Research map / Vision-language and multimodal / Multimodal LLMs
Passing the Driving Knowledge Test
Maolin Wei, Wanzhou Liu, Eshed Ohn-Bar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DriveQA finetuning (ablation) | Gemma-2 9B · DriveQA-T average | accuracy: 76.91 -> 87.28 | Improved |
| add DriveQA finetuning (ablation) | Llama-3.1 8B · DriveQA-T average | accuracy: 60.79 -> 87.62 | Improved |
| add CoT and RAG context (ablation) | Gemma-2 2B · DriveQA-T average | accuracy: 44.15 -> 56.62 | Improved |
| add RAG context to CoT (ablation) | Gemma-2 2B · DriveQA-T explanations | BLEU-4: 0.1098 -> 0.1704 | Improved |
| add RAG context to CoT after finetuning (ablation) | Gemma-2 2B · DriveQA-T explanations | BLEU-4: 0.3623 -> 0.2934 | Hurt |
| finetune on DriveQA vs GPT-4o for explanations | Gemma-2 9B · DriveQA-T explanations | ROUGE-L: 0.5393 -> 0.5528 | Improved |
| add DriveQA finetuning (ablation) | Mini-InternVL 2B · DriveQA-V average | accuracy: 41.82 -> 86.59 | Improved |
| add DriveQA finetuning (ablation) | LLaVA-1.5 7B · DriveQA-V average | accuracy: 23.52 -> 69.67 | Improved |
| add negative-sampling hard distractors (ablation) | GPT-4o · DriveQA-V (Signs) | accuracy: 94.10 -> 79.40 | Hurt |
| add negative-sampling hard question set (ablation) | Llama-3.1 8B · DriveQA-T hard subset | accuracy: 55.89 -> 39.87 | Hurt |
| finetune on DriveQA, test on real Mapillary signs | VILA-1.5 8B · Real-world Mapillary signs | accuracy: 34.38 -> 60.86 | Improved |
| pretrain on DriveQA for trajectory planning (ablation) | InternVL-2.5-8B · nuScenes | avg L2 (m): 3.06 -> 2.71 | Improved |
Closest papers by meaning
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.91)
- STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.88)
- MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding (CVPR 2024 · similarity 0.88)
- SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.88)
- NuScenes-QA: A Multi-Modal Visual Question Answering Benchmark for Autonomous Driving Scenario (AAAI 2024 · similarity 0.88)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
