# DriveReferee: Geometric Safety Verdicts Need Not Be Learned for Driving World-Action Models

arXiv: https://arxiv.org/abs/2609.22762 · HTML: https://mlsanity.com/paper/2609.22762/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in World models for autonomous driving, one of 11 directions in World models.
- Builds on 19 of its 35 references that are in the map.
- First posted on arXiv 19 Sep 2026.

## Placed in

- World models / [World models for autonomous driving](https://mlsanity.com/wm/world-models-for-autonomous-driving/)
- End-to-end autonomous driving / [Generative simulation for driving](https://mlsanity.com/driving/generative-simulation-for-driving/)

## Builds on (most cited first, 19 of 19)

- [NuPlan: A closed-loop ML-based planning benchmark for autonomous vehicles](https://mlsanity.com/paper/2106.11810/) (2021)
- [Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillation](https://mlsanity.com/paper/2406.06978/) (2024)
- [Cosmos 3: Omnimodal World Models for Physical AI](https://mlsanity.com/paper/2606.02800/) (2026)
- [On a Formal Model of Safe and Scalable Self-driving Cars](https://mlsanity.com/paper/1708.06374/) (2017)
- [Hydra-MDP++: Advancing End-to-End Driving via Expert-Guided Hydra-Distillation](https://mlsanity.com/paper/2503.12820/) (2025)
- [SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving](https://mlsanity.com/paper/2603.29163/) (2026)
- [DriveVA: Video Action Models are Zero-Shot Drivers](https://mlsanity.com/paper/2604.04198/) (2026)
- [DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning](https://mlsanity.com/paper/2604.01765/) (2026)
- [DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving](https://mlsanity.com/paper/2605.28544/) (2026)
- [EvaDrive: Evolutionary Adversarial Policy Optimization for End-to-End Autonomous Driving](https://mlsanity.com/paper/2508.09158/) (2025)
- [MindDrive: An All-in-One Framework Bridging World Models and Vision-Language Model for End-to-End Autonomous Driving](https://mlsanity.com/paper/2512.04441/) (2025)
- [Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation](https://mlsanity.com/paper/2606.15869/) (2026)
- [UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation](https://mlsanity.com/paper/2607.05133/) (2026)
- [Beyond Imitation: Learning Safe End-to-End Autonomous Driving from Hard Negatives](https://mlsanity.com/paper/2605.19771/) (2026)
- [Test-Time Trajectory Optimization for Autonomous Driving](https://mlsanity.com/paper/2606.07170/) (2026)
- [Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving](https://mlsanity.com/paper/2605.21139/) (2026)
- [ForgeDrive: Bidirectional Cross-Conditioning for Unified Visual-Action Generation in Autonomous Driving](https://mlsanity.com/paper/2606.31226/) (2026)
- [PerceptDrive: Perception Prior World-Action Modeling with Adaptive Expert Routing for End-to-End Autonomous Driving](https://mlsanity.com/paper/2607.20175/) (2026)
- [DriveVer: Lightweight Trajectory Evaluator as Test-Time Verifier for Autonomous Driving](https://mlsanity.com/paper/2607.00399/) (2026)

## Abstract (arXiv)

Generative world-action models (WAMs) jointly generate future video and vehicle actions, while their action branches remain primarily optimized by expert imitation. Yet imitation provides no explicit closed-loop geometric verdict for generated trajectories, making verification important during both training and deployment. Closed-loop evaluators can check collision and drivable-area violations, but require privileged scene state unavailable at deployment. Existing approaches often close this gap by learning a verifier from sensor features. For these geometric checks, the rule itself is explicit. For example, collision is determined by whether the rolled-out ego footprint overlaps occupied vehicle space. What is unavailable at deployment is the scene state needed to apply the rule. We introduce DriveReferee, which uses a learned geometry readout to predict the scene representation from camera observations and executes the geometric safety rule directly rather than learning it. The resulting analytic referee evaluates collision and drivable-area safety from a scene state and candidate trajectory. During training, it scores self-sampled trajectories on ground-truth state and distills the resulting preferences into the WAM policy. At deployment, the same referee evaluates generated trajectories on this predicted state and selects a safer alternative when needed. The analytic referee requires no verdict-specific training, and its decisions follow an explicit geometric rule. Under matched candidates and inference budgets, it matches or outperforms all learned-verifier and heuristic baselines. Given the same predicted state and trajectory, learning the verdict provides no measurable downstream gain despite requiring tens of thousands of evaluator-labeled training examples. On the full NAVSIM navtest, DriveReferee reaches 92.02 PDMS with single-camera visual input and no external training data.