mlsanityData to 7 Oct 2026

Devil is in Narrow Policy: Unleashing Exploration in Driving VLA Models

arXiv 2603.06049 ↗First posted 6 Mar 2026cs.CV cs.ROcited by 14

On the map: End-to-end autonomous driving → Reasoning VLA and VLM driving policies

also in 1 more
Abstract (arXiv)

We identify a fundamental Narrow Policy limitation undermining the performance of autonomous VLA models, where driving Imitation Learning (IL) tends to collapse exploration and limit the potential of subsequent Reinforcement Learning (RL) stages, which often saturate prematurely due to insufficient feedback diversity. Thereby, we propose Curious-VLA, a framework that alleviates the exploit-explore dilemma through a two-stage design. During IL, we introduce a Feasible Trajectory Expansion (FTE) strategy to generate multiple physically valid trajectories and a step-wise normalized trajectory representation to adapt this diverse data. In the RL stage, we present Adaptive Diversity-Aware Sampling (ADAS) that prioritizes high-diversity samples and introduce Spanning Driving Reward (SDR) with a focal style weighting to amplify reward's value span for improving sensitivity to driving quality. On the Navsim benchmark, Curious-VLA achieves SoTA results (PDMS 90.3, EPDMS 85.4) and a Best-of-N PDMS of 94.8, demonstrating its effectiveness in unlocking the exploratory potential of VLA models. Code: https://github.com/Mashiroln/curious_vla.git.

Key facts. Cited by 14 papers in the map (data through 7 Oct 2026). Placed in Reasoning VLA and VLM driving policies, one of 9 directions in End-to-end autonomous driving. Builds on 44 of its 53 references that are in the map. First posted on arXiv 6 Mar 2026.

Builds on

44 of 53 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025AdaThinkDrive: Adaptive Thinking via Reinforcement Learning for Autonomous Driving17 Sep
  2. 2025AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning16 Jun
  3. 2025Poutine: Vision-Language-Trajectory Pre-Training and Reinforcement Learning Post-Training Enable Robust End-to-End Autonomous Driving12 Jun
  4. 2025OpenDriveVLA: Towards End-to-end Autonomous Driving with Large Vision Language Action Model30 Mar
  5. 2025ORION: A Holistic End-to-End Autonomous Driving Framework by Vision-Language Instructed Action Generation25 Mar

Other references · 39

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2017Proximal Policy Optimization Algorithmscited by 14,441
  3. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  4. 2025Qwen2.5-VL Technical Reportcited by 5,418
  5. 2017Deep reinforcement learning from human preferencescited by 3,846
  6. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
  7. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
  8. 2024LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Modelscited by 1,442
  9. 2019ZeRO: Memory Optimizations Toward Training Trillion Parameter Modelscited by 1,278
  10. 2022Planning-oriented Autonomous Drivingcited by 1,068
  11. 2023VAD: Vectorized Scene Representation for Efficient Autonomous Drivingcited by 613
  12. 2023DriveLM: Driving with Graph Visual Question Answeringcited by 527
27 more
  1. 2021NuPlan: A closed-loop ML-based planning benchmark for autonomous vehiclescited by 485
  2. 2022TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Drivingcited by 477
  3. 2025The Entropy Mechanism of Reinforcement Learning for Reasoning Language Modelscited by 463
  4. 2022ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learningcited by 365
  5. 2024NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarkingcited by 343
  6. 2024DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Drivingcited by 324
  7. 2024Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillationcited by 244
  8. 2023Is Ego Status All You Need for Open-Loop End-to-End Autonomous Driving?cited by 242
  9. 2023Scene as Occupancycited by 208
  10. 2025Agentic Reinforced Policy Optimizationcited by 166
  11. 2025End-to-End Driving with Online Trajectory Evaluation via BEV World Modelcited by 145
  12. 2023LingoQA: Visual Question Answering for Autonomous Drivingcited by 144
  13. 2025Pseudo-Simulation for Autonomous Drivingcited by 141
  14. 2025DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Drivingcited by 127
  15. 2025DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Drivingcited by 125
  16. 2025Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rolloutscited by 108
  17. 2023Visual Point Cloud Forecasting enables Scalable Autonomous Drivingcited by 97
  18. 2024OpenEMMA: Open-Source Multimodal Model for End-to-End Autonomous Drivingcited by 93
  19. 2025DriveSuprim: Towards Precise Trajectory Selection for End-to-End Planningcited by 91
  20. 2025GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Modelscited by 81
  21. 2025ARTEMIS: Autoregressive End-to-End Trajectory Planning with Mixture of Experts for Autonomous Drivingcited by 66
  22. 2024CarLLaVA: Vision language models for camera-only closed-loop drivingcited by 57
  23. 2024DriveGPT: Scaling Autoregressive Behavior Models for Drivingcited by 29
  24. 2025CDE: Curiosity-Driven Exploration for Efficient Reinforcement Learning in Large Language Modelscited by 28
  25. 2025ImagiDrive: A Unified Imagination-and-Planning Framework for Autonomous Drivingcited by 24
  26. 2025OccVLA: Vision-Language-Action Model with Implicit 3D Occupancy Supervisioncited by 16
  27. 2025EvaDrive: Evolutionary Adversarial Policy Optimization for End-to-End Autonomous Drivingcited by 12

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 14.

14 papers.

Sort:

By area:

  1. 2026EponaV2: Driving World Model with Comprehensive Future Reasoningcited by 12
  2. 2026CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planningcited by 9
  3. 2026Auto-JEPA: A Latent World Model of Continuous Intent for End-to-End Autonomous Drivingcited by 4
  4. 2026CRAFT: Counterfactual-to-Interactive Reinforcement Fine-Tuning for Driving Policiescited by 2
  5. 2026HiDrive: A Closed-Loop Benchmark for High-Level Autonomous Drivingcited by 1
  6. 2026Driving Intents Amplify Planning-Oriented Reinforcement Learningcited by 1
  7. 2026Teaching Vision-Language-Action Models What to See and Where to Lookcited by 1
  8. 2026SPVC: Structured and Panoptic Video Fixing for Cross-Dataset Driving Scene Renderingcited by 0
  9. 2026Aligning Multi-Trajectory Supervision with Policy Optimization for VLA Drivingcited by 0
  10. 2026CoLT-Drive: Counterfactual Long-Tail Benchmarking and Knowledge-Preserving Adaptation for Driving Affordance Predictioncited by 0
  11. 2026GRAVA: Grounded Reasoning-to-Action Representation and Learning for Autonomous Drivingcited by 0
  12. 2026Planning in the Backbone: DiffAdapterVLA for Native Continuous Trajectory Generation with Driving VLMscited by 0
2 more
  1. 2026FocusDrive: Reasoning with Visual Focus for Autonomous Drivingcited by 0
  2. 2026Many Ways to Succeed: Diversity-Driven RL Fine-Tuning for VLA Generalizationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Unleashing VLA Potentials in Autonomous Driving via Explicit Learning from Failures
  2. 2026ExploreVLA: Dense World Modeling and Exploration for End-to-End Autonomous Driving
  3. 2025V-Max: A Reinforcement Learning Framework for Autonomous Driving
  4. 2025DIVER: Reinforced Diffusion Breaks Imitation Bottlenecks in End-to-End Autonomous Driving
  5. 2025VDRive: Leveraging Reinforced VLA and Diffusion Policy for End-to-end Autonomous Driving
  6. 2025IRL-VLA: Training an Vision-Language-Action Policy via Reward World Model
  7. 2025MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning
  8. 2025DiffusionDriveV2: Reinforcement Learning-Constrained Truncated Diffusion Modeling in End-to-End Autonomous Driving