ECCV 2026 | CritiqueDriveVLM: From Slow, Tool-Heavy Reasoning to 416ms Driving VLM

CritiqueDriveVLM: From Verifier-Guided Reinforcement Learning to Latent Thought Distillation for Autonomous Driving Zhaohong Liu, Hao Ye, Xianlin Zhang, Mengshi Qi Beijing University of Posts and Telecommunications ECCV 2026 Paper: arxiv.org/abs/2607.04179 Code: github.com/MICLAB-BUPT/CritiqueDriveVLM CritiqueDriveVLM is a three-stage framework for autonomous driving vision-language models: warm-up SFT and verifier construction, critique-driven multi-turn reinforcement learning, and latent thought distillation. Our Teacher model reaches 76.54 MCQ on DriveLMM-o1 without external tools; our distilled Student answers in 28.8 tokens and 416 ms.