeled training data, but existing approaches typically optimize a large fraction of the model parameters. This raises a natural question: can effective test-time adaptation emerge when both the reward signal and the optimization space are severely restricted? We answer this question with label-free bias-only TTRL, which uses majority-vote pseudo-labels as rewards and optimizes only ~100K bias parameters while keeping the pretrained backbone frozen. On MATH-500, our approach reaches 76.67% accuracy, matching the labeled bias-steering baseline while optimizing 76,000× fewer parameters than full-parameter TTRL. The same training procedure improves performance across vision-language and audio reasoning tasks, including MathVista, AI2D, LogicVista, and MMAU. We further show that the learned steering vectors transfer to 4,500 held-out MATH problems, indicating that the adaptation is not limited to the problems used during test-time optimization. Finally, we analyze why this highly restricted adaptation can work, showing that majority-vote reliability improves with rollout consensus and that bias subspaces with greater accessible gradient energy exhibit stronger downstream trainability.
Method
Label-free bias-only TTRL sits at the intersection of two restrictions applied together for the first time: it removes the need for labeled rewards, the way TTRL does, while also restricting optimization to a tiny per-layer additive bias subspace, the way bias-only steering does — rather than requiring either a labeled reward or a full-parameter update.
| Optimization space | ~100K additive bias parameters per model (one vector per MLP layer), backbone frozen. |
| Reward | Majority-vote pseudo-labels from the model's own rollouts — no ground-truth labels. |
| Modalities | Text (MATH-500), vision-language (MathVista, AI2D, LogicVista), audio (MMAU). |
| Models | Qwen2.5-7B / Qwen2.5-Math-7B, Qwen2.5-VL-7B-Instruct, Qwen2.5-Omni-7B. |
Key Findings
-
1. Competitive with Far More Expensive Alternatives
On MATH-500, label-free bias-only TTRL reaches 76.67% / 79.50% accuracy on Qwen2.5-7B / Qwen2.5-Math-7B, within one point of a labeled bias-steering reproduction, while training 76,000× fewer parameters than full-parameter TTRL. -
2. Removing Labels Costs Almost Nothing
The label-free reward loses only a small amount of accuracy relative to the same ~100K-parameter budget trained with ground-truth labels. -
3. Not Uniformly Beaten by Larger Trainable Subspaces
Compared against LoRA at r=16 and r=64, bias-only TTRL wins on LogicVista and MathVista and remains competitive elsewhere — a much smaller subspace is not systematically worse. -
4. Transfers Beyond the Training Problems
Frozen checkpoints evaluated on 4,500 held-out, verified-disjoint MATH problems still gain 22–25 points over baseline, evidence the adaptation is not just memorizing majority-vote consensus on the training set. -
5. Generalizes Across Modalities with One Recipe
The same training procedure, only the prompt and grading function changed, improves MathVista, AI2D, LogicVista, and MMAU over their untrained baselines. -
6. Equal Parameter Count ≠ Equal Trainability
Across nine single-layer bias subspaces of identical size, accessible gradient energy correlates with downstream accuracy (Spearman ρ=0.917, exact two-tailed permutation test, p=0.0013) — which bias subspace you pick matters, not just how large it is.
Main Results
| Model | Benchmark | Baseline | Bias-only TTRL (ours) |
|---|---|---|---|
| Qwen2.5-Math-7B | MATH-500 | 56.0 | 79.50 ± 0.70 |
| Qwen2.5-7B | MATH-500 | 45.9 | 76.67 ± 1.57 |
| Qwen2.5-VL-7B-Instruct | AI2D | 73.58 | 78.29 ± 0.03 |
| Qwen2.5-VL-7B-Instruct | LogicVista | 37.50 | 43.38 ± 1.01 |
| Qwen2.5-VL-7B-Instruct | MathVista | 61.60 | 65.40 ± 1.77 |
| Qwen2.5-Omni-7B | MMAU | 57.40 | 60.87 ± 0.37 |
Mean ± standard deviation over 3 independent training seeds. Full comparisons against LoRA and full-parameter fine-tuning are in the paper.
Training Recipe
All training and evaluation scripts, organized per benchmark, share the same recipe:
python src/train_ttrl_steer.py \
--model_id Qwen/Qwen2.5-Math-7B \
--dataset math --num_steps 300 --G 64 --lr 5e-4 \
--use_vllm --output_dir outputs/math500_math7b
The repository README lists the full set of launch commands across all five benchmarks. Trained bias vectors for each benchmark are released on Hugging Face.
Resources
arXiv Preprint
Read the full paper on label-free bias-only test-time reinforcement learning.
Hugging Face Checkpoints
Trained bias steering vectors for each of the five benchmarks covered in the paper.
BibTeX
@article{vakada2026labelfree,
title = {Label-Free Steering: Compressing Test-Time Reinforcement Learning into Bias-Only Subspaces},
author = {Vakada, Naveen and Li, Mingyuan and Ji, Shaoxiong},
journal = {arXiv preprint arXiv:2609.18587},
eprint = {2609.18587},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2609.18587},
year = {2026}
}
Funding
HAIF is co-funded by the European Union's Horizon Europe research and innovation programme's Marie Skłodowska-Curie Action.



