Qwen-VL-PRM-3B is a process reward model finetuned from Qwen2.5-3B-Instruct on approximately 300,000 examples. It demonstrates strong test-time scaling performance improvements on various advanced multimodal reasoning benchmarks when used with Qwen2.5-VL and Gemma-3 models despite being trained mainly on abstract reasoning datasets and elementary reasoning datasets.
| Model | MMMU | PuzzleVQA | AlgoPuzzleVQA | MathVista | MathVision | Overall |
|---|---|---|---|---|---|---|
| GPT-4o | 70.7 | 60.0 | 57.8 | 30.9 | 31.2 | 50.1 |
| o1 | 78.2 | 78.9 | 54.4 | 73.9 | 60.3 | 69.1 |
| o3 | 82.9 | 84.1 | 62.3 | 86.8 | -- | -- |
| Model | MMMU | PuzzleVQA | AlgoPuzzleVQA | MathVista | MathVision | Overall |
|---|---|---|---|---|---|---|
| Qwen-2.5-VL-3B | 51.7 | 34.5 | 25.7 | 60.0 | 21.2 | 38.6 |
| + VL-PRM-7B | 53.7 (+2.0) | 44.9 (+10.5) | 28.3 (+2.6) | 64.1 (+4.1) | 21.8 (+0.6) | 42.6 (+4.0) |
| Qwen-2.5-VL-7B | 55.0 | 48.0 | 29.1 | 67.8 | 24.2 | 44.8 |
| + VL-PRM-3B | 57.6 (+2.6) | 55.5 (+7.5) | 33.8 (+4.7) | 70.0 (+2.2) | 26.1 (+1.9) | 48.6 (+3.6) |
| + VL-PRM-7B | 57.4 (+2.4) | 54.8 (+6.8) | 35.3 (+6.2) | 71.0 (+3.2) | 26.2 (+2.0) | 48.9 (+4.1) |
| Qwen-2.5-VL-32B | 66.0 | 46.2 | 26.9 | 76.9 | 36.7 | 50.5 |
| + VL-PRM-3B | 67.0 (+1.0) | 67.1 (+20.8) | 41.6 (+14.7) | 77.7 (+0.8) | 40.5 (+3.8) | 58.7 (+8.2) |
| + VL-PRM-7B | 67.6 (+1.6) | 66.8 (+20.6) | 44.2 (+17.3) | 78.3 (+1.4) | 40.1 (+3.2) | 59.4 (+8.9) |
| Model | MMMU | PuzzleVQA | AlgoPuzzleVQA | MathVista | MathVision | Overall |
|---|---|---|---|---|---|---|
| Gemma-3-12B | 57.6 | 45.0 | 29.1 | 58.9 | 28.1 | 43.7 |
| + VL-PRM-3B | 60.4 (+2.8) | 57.7 (+12.7) | 39.7 (+10.6) | 60.3 (+1.4) | 33.8 (+5.7) | 50.4 (+6.7) |
| + VL-PRM-7B | 60.2 (+2.6) | 59.0 (+12.0) | 41.1 (+4.5) | 63.3 (+4.4) | 33.9 (+5.8) | 51.5 (+7.8) |
| Gemma-3-27B | 62.9 | 50.8 | 29.9 | 61.6 | 32.4 | 47.5 |
| + VL-PRM-3B | 65.5 (+2.6) | 67.4 (+16.6) | 40.3 (+10.4) | 65.4 (+3.8) | 39.8 (+7.4) | 55.7 (+8.2) |
| + VL-PRM-7B | 64.5 (+1.6) | 67.6 (+16.8) | 41.1 (+11.2) | 65.2 (+3.6) | 40.9 (+8.5) | 55.9 (+8.4) |
1@misc{ong2025vlprms,
2 title={Training Vision-Language Process Reward Models for Test-Time Scaling in Multimodal Reasoning: Key Insights and Lessons Learned},
3 author={Brandon Ong, Tej Deep Pala, Vernon Toh, William Chandra Tjhi, and Soujanya Poria},
4 year={2025},
5 eprint={2509.23250},
6 archivePrefix={arXiv},
7 primaryClass={cs.AI},
8 url={https://arxiv.org/pdf/2509.23250},
9}