VLAA-Thinker-Qwen2.5VL-7B: benchmark results
UC Santa Cruz VLAA's VLAA-Thinker Qwen2.5-VL 7B, a vision-language reasoning model built on Qwen2.5-VL-7B and trained with SFT/RL for R1-style reasoning, from their 'SFT or RL?' study. Provider: UCSC-VLAA. Released 2025-03-31. Access: Open.
Aggregate Rating 1567 ± 22, rank #556 of 1654 rated models, from 99 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LMM Reasoning - WeMath - Correspondence of Coordinates and Positions | 93.3 | Accuracy (%) | 100 |
| Open LMM Reasoning - LogicVista - inductive | 43 | Accuracy (%) | 98.1 |
| Open LMM Reasoning - MathVerse - Angle | 54.9 | Accuracy (%) | 95.9 |
| Open LMM Reasoning - MathVerse - Plane Geometry | 53.3 | Accuracy (%) | 95.9 |
| Open LMM Reasoning - MathVision - topology | 34.8 | Accuracy (%) | 94.9 |
| Open LMM Reasoning - MathVerse - Applied | 51.4 | Accuracy (%) | 94.4 |
| OpenVLM MathVista - FQA | 75.5 | Accuracy (%) | 92.5 |
| OpenVLM MathVista - STA | 84.4 | Accuracy (%) | 91.3 |
| Open LMM Reasoning - MathVerse | 48.2 | Accuracy (%) | 90.9 |
| Open LMM Reasoning - WeMath - Calculation of Plane Figures | 82 | Accuracy (%) | 90.1 |
| Open LMM Reasoning - MathVerse - Length | 56 | Accuracy (%) | 89.8 |
| Open LMM Reasoning - WeMath - Understanding of Solid Figures | 67.7 | Accuracy (%) | 87.7 |
Interactive version: theaggregate.ai/model?slug=vlaa-thinker-qwen2-5vl-7b · How It Works · Data refreshed daily, snapshot 2026-10-24.