Motion Reasoning on ReMoT 16k (test)
31.5Navigation: Camera (Overall)SFT↔GRPO
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SFT↔GRPOTraining=Interleaved (SFT and GRPO), Epochs=22026.02 | 31.5 | 68.8 | 23.3 | 40 | 93.3 | 79.8 | 89.5 | 45.2 | 65.1 | 64.5 | 83.3 | 39.9 | 67.7 | |
| GRPOTraining=GRPO only2026.02 | 27 | 62.4 | 24.3 | 44.3 | 94.1 | 54.5 | 74.3 | 27.3 | 44.7 | 61.3 | 79.2 | 33.6 | 61.6 | |
| GRPOBase Model=Qwen3-VL-CoT-4B2026.02 | 27 | 62.4 | 24.3 | 44.3 | 94.1 | 54.5 | 74.3 | 27.3 | 44.7 | 61.3 | 79.2 | 33.6 | 61.6 | |
| SFT→GRPOTraining=Sequential (SFT then GRPO)2026.02 | 26.6 | 62.1 | 22.5 | 35.8 | 82.5 | 57.3 | 78.8 | 32.3 | 60.2 | 62.9 | 82.3 | 35 | 63.3 | |
| SFT→GRPOBase Model=Qwen3-VL-CoT-4B, Training=Sequential2026.02 | 26.6 | 62.1 | 22.5 | 35.8 | 82.5 | 57.3 | 78.8 | 32.3 | 60.2 | 62.9 | 82.3 | 35 | 63.3 | |
| Gemini-2.5-ProAccess=Closed-source2026.02 | 25 | 62.5 | 25 | 25 | 75 | 12.5 | 41.7 | 0 | 16.7 | 25 | 50 | 26.4 | 49.1 | |
| SFT↔GRPOTraining=Interleaved (SFT and GRPO), Epochs=12026.02 | 21.4 | 61.2 | 26.7 | 46.7 | 82.7 | 68.6 | 82 | 45.2 | 65.1 | 69.4 | 86.6 | 38 | 64 | |
| SFT↔GRPOBase Model=Qwen3-VL-CoT-4B, Training=Alternating2026.02 | 21.4 | 61.2 | 26.7 | 46.7 | 82.7 | 68.6 | 82 | 45.2 | 65.1 | 69.4 | 86.6 | 38 | 64 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.02 | 4.8 | 34.7 | 0 | 23.9 | 0 | 4 | 36.6 | 8.1 | 31.7 | 0 | 16.7 | 5.1 | 25.4 | |
| Qwen2.5-VL-7BParameters=7B2026.02 | 4.8 | 34.7 | 0 | 23.9 | 0 | 4 | 36.6 | 8.1 | 31.7 | 0 | 16.7 | 5.1 | 25.4 | |
| InternVL3-8BBackbone=InternVL3-8B2026.02 | 2.8 | 23.8 | 15 | 30.6 | 51.7 | 1.6 | 28.5 | 8.1 | 38.7 | 0 | 18.3 | 12.2 | 28.9 | |
| InternVL3-8BParameters=8B2026.02 | 2.8 | 23.8 | 15 | 30.6 | 51.7 | 1.6 | 28.5 | 8.1 | 38.7 | 0 | 18.3 | 12.2 | 28.9 | |
| Qwen3-VL-CoT-4BBackbone=Qwen3-VL-CoT-4B, Reasoning=CoT2026.02 | 2.4 | 25.9 | 22.5 | 35.8 | 79 | 15.3 | 46.2 | 3.2 | 35.5 | 4.8 | 22 | 20.7 | 38.9 | |
| Qwen3-VL-CoT-4BParameters=4B2026.02 | 2.4 | 25.9 | 22.5 | 35.8 | 79 | 15.3 | 46.2 | 3.2 | 35.5 | 4.8 | 22 | 20.7 | 38.9 | |
| LLaVA-One-VisionBackbone=LLaVA-One-Vision2026.02 | 2 | 30.4 | 17.5 | 32.9 | 0 | 10.5 | 37.6 | 21 | 47.3 | 0 | 14.8 | 9.7 | 27.9 | |
| LLaVA-One-Vision2026.02 | 2 | 30.4 | 17.5 | 32.9 | 0 | 10.5 | 37.6 | 21 | 47.3 | 0 | 14.8 | 9.7 | 27.9 | |
| InternVL3-2BBackbone=InternVL3-2B2026.02 | 1.6 | 21.5 | 20 | 31.3 | 60.5 | 0.8 | 29.8 | 6.5 | 35 | 0 | 17.7 | 14.9 | 29.3 | |
| InternVL3-2BParameters=2B2026.02 | 1.6 | 21.5 | 20 | 31.3 | 60.5 | 0.8 | 29.8 | 6.5 | 35 | 0 | 17.7 | 14.9 | 29.3 | |
| GPT-5-ChatAccess=Closed-source2026.02 | 0 | 31.3 | 0 | 0 | 75 | 0 | 37.5 | 25 | 41.7 | 0 | 33.4 | 10.4 | 33.3 | |
| SFTTraining=SFT only2026.02 | 0 | 2.4 | 23.3 | 40 | 37.9 | 0 | 10.5 | 0 | 9.7 | 0 | 9.2 | 10.4 | 15.4 |