The Score That Shapes the Forecast

TL;DR for operators

The World Action Critic (WAC) Model¹ tests whether evaluating an action’s consequences should influence how an AI system represents its predicted future, rather than merely score that future afterward.

The strongest evidence comes from a controlled NAVSIM-v1 ablation: allowing critic gradients to update predictive features produces 93.3 PDMS, compared with 92.3 when those gradients are detached. WAC also improves candidate-ranking metrics against Drive-JEPA, although the aggregate planning-score advantage is only 0.1 point and the comparison uses different candidate pools.

For autonomous-driving teams and developers of candidate-based decision systems, this supports testing joint prediction-and-evaluation training. It does not establish universally better decisions. The approach depends on externally supplied quality labels, adequate candidate-distribution coverage, and predictions that have not been directly verified for rejected actions. Real-world safety and inference efficiency remain untested.

Accurate Future Prediction Can Still Lose the Decision

Consider an autonomous-driving planner evaluating several trajectories. One offers faster progress, another avoids a possible collision, and a third remains comfortably within the drivable area.

The planner must anticipate what each action could produce, then decide which consequences matter. Accurate environmental prediction helps, but it cannot guarantee a good choice if the internal representation omits a detail critical to safety or compliance.

This matters because predictive models compress future information. Training primarily to reconstruct observations can favor prominent environmental variations while retaining less information about smaller features that determine action quality. A scoring module attached afterward cannot reliably recover information already discarded.

WAC addresses this mismatch by letting consequence-evaluation errors influence the predictive representations themselves.

Its controlled experiment provides a reason to investigate that design. With recurrent prediction and factual future supervision retained, removing the gradient connection between the evaluator and predictive features lowers planning performance by 1.0 PDMS point (Section 4.5, Table 3). The difference isolates a training relationship that an ordinary final-score comparison would obscure.

The Critic Changes Learning Without Entering the Prediction Stream

WAC integrates three functions: an actor proposes alternative actions, a world model predicts action-conditioned future states, and a critic evaluates the resulting consequences.

In its driving implementation, the system uses two front-camera observations with ego-state and command information. It generates 32 candidate trajectories, each covering four seconds through eight future poses. Candidate-specific recurrent processing predicts latent future states over two rollout steps.

A predictive latent representation is a compressed internal description of a possible future. WAC’s central architectural choice concerns which information that description learns to retain.

The recurrent Transformer uses asymmetric attention. Critic tokens can attend to intermediate predictive features, but prediction tokens cannot directly attend to historical critic tokens. This preserves a directional separation in the forward computation.

Training remains coupled. Critic prediction errors propagate gradients into the features used by the world model, allowing consequence supervision to influence what those features encode.

The distinction is essential: the critic does not need to feed its tokens directly into future-state prediction to change how future-state representations are learned.

The paper’s first theorem explains the potential benefit in a restricted linear setting. Prediction-only learning retains directions associated with large future-state variance. Adding critic supervision changes the representation-selection criterion to include decision relevance:

Here, $\Sigma$ captures future-state covariance and $vv^{\top}$ emphasizes the direction relevant to the critic target. Under the theorem’s linearity, covariance, squared-loss, and representation-capacity assumptions, this can preserve decision-relevant features that prediction-only training would discard (Section 3.4, Theorem 1).

It establishes a mechanism under specified assumptions, not an optimality guarantee for the full nonlinear Transformer.

Rejected Actions Receive Quality Labels, Not Observed Futures

Joint learning faces a fundamental data constraint. A driving log records what happened after the executed action; it cannot reveal the actual future that would have followed every rejected trajectory.

WAC addresses this through two complementary supervision channels.

For recorded actions, future observations provide factual latent targets. The model learns to predict those targets, using stop-gradient representations and, in the driving implementation, Smooth L1 losses for scene and ego-state predictions.

For proposed alternatives, an external NAVSIM evaluator supplies action-quality targets. The critic learns from these labels even though the alternatives have no observed future-state targets. Training also includes best-of-N action regression and jointly optimizes proposal, prediction, and critic objectives (Section 3.3, Equations 5–8).

This extends quality supervision to counterfactual actions without claiming that their predicted futures have been observed or independently verified.

It also creates an evaluation dependency. If the external scoring system systematically misrepresents desired outcomes, the learned critic can reproduce that mismatch, and its gradients can influence the predictive features. More integrated training therefore makes evaluator reliability a representation-learning concern, not merely a final-ranking concern.

The Largest Controlled Difference Is Not the Headline Benchmark Margin

WAC reports the highest aggregate PDMS in its NAVSIM-v1 comparison: 93.4 versus 93.3 for Drive-JEPA and 93.1 for DrivoR (Section 4.3, Table 1).

WAC also records the highest reported Ego Progress, at 91.0, alongside No at-fault Collisions of 99.0, Drivable Area Compliance of 98.3, Time to Collision of 95.9, and comfort of 100.0. It does not lead every safety or compliance component.

The more informative evidence comes from the within-WAC ablation.

Source: Section 4.5, Table 3. All variants use a four-pose prediction interval with focal weighting disabled.

These controlled tests serve different purposes. Gradient detachment tests whether consequence supervision reaching predictive features contributes to planning performance. The one-shot variant tests recurrence. Removing factual future targets tests whether observed dynamics supervision remains valuable alongside critic feedback.

All three contribute in this experimental configuration. A separate training-configuration ablation finds that adding focal weighting raises the selected two-step model from 93.3 to the reported 93.4 (Section 4.5, Table 4). That additional 0.1 point should not be attributed to critic-gradient propagation.

Candidate-scoring diagnostics offer another perspective. Relative to Drive-JEPA, WAC increases Spearman ranking correlation from 0.468 to 0.551, raises pairwise ranking accuracy from 74.02% to 77.11%, and reduces positive overestimation of low-quality candidates from 44.16 to 34.02 (Section 4.4, Table 2; Appendix C).

The final measure concerns candidates whose evaluator PDMS is below 50. Its reduction is particularly relevant to avoiding inflated assessments of poor alternatives.

However, each planner is evaluated on its own candidate pool. The comparison therefore reflects differences in both the proposals available and how they are scored; it does not isolate critic quality on identical actions.

The Selection Guarantee Ends at the Candidate Set

WAC’s second theorem bounds expected selection regret relative to the best proposed candidate, not the globally best feasible action.

The bound depends on three quantities: critic fitting error, discrepancy between the external evaluator and intended task quality, and coverage of the test candidate distribution by the supervised distribution. It also scales with the square roots of candidate count and the coverage constant (Section 3.4, Theorem 2; Appendix A.2).

This distinguishes two failure sources. The critic may estimate evaluator scores poorly, or the evaluator may accurately express a target that does not match the intended objective. A strong critic cannot eliminate the second problem.

Candidate generation introduces a further limitation. Even perfect ranking cannot select an action that was never proposed. Increasing candidate count also does not automatically strengthen the stated regret bound.

What Decision-System Teams Should Measure Separately

Cognaptus inference: For autonomous-driving planners and other systems selecting among generated actions, WAC provides a concrete architectural hypothesis worth testing: consequence feedback may be more effective when it shapes intermediate prediction features than when evaluation is trained only downstream.

The corresponding evaluation should distinguish three outcomes:

1. Selected-action quality: Does the final chosen action improve under the intended task objective, including its safety and compliance components?
2. Candidate-ranking quality: Does the critic order the same candidate actions more accurately when proposal quality is held fixed?
3. Poor-action overestimation: Does the system avoid assigning excessively favorable scores to clearly undesirable candidates?

Teams would also need to test evaluator-target alignment and whether production candidate distributions remain covered by the critic’s supervised training distribution.

The current evidence does not settle deployment economics. WAC uses recurrent candidate-conditioned computation, but the paper establishes neither a real-time latency advantage nor lower inference cost.

Decision-Aware Prediction Still Requires Independent Validation

The empirical findings are confined to NAVSIM-v1 autonomous-driving evaluation. Published baseline comparisons use potentially different sensing and training configurations, and the paper does not supply numerical confidence intervals or significance tests alongside its main tables, despite describing paired-bootstrap resampling in Appendix C.

Qualitative driving examples help illustrate trajectory ordering around collision risk, road compliance, and progress, but they do not establish real-world collision-risk reductions.

WAC’s contribution is therefore best understood as evidence for a particular learning mechanism: allowing action-quality feedback to influence predictive representations can improve simulated planning, while factual future supervision and recurrence continue to matter.

For systems that must choose among uncertain alternatives, the architectural lesson is to test what consequence evaluation teaches the predictor—not merely how accurately the evaluator scores its output. Whether that translates into reliable operational decisions depends on the quality of the external objective, the alternatives actually proposed, and validation beyond the benchmark.

Cognaptus: Automate the Present, Incubate the Future.