Visual Question Answering on DriveQA-V CARLA Signs (test)
75.7Regulatory AccuracyTraffic-MLLM
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Traffic-MLLMParams=4B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 75.7 | 74.8 | 72.1 | 70.6 | 74.5 | |
| TSR-MLLMParams=4B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 74.3 | 78.2 | 62.7 | 81.2 | 74.9 | |
| Mini-InternVLParams=2B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 64.1 | 55.3 | 65.8 | 45 | 59.6 | |
| LLaVA-1.6-mistralParams=7B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 42.6 | 43 | 52.8 | 37.5 | 43.4 | |
| VILA-1.5Params=8B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 25.3 | 23.3 | 27.8 | 21.5 | 24.6 | |
| LLaVA-1.5Params=7B, Decoding Protocol=greedy letter decoding, Evaluation Mode=zero-shot (no fine-tuning on DriveQA)2026.07 | 23.5 | 26.6 | 22.3 | 21.1 | 24.1 |