Research map / Vision-language and multimodal / Multimodal LLMs
Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data and Metric Perspectives
Shaoyuan Xie, Lingdong Kong, Yuhao Dong, Chonghao Sima, Wenwei Zhang, Qi Alfred Chen, Ziwei Liu, Liang Pan
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 |
|---|---|---|---|
| remove the image so the prompt is text-only (ablation) | GPT-4o · DriveBench behavior | GPT score: 45.40 -> 50.03 | No clear effect |
| remove the image so the prompt is text-only (ablation) | GPT-4o · DriveBench perception | GPT score: 35.37 -> 36.48 | No clear effect |
| add the image back to the prompt (vs text-only) (ablation) | Qwen2-VL-72B · DriveBench prediction | GPT score: 5.57 -> 49.35 | Improved |
| add the image back to the prompt (vs fully black) (ablation) | Qwen2-VL-72B · DriveBench perception MCQ | accuracy: 23.5 -> 60.0 | Improved |
| add the image back to the prompt (vs fully black) (ablation) | GPT-4o · DriveBench perception MCQ | accuracy: 59.5 -> 59.0 | No clear effect |
| add the image back to the prompt (vs fully black) (ablation) | Qwen2-VL-72B · DriveBench behavior MCQ | accuracy: 36.5 -> 23.0 | Hurt |
| add the image back to the prompt (vs fully black) (ablation) | Phi-3 · DriveBench perception MCQ | accuracy: 17.5 -> 54.5 | Improved |
| apply corruptions to images (avg of 15) | Human annotators · DriveBench behavior MCQ | GPT score: 69.51 -> 54.09 | Hurt |
| apply corruptions to images (avg of 15) | GPT-4o · DriveBench perception | GPT score: 35.37 -> 35.25 | No clear effect |
| name the corruption type explicitly in the question (ablation) | LLaVA-NeXT · DriveBench MCQ (Snow) | accuracy change: -31.95 | Hurt |
| name the corruption type explicitly in the question (ablation) | GPT-4o · DriveBench MCQ (Dark) | accuracy change: -12.98 | Hurt |
| name the corruption type explicitly in the question (ablation) | LLaVA-1.5-7B · DriveBench MCQ (Bright) | accuracy change: 0.26 | No clear effect |
Closest papers by meaning
- DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution (NeurIPS 2026 · similarity 0.91)
- CoC-VLA: Delving into Adversarial Domain Transfer for Explainable Autonomous Driving via Chain-of-Causality Visual-Language-Action Model (NeurIPS 2025 · similarity 0.91)
- STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.91)
- DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.90)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.90)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 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.
