Research map / Vision-language and multimodal / Multimodal LLMs
Fine-Grained Evaluation of Large Vision-Language Models in Autonomous Driving
Yue Li, Meng Tian, Zhenyu Lin, Jiangtong Zhu, Dechang Zhu, Haiqiang Liu, Yueyi Zhang, Zhiwei Xiong, Xinhai Zhao
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 |
|---|---|---|---|
| train IVL2-4B on all domain data jointly instead of TKU data only (ablation) | IVL2-4B · VLADBench | TKU score gain: 8.67 -> 14.50 | Improved |
| train on GER or EDP data only (vs all data) (ablation) | IVL2-4B · VLADBench TGG domain | score gain | Hurt |
| train on all domain data jointly vs individual domain data (ablation) | IVL2-4B · VLADBench | score gain across domains | Improved |
| add domain-specific AD data to InternVL2-4B (DriveLM-4B) | IVL2-4B · VLADBench TAC domain | score | Hurt |
| add traffic-knowledge tips to TGG questions (ablation) | VLMs · VLADBench signal element relation | improvement rate: up to 37% | Improved |
| add traffic-knowledge tips to TGG questions (ablation) | VLMs · VLADBench lane element relation | improvement rate: up to 73% | Improved |
| scale language model to 72B with unchanged SigLIP encoder (OV, LV) | OneVision/LLaVA-Video · VLADBench | Overall score: 52.15 -> 52.64 (OV); 53.63 -> 54.04 (LV) | Mixed |
| scale model to 72B-76B generally | Qwen/InternVL/OV/LV · VLADBench | Overall score: 53.63 -> 61.03 (best small LV-7B vs QW2.5-72B) | Mixed |
| train with GER data for motion planning (vs TKU, EDP etc.) (ablation) | IVL2-4B · DriveLM-nuScenes val (ST-P3) | Avg L2 (m): 0.65 vs 0.70 (TKU), 0.75 (EDP) | Improved |
| train with TGG or TAC data for motion planning (ablation) | IVL2-4B · DriveLM-nuScenes val (ST-P3) | Avg L2 (m): 0.81 (TGG), 0.87 (TAC) vs 0.65 (GER) | Mixed |
| train with all domain data for motion planning (ablation) | IVL2-4B · DriveLM-nuScenes val (ST-P3) | Avg collision (%): 0.33 vs 0.40 (TKU only) | Improved |
Closest papers by meaning
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.92)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.92)
- STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.91)
- MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding (CVPR 2024 · similarity 0.91)
- nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving (NeurIPS 2026 · similarity 0.90)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.90)
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.
