Research map / Image generation / Fast diffusion sampling
WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving
Yifang Xu, Jiahao Cui, Zhihao Zhu, Hanlin Shang, Shan Luan, Mingwang Xu, Feipeng Cai, Neng Zhang, Yaoyi Li, Jia Cai, Siyu Zhu
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 |
|---|---|---|---|
| use dedicated numerical tokenizer instead of Janus text tokenizer (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 76.2 -> 81.1 | Improved |
| add metric-aligned triplet-margin embeddings (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 81.1 -> 83.4 | Improved |
| add large-scale VQA pretraining (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 83.4 -> 86.7 | Improved |
| add simulator-guided GRPO (after pretraining) (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 86.7 -> 90.3 | Improved |
| add VQA pretraining on top of GRPO (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 86.9 -> 90.3 | Improved |
| increase pretraining data from 0 to 0.65M VQA (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: +1.9 (numerical tokenizer), +5.2 (text tokenizer) | Improved |
| increase pretraining data from 0.65M to 6.5M (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: +1.4 (numerical tokenizer), +2.6 (text tokenizer) | Improved |
| pretrain for 3 epochs versus 0 (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: +3.3 | Improved |
| GRPO group size 3 instead of 2 (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 89.2 -> 90.3 | Improved |
| GRPO group size 4 instead of 3 (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 90.3 -> 89.6 | Hurt |
| reward weights favoring TTC (5:20:2) instead of 5:5:2 (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 90.3 -> 89.7 | Hurt |
| reward weights favoring EP (20:5:2) instead of 5:5:2 (ablation) | WAM-Flow · NAVSIM-v1 | PDMS: 90.3 -> 90.0 | Hurt |
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.
- DiffusionDrive — DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025) · better on 1 of 1 shared result
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
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory (NeurIPS 2026 · similarity 0.87)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.87)
- GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving (CVPR 2025 · similarity 0.87)
- ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving (NeurIPS 2026 · similarity 0.86)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.86)
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.
