Target Viewpoint Reproduction on TVRBench (test)
51.4Overall PerformanceMulti-turn GRPO
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Multi-turn GRPOMem=VA, CoT=no, Init=VA-SFT2026.05 | 51.4 | 81.6 | 64 | 34.4 | 25.6 | 24.9 | 51.4 | 0 | |
| VA-SFTMem=VA, CoT=no, Init=Qwen3.5-9B2026.05 | 50.8 | 82.4 | 68.8 | 27.2 | 24.8 | 25.3 | 50.8 | 0 | |
| AO-SFTMem=AO, CoT=no, Init=Qwen3.5-9B2026.05 | 44.2 | 61.6 | 60.8 | 32 | 22.4 | 25.9 | 48 | 7.9 | |
| VA-CoT-SFTMem=VA, CoT=yes, Init=Qwen3.5-9B2026.05 | 35.6 | 68 | 51.2 | 12.8 | 10.4 | 28.9 | 35.6 | 0 | |
| Single-turn GRPOMem=AO, CoT=no, Init=AO-SFT2026.05 | 26.2 | 48.8 | 35.2 | 12 | 8.8 | 28.2 | 33 | 20.6 | |
| AO-CoT-SFTMem=AO, CoT=yes, Init=Qwen3.5-9B2026.05 | 24.8 | 40.8 | 38.4 | 8.8 | 11.2 | 30.5 | 25.4 | 2.4 | |
| Qwen3.5-9BMem=AO2026.05 | 2.8 | 6.4 | 4.8 | 0 | 0 | 29.9 | 27 | 89.6 | |
| Qwen3.5-9BMem=VA2026.05 | 0 | 0 | 0 | 0 | 0 | 34.3 | 2.4 | 100 |