Research map / Large language models / Reasoning and chain of thought
CritiqueDriveVLM: From Verifier-Guided Reinforcement Learning to Latent Thought Distillation for Autonomous Driving
Zhaohong Liu, Hao Ye, Xianlin Zhang, Mengshi Qi
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 accuracy-reward GRPO after warm-up SFT (ablation) | Qwen3-VL-8B SFT · DriveLMM-o1 | MCQ (%): 62.86 -> 64.49 | Improved |
| add multi-dimensional verifier rewards to single-turn GRPO (ablation) | Qwen3-VL-8B GRPO · DriveLMM-o1 | MCQ (%): 64.49 -> 67.46 | Improved |
| add multi-turn interaction with step-decay penalty (ablation) | Qwen3-VL-8B GRPO + verifier · DriveLMM-o1 | MCQ (%): 67.46 -> 76.54 | Improved |
| add multi-turn interaction with step-decay penalty (ablation) | Qwen3-VL-8B GRPO + verifier · DriveLMM-o1 | Overall reasoning (%): 75.10 -> 80.48 | Improved |
| add multi-turn interaction with step-decay penalty (ablation) | Qwen3-VL-8B GRPO + verifier · DriveLMM-o1 | Risk assessment (%): 75.30 -> 81.63 | Improved |
| add warm-up SFT on DriveLMM-o1 CoT data (ablation) | Qwen3-VL-8B · DriveLMM-o1 | MCQ (%): 55.54 -> 62.86 | Mixed |
| add latent alignment loss (Latent Thought Distillation) to CoT-free student (ablation) | Qwen3-VL-8B (SFT) CoT-free · DriveLMM-o1 | MCQ (%): 61.73 -> 68.59 | Improved |
| distill into CoT-free student instead of running teacher with CoT (ablation) | Ours-Teacher · DriveLMM-o1 | Avg. inference time (ms): 3482 -> 416 | Improved |
| distill into CoT-free student instead of running teacher with CoT | Ours-Teacher · DriveLMM-o1 | MCQ (%): 76.54 -> 68.59 | Hurt |
| remove CoT from Qwen3-VL-8B (SFT) at inference | Qwen3-VL-8B (SFT) · DriveLMM-o1 | MCQ (%): 62.86 -> 61.73 | Hurt |
| use verifier-critique multi-turn RL instead of external tool calls (AgentThink) | Ours-Teacher vs AgentThink · DriveLMM-o1 | Avg. inference time (ms): 5416 -> 3482 | Improved |
| tool-free multi-turn RL teacher versus tool-augmented AgentThink | Ours-Teacher vs AgentThink · DriveLMM-o1 | MCQ (%): 71.35 -> 76.54 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Mulberry-7B — Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search (NeurIPS 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.93)
- $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.91)
- DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.91)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.91)
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (CVPR 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.
