Complex Reasoning on TVbench
54.7AccuracyQwen3-VL-8B + SynRL
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3-VL-8B + SynRLModel Scale=8B, SynRL Training=true2026.03 | 54.7 | |
| Qwen3-VL-8BModel Scale=8B, SynRL Training=false2026.03 | 54 | |
| Qwen3-VL-4B + SynRLModel Scale=4B, SynRL Training=true2026.03 | 53.6 | |
| Qwen3-VL-4BModel Scale=4B, SynRL Training=false2026.03 | 51.9 |