Risk Decomposition for Hallucination-Aware Video Sampling

Chenyan Zheng, Fan Wu, Ruibo Li, Jiaxing Li, Minghao Yang, Bang Zhang, Guosheng Lin

TL;DR

This paper introduces a two-dimensional framework to categorize physical hallucinations in video models into "low-density sampling risk" and "high-density drifting risk." By addressing both risks simultaneously through a novel sampler (HA-Sampler) and a likelihood-based evaluator (HA-Scorer), the authors significantly improve the physical plausibility of generated videos across multiple state-of-the-art backbones.

What Problem Does This Paper Address?

Modern video generative models produce visually stunning results, but they frequently "hallucinate" impossible physics—objects might teleport, disappear, or gain energy out of nowhere. While researchers have tried to fix this using internal model signals (refining the model's own "preferences") or external physics checkers (guiding the model toward real-world laws), there hasn't been a unified theory explaining why these hallucinations happen or how these two approaches interact.

The authors identify a fundamental gap: a model might fail because it samples from an "unfamiliar" low-probability region (low-density sampling risk), or it might fail because the regions it finds "familiar" are actually physically incorrect (high-density drifting risk). Standard methods often tackle one but ignore the other.

Figure 1: Overview of our two-dimensional physics hallucination risk framework. Left: Low- density sampling risk captures generations assigned low likelihood by the video model, while high- density drifting risk captures physical errors in high-likelihood generations. Middle: Model-internal physics-aware signals steer sampling toward higher-density regions of the learned distribution to reduce low-density sampling risk. Right: External physics-aware guidance helps correct the drifting between model-preferred generation and physical plausibility to address high-density drifting risk.

Methods

The paper proposes a dual-pronged approach based on their new risk framework:

  1. Two-Dimensional Risk Framework:
  2. Low-Density Sampling Risk: Captures cases where the model generates something it doesn't "believe" in (low likelihood), like a ball suddenly jumping positions.
  3. High-Density Drifting Risk: Captures cases where the model generates something it thinks is likely, but which violates physics, like a ball bouncing progressively higher.
  4. HA-Sampler (Hallucination-Aware Sampler): A training-free sampling method that integrates external physical guidance directly into the latent refinement process. It uses gradients from an external physics loss to steer the noise perturbation, ensuring the model stays in high-likelihood regions that are also physically valid.
  5. HA-Scorer (Hallucination-Aware Scorer): A method to evaluate generated videos by calculating their exact log-likelihood under the model. It leverages Continuous Normalizing Flow (CNF) mathematics, using a coarse ODE solver and a finite-difference Hutchinson estimator to make likelihood calculation computationally feasible for high-dimensional video data.
Figure 2: Overview of our methods. (a) HA-Sampler uses an external loss on the predicted clean video to guide the perturbation noise. The guided noise is combined with the original endpoint prediction before denoising continues. Grey dashed lines denote the flow of gradients. (b) HA- Scorer integrates the clean latent z0 to the noise endpoint z1 under the fixed base velocity field using a coarse Heun solver. Finite-difference Hutchinson estimator provides the divergence integral. In step 3, zt and zt+1 fall on the same velocity line.

Datasets

The methods were evaluated on two primary benchmarks:

  • Physics-IQ: A benchmark for image-to-video generation that compares model continuations against real-world physical recordings using metrics like Spatial IoU and MSE.
  • VideoPhy: A text-to-video benchmark consisting of 344 prompts designed to test "Physical Commonsense" (PC) across different material interactions (Solid-Solid, Fluid-Fluid, etc.).

Key Insights and Ideas

  • Complementary Strengths: The "aha moment" is that internal refinement and external guidance are not redundant. Refinement acts like a "denoising" step for the model's intent, while guidance acts as a "correction" toward reality. Using one without the other often leads to a trade-off where you either get high-likelihood nonsense or low-likelihood physics.
  • Likelihood as a Proxy for Risk: The authors demonstrate that a model’s internal likelihood is a surprisingly strong indicator of physical consistency. If a model "knows" a video is unlikely, that video is probably a hallucination.
  • Training-Free SOTA: Both HA-Sampler and HA-Scorer are "plug-and-play" and do not require retraining the massive underlying video models, making them highly practical for existing pipelines.

Key Results

  • Benchmark Dominance: HA-Sampler achieved the highest overall scores on both Physics-IQ and VideoPhy. On the Wan2.2-TI2V-5B backbone, it exceeded the base model's Physics-IQ score by +2.10 points.
  • Superior Selection: HA-Scorer proved more effective at picking the most physically plausible video from a pool of candidates than standard Vision-Language Models (VLMs) like Qwen-VL or Gemma.
  • Qualitative Improvement: Qualitative results show that HA-Sampler successfully prevents common errors like objects changing color mid-motion or disappearing into thin air, which were prevalent in baseline "Predict-and-Perturb" (P&P) methods.

Limitations and Future Work

  • Diversity vs. Plausibility: The authors note a potential trade-off: by pushing the model toward high-density, physically valid regions, the diversity of generated outputs might decrease. Preserving creative "long-tail" but valid generations remains an open challenge.
  • Computational Overhead: While HA-Scorer is optimized, it still requires multiple model forward passes (NFE), which may be slow for real-time applications.
  • Future Applications: The authors suggest these risk measures could be used to filter synthetic data for training robotics models, ensuring that the "world simulators" robots learn from are physically accurate.

Summary updated

Abstract

Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Existing methods leverage model-internal or external signals to improve physical plausibility, but their roles and limitations require a unified explanation. We propose a two-dimensional physical-hallucination risk framework that defines low-density sampling risk and high-density drifting risk. The first risk captures generations assigned low likelihood by the video model, for example, a ball suddenly teleporting despite the model favoring continuous motion, while the second risk captures physical inconsistencies in high-likelihood generations, for example, a ball bouncing progressively higher without additional energy. Our analysis explains how model-internal signals steer sampling toward higher-density regions of the learned distribution to address low-density sampling risk. However, higher model likelihood does not guarantee physical plausibility, motivating external guidance to address high-density drifting risk. Based on this framework, we develop Hallucination-Aware Sampler (HA-Sampler), which applies external physical guidance within latent refinement to generate physically plausible videos, and HA-Scorer, which derives a likelihood-based signal from the generator itself for evaluating generated videos using continuous normalizing flow (CNF), with coarse ODE integration and finite-difference divergence estimates. Experiments on Physics-IQ and VideoPhy benchmarks support our theoretical analysis and demonstrate that our methods outperform the compared sampling and scoring SOTA baselines.
BibTeX
@misc{zheng2026risk,
  author = {Chenyan Zheng and Fan Wu and Ruibo Li and Jiaxing Li and Minghao Yang and Bang Zhang and Guosheng Lin},
  title = {{Risk Decomposition for Hallucination-Aware Video Sampling}},
  year = {2026},
  eprint = {2610.26069},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.26069},
}
Figure