AI
DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution
Publication: cs.AI updates on arXiv.org
Author: Chengkai Xu, Jiaqi Liu, Yicheng Guo, Peng Hang, Jian Sun
Source published: 29/09/2026, 04:00:00 UTC
Briefing created: 29/09/2026, 05:45:06 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:24 UTC.
Briefings
The paper introduces DriveHierarchy, a benchmark organizing VLM-based autonomous driving into four ranks: perceptual grounding, contextual memory, mental reasoning, and closed-loop execution. It integrates 84,279 frames and 76,798 question-answer pairs from open-source datasets into a unified open-loop benchmark, alongside a closed-loop simulation platform with 100 curated scenarios. Experiments on 15 VLMs demonstrate that the benchmark captures structured but non-redundant capability variation and provides a practical basis for diagnosing and optimizing model performance.
Key points
- DriveHierarchy organizes VLM-based autonomous driving into four distinct ranks: perceptual grounding, contextual memory, mental reasoning, and closed-loop execution.
- The benchmark integrates 84,279 frames and 76,798 question-answer pairs from open-source datasets into a unified open-loop benchmark with a closed-loop simulation platform.
Context and limitations
The benchmark was evaluated on 15 VLMs, and the closed-loop simulation platform was constructed from a real-world road network with 100 curated driving scenarios.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 05:45:06 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigytgi5av4o6ket2giiyo3yi33nhcirjjygbrlgmf57gwf66kfjy4