Driving world models fail at true counterfactual prediction; a training-free fix recovers the lost events

How Can Driving World Models Do Counterfactual Prediction?

Jiaru Zhang, Can Cui, Yi Xu, Xin Ye, Ruqi Zhang, Ziran Wang

cs.CV

2026-08-12

Direct prediction ignores what actually happened next, dropping episode-specific events. Transporting observed evidence into the counterfactual view lifts recovered fraction from about 0.3 to 0.7.

What problem this solves

Many driving world models (Vista, Drive-WM, plus industrial models from Waymo and Google DeepMind) claim to do counterfactual prediction: given a recorded drive, they answer "what if the ego had accelerated or braked instead" by generating a video. The field assumes that feeding the model an alternative action is enough.

This paper argues that the standard practice is fundamentally broken. A counterfactual asks what would have happened in this specific realized episode under a different action, so the answer must stay tied to that episode. But the model only gets the shared history (the prefix common to both actions) and the alternative action. It never gets the factual continuation, the frames actually recorded after the action point. The car that emerges from a side street is not yet visible in the shared history, so the model has no reason to preserve it and can produce a fluent, plausible video that omits the event. The video looks good but answers the wrong question.

The authors name the failure with Pearl's ladder of causation. Direct prediction sits at rung 2 (intervention); a true counterfactual lives at rung 3 and requires abduction, using the factual continuation to infer the state of this particular world. Direct prediction skips that step.

Method

After the diagnosis they offer a deliberately simple, training-free pipeline in four stages that maps onto abduction, action, and prediction.

Abduce: a frozen Depth Anything V2 estimates depth for each frame of the factual continuation, lifts every pixel to a 3D point, and rebuilds the observed part of the world.
Transport: because the camera is rigidly mounted, the executed and target trajectories give the relative pose between the camera that did film the scene and the one that would have. The 3D points are reprojected into the counterfactual view, producing a warped evidence image and a support mask marking which pixels have factual backing. Residual holes are filled from nearby frames (multi-frame filling).
Complete: regions backed by evidence are held fixed, and the frozen world model fills the rest. For the diffusion model, sampling starts midway from a noisy encoding and restores the evidence region at the matching noise level after each denoising step; for the autoregressive VQ model, tokens covered by evidence are pinned while the rest generate normally.
Combine: passing through the model blurs the transported pixels, so this stage restores the reliable transported pixels and blends the boundary.

Everything runs at inference time with all weights frozen.

Results

Real driving can never supply a counterfactual answer (each episode happens once), so the authors turn to the CARLA simulator, where the same world can be replayed under a different action to yield a counterfactual ground truth. The benchmark has 186 cases, 72 placements, 3 towns, and 3 scenario types (side-street emergence, lead cut-in, lead brake), at 10 fps, 576x320, with a 15-frame history. Each case runs the world three times: the factual log (event occurs), the counterfactual ground truth (target trajectory, event still occurs), and a null reference (target trajectory, event never occurs). Two metrics: recovered fraction (Rec, computed with DINOv2 and CLIP; 0 means the event is fully omitted, 1 means the counterfactual reference is reproduced) and LPIPS against the counterfactual ground truth.

Two frozen backbones are tested: Vista (diffusion) and DrivingWorld (autoregressive VQ). Overall numbers:

Model	Metric	Direct	Ours
Vista	Rec (DINOv2)	0.38	0.70
Vista	Rec (CLIP)	0.33	0.65
Vista	LPIPS down	0.423	0.169
DrivingWorld	Rec (DINOv2)	0.31	0.67
DrivingWorld	Rec (CLIP)	0.24	0.64
DrivingWorld	LPIPS down	0.291	0.211

Every direct-prediction recovered fraction sits below 0.5, meaning it lands closer to the event-free null reference. On side-street cases alone, Vista's Rec (DINOv2) climbs from 0.29 to 0.75. Ablations show the event signal comes mostly from Transport, while Complete and Combine remove seams and cut LPIPS. A key control: transport only helps with evidence from the correct time and episode. The wrong time drops the recovered fraction back near direct prediction; another episode still recovers the event but LPIPS jumps because the geometry is wrong.

Why it matters

For anyone building driving world models or simulators, this is a wake-up call: the widely advertised "counterfactual ability" does not hold in the strict causal sense, and direct prediction is only intervention prediction. The claim is pinned down with a clean same-backbone, same-history, same-action comparison that removes suspicion about the models themselves.

The method is an offline analysis tool, not an online decision maker. It reads the factual continuation, which exists only after the fact, so it fits retrospective uses like incident review, safety auditing, and liability assessment. Being training-free and weight-frozen means gains transfer automatically as depth models and world models improve.

Limitations

The authors' main stated limit is the open-loop setting: surrounding agents follow scripts and do not react to the ego. Over longer horizons they would react (a pedestrian might stop if the ego slows), and transported evidence would freeze behavior that the new action should have changed, such as a pedestrian who would have stopped but keeps walking. Extending to closed-loop, reactive agents is the explicit next step. The method also reads the factual continuation, so it cannot serve decision time, when the outcome is still unobserved.

What stays open after reading: both world models are evaluated outside their training render domain (they were not trained on CARLA). The causal analysis does not depend on the render domain, but a model trained on the benchmark's domain would be more convincing. Monocular depth error degrades the transported evidence and leaves seams, and the quality of unsupported holes inherits the backbone's completion ability; the paper does not quantify depth-error sensitivity.

Terms

Counterfactual prediction: asking, for a specific episode that already happened, what would have occurred under a different action, with the answer tied to that episode.
Pearl's ladder of causation: observation (rung 1), intervention (rung 2), and counterfactual (rung 3); higher rungs are generally not identifiable from lower-rung information alone.
Abduction: the step that uses observed facts to infer the underlying world state, the key to counterfactual prediction.
Factual continuation: the frames actually recorded after the action point, evidence that direct prediction ignores but a counterfactual must use.
Recovered fraction: whether a prediction is closer to the counterfactual ground truth or the event-free null; 0 is full omission, 1 reproduces the reference.