Road Reasoning on CRS (Compositional Road Scenes)
95.5Total Score (CRS)Human expert
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Human expertNotes=Randomly selected 300 samples2026.05 | 95.5 | — | — | |
| Qwen3-VL-4b-InstructSupervised fine-tune=true, Chain-of-Thought (CoT)=true2026.05 | 71.3 | 72.2 | 67.5 | |
| Qwen3-VL-4b-InstructSupervised fine-tune=true2026.05 | 71.1 | 71.8 | 68 | |
| Qwen3-VL-2b-InstructSupervised fine-tune=true, Chain-of-Thought (CoT)=true2026.05 | 69.6 | 71.5 | 62 | |
| Qwen3-VL-2b-InstructSupervised fine-tune=true2026.05 | 69.2 | 71.5 | 60 | |
| GPT-5.42026.05 | 62.1 | 66.7 | 43.5 | |
| Qwen3-VL-4b-InstructSupervised fine-tune=true, Input modality=without visual input (- images)2026.05 | 58.7 | 58.6 | 59 | |
| Gemini-3.1-Pro2026.05 | 58.4 | 60.9 | 48 | |
| Gemma-4-31b-IT2026.05 | 53.9 | 57.6 | 39 | |
| Claude-Sonnet-4.52026.05 | 52.1 | 54.8 | 41 | |
| Gemini-3.1-Flash-Lite2026.05 | 51.9 | 55.4 | 37.5 | |
| Qwen3-VL-235b-a22b-Instruct2026.05 | 49.9 | 51.2 | 44.5 | |
| Qwen3-VL-30b-a3b-Thinking2026.05 | 46 | 48.4 | 36 | |
| Qwen3-VL-4b-Instruct2026.05 | 44.3 | 46.6 | 35 | |
| Qwen3-VL-2b-Instruct2026.05 | 42 | 42.9 | 38.5 | |
| GPT-4o-mini2026.05 | 41.7 | 44.5 | 30.5 | |
| Kimi-k2.62026.05 | 41.7 | 43.9 | 33 | |
| Molmo2-8B2026.05 | 40.5 | 41 | 38.5 |