World Models That Know When They Don't Know: Controllable Video Generation with Calibrated Uncertainty

Abstract: Recent advances in generative video models have led to significant breakthroughs in high-fidelity video synthesis, specifically in controllable video generation where the generated video is conditioned on text and action inputs, e.g., in instruction-guided video editing and world modeling in robotics. Despite these exceptional capabilities, controllable video models often hallucinate - generating future video frames that are misaligned with physical reality - which raises serious concerns in many tasks such as robot policy evaluation and planning. However, state-of-the-art video models lack the ability to assess and express their confidence, impeding hallucination mitigation. To rigorously address this challenge, we propose C3, an uncertainty quantification (UQ) method for training continuous-scale calibrated controllable video models for dense confidence estimation at the subpatch level, precisely localizing the uncertainty in each generated video frame. Our UQ method introduces three core innovations to empower video models to estimate their uncertainty. First, our method develops a novel framework that trains video models for correctness and calibration via strictly proper scoring rules. Second, we estimate the video model's uncertainty in latent space, avoiding training instability and prohibitive training costs associated with pixel-space approaches. Third, we map the dense latent-space uncertainty to interpretable pixel-level uncertainty in the RGB space for intuitive visualization, providing high-resolution uncertainty heatmaps that identify untrustworthy regions. Through extensive experiments on large-scale robot learning datasets (Bridge and DROID) and real-world evaluations, we demonstrate that our method not only provides calibrated uncertainty estimates within the training distribution, but also enables effective out-of-distribution detection.

• The paper introduces a new framework that employs calibrated confidence estimates at the subpatch level using proper scoring rules for controllable video generation.
• It integrates latent-space uncertainty quantification with a transformer-based probe to generate interpretable RGB heatmaps that localize hallucinations and out-of-distribution regions.
• Empirical evaluations on robotics benchmarks show low calibration errors (ECE, MCE) and effective identification of artifacts, enhancing model trustworthiness in safety-critical applications.

World Models That Know When They Don't Know: Controllable Video Generation with Calibrated Uncertainty ((2512.05927))

Motivation and Problem Statement

Controllable video generation models, particularly those conditioned on action or text, have become central to high-fidelity visual modeling in robotics and interactive tasks. Despite strong generative and controllable capabilities, these models are prone to hallucinations—generating frames that are inconsistent with the underlying physical dynamics—which limits their utility in applications such as autonomous systems, scalable robot policy evaluation, and visual foresight planning. Critically, contemporary video models offer no internal quantification of their uncertainty, which compromises trustworthiness, particularly in OOD scenarios or when hallucinated artifacts arise. This paper introduces a new framework for uncertainty quantification (UQ) in action-conditioned video generation, achieving dense, calibrated confidence predictions localized at the subpatch level. The focus is on practical calibration and interpretability, as well as the ability to facilitate OOD detection and precise localization of hallucinations.

Methodology

The proposed approach, $\mathcal{C}^3$ (Calibrated Controllable Confidence), is designed for continuous-scale, dense UQ in video models. It is architected around three primary innovations:

1. Training with Proper Scoring Rules: Instead of ad hoc loss functions, the model employs strictly proper scoring rules (Brier score, BCE, or cross-entropy) for simultaneous calibration and accuracy. This ensures, by design, that predicted confidence values are statistically well-calibrated and reflect true epistemic irreducibility, even in high-dimensional action-conditional spaces.
2. Latent Space Uncertainty Quantification: Recognizing the complexity and computational cost of pixel-space UQ, uncertainty is predicted in the compressed latent representations of the video model. A dedicated transformer-based probe, $f_\phi$, processes internal DiT features and action/timestep embeddings to output subpatch confidence estimates, dramatically reducing training and inference cost while maintaining expressivity.
3. Decoding and Visualization of Confidence: Confidence estimates are mapped from latent vectors to interpretable RGB heatmaps for spatial localization of uncertainty. The mapping leverages latent space color maps built from monochromatic frames, interpolated to produce pixel-level uncertainty visualizations that are physically and semantically coherent.

The architecture (Figure 1) integrates a VQ-VAE encoder for constructing the latent space, a diffusion transformer (DiT) for conditional generation, and a UQ probe for dense, channel-wise confidence prediction.

Figure 1: The model simultaneously produces a video sequence and interpretable uncertainty heatmaps from its latent representations, with an independent UQ probe acting on video latents.

The uncertainty objectives are framed as classification tasks: fixed-threshold (single $\varepsilon$), multi-class (confidence bins), or continuous-scale ($\varepsilon$ as query), all supervised with proper scoring rules. The approach generalizes to multiple model architectures and is compatible with diffusion and flow-based video models as well as GAN/RNN-based architectures.

Empirical Evaluation

Datasets and Benchmarks

Evaluation is conducted primarily on the Bridge and DROID datasets, which provide multi-environment, real-world robot manipulation scenarios under diverse observation conditions (e.g., varying camera views, object sets, backgrounds).

Calibration Analysis

Calibration is assessed using ECE (Expected Calibration Error) and MCE (Maximum Calibration Error) across all video frame subpatches. Comparisons are performed between the proposed model variants (fixed-scale, multi-class, and continuous-scale). Results indicate consistently low ECE and MCE, verifying the models' well-calibrated predictive uncertainty. The continuous-scale model offers the best expressivity, with marginal trade-offs in raw calibration due to its flexibility.

Figure 2: Quantitative ECE and MCE analysis demonstrates consistently low calibration error, and reliability diagrams show model confidence closely tracks empirical correctness.

Reliability diagrams further reveal models trend toward conservativeness (mild underconfidence) in low-confidence regimes—desirable in safety-critical systems. Fine-grained analysis (Figure 3) demonstrates precise calibration across error thresholds, with the model appropriately increasing uncertainty as error thresholds become stricter.

Figure 3: The method remains well-calibrated at various thresholds, with greater conservativeness at fine-grained accuracy levels.

Interpretability of Uncertainty

Spatial heatmap visualizations correlate uncertainty with error regions in generated videos. Features such as robot-object interactions, occlusions, and ambiguous dynamics correspond to high-uncertainty responses, especially in hallucinated or OOD regions.

Figure 4: Confidence heatmaps reveal heightened uncertainty around robot motion and interaction—well-aligned with known challenges in robotic vision.

Hallucinated artifacts (unreal objects or dynamic inconsistencies) are reliably associated with localized high-uncertainty, supporting practical deployment for downstream applications like policy verification or post-generation filtering.

Figure 5: Hallucinated video segments are consistently flagged as uncertain by the UQ probe, enabling artifact localization and mitigation.

Further, quantitative analysis using Shepherd's Pi confirms negative correlation between model confidence and error magnitude in most conditions, illustrating faithfulness and interpretability. Failure to observe this correlation only occurs with imbalanced binning in multi-class variants.

OOD Detection

OOD detection is evaluated under shifts in background, lighting, clutter, object configuration, and robot morphology. In all scenarios, the model expresses higher uncertainty in OOD patches while retaining overall calibration, enabling trustworthy error signaling in unfamiliar environments.

Figure 6: OOD detection—regions with previously unseen background/lighting are identified as uncertain, supporting robust change detection in deployment.

Aggregate ECE and MCE in OOD conditions remain low, verifying robustness of calibration outside the training distribution.

Generalization to Diverse Embodiments

Experiments on the DROID dataset, which features different robot platforms and richer scene diversity, yield comparable calibration and interpretability metrics. The method scales to multi-view input, generalizes across robot types, and remains practical for scenarios with limited prior about test domain appearance.

Figure 7: Outlier/hallucinated robot gripper segments are captured as uncertain in new robot/platform ensembles, showcasing multi-environment robustness.

Theoretical Implications

By integrating proper scoring rules into dense, continuous-valued UQ in high-dimensional generative models, the work provides a principled bridge between modern Bayesian calibration theory and practical generative modeling. Latent-space UQ establishes a computationally tractable alternative to ensemble- or variance-based Bayesian methods, avoiding pixel-space pathologies and enabling seamless integration with transformer architectures. The subpatch resolution further supports modular, composable uncertainty integration in planning/planning-under-uncertainty tasks.

Practical Implications and Future Directions

The framework targets central challenges in trustworthy policy evaluation, world modeling for model-based RL, and scalable simulation-to-real transfer, where artifact localization and OOD detection are paramount. Its design is conducive to deeper integration in interactive and safety-critical robotics, data-efficient RL pipelines, and conditional planning frameworks. Future developments should focus on:

• Expanding the diversity and representativeness of training data to tighten calibration guarantees under extreme OOD shifts.
• Scalable uncertainty propagation for long-horizon, temporally-consistent video synthesis.
• Joint training or tighter architecture integration of the UQ probe into large generative models to explore potential for efficiency gains or further synergy.

Conclusion

This work delivers a framework for training controllable video models with robust, dense, and calibrated estimation of their own uncertainty. It enables spatially and temporally resolved identification of hallucinations, interpretable visualization of confidence, and high-precision detection of distributional shift. By leveraging proper scoring rules in latent UQ, it offers a practical and theoretically sound solution to a critical failure mode of modern world models. The approach generalizes across hardware, tasks, and environments—and lays out a foundation for future integration of trustworthy generative modeling in robotics and AI planning.