YoCausal: How Far is Video Generation from World Model? A Causality Perspective

arXiv:2605.30346 · cs.CV · Submitted 2026-05-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "YoCausal: How Far is Video Generation from World Model? A Causality Perspective".

Jane: Video generation models are being evaluated for their capacity to understand causality, and this paper introduces YoCausal,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, wrapping up our discussion on "YoCausal: How Far is Video Generation from World Model? A Causality Perspective," the paper really hammers home that there's a substantial gap between current video generation models and what we consider human causal cognition <ref:2605.30346#pg2>.

Jane: It’s clear from the title and the authors that this research is specifically designed to probe whether these models are just mimicking statistical temporal flow or if they're actually learning the underlying mechanics of cause and effect <ref:2605.30346#pg1>.

Lu: The authors successfully introduced a two-level benchmark, YoCausal, that uses real-world videos to test this concept in a scalable way <ref:2605.30346#pg2>.

Meng: Essentially, the paper demonstrates that while scaling models help improve their ability to reason causally, they still haven't achieved true human-level causal understanding on its own.

Tom: The implication is that for the world of AI applications, we need metrics like the ones proposed in this paper to ensure we are aiming for a deeper level of modeling than just temporal prediction <ref:2605.30346#pg2>.

Jane: It really emphasizes that causality isn't something you can just patch in; it has to be learned fundamentally, and this is a crucial distinction for the future of this field.

Lu: This study offers a clear path forward by providing a method to measure where we are relative to human performance on causal reasoning <ref:2605.30346#pg2>.

Tom: So, in summary, the authors of "YoCausal: How Far is Video Generation from World Model? A Causality Perspective" have created a rigorous two-level benchmark that shows that video generation models are currently better at perceiving the arrow of time than they are at grasping true causal relationships <ref:2605.30346#pg2>.

Jane: It’s a really important clarification for listeners to hear: simply seeing the direction of time doesn't equate to understanding causality <ref:2605.30346#pg1>.

Lu: The benchmark provides a concrete way to measure the gap between current AI and human causal cognition across diverse domains like physics, human action, and animal action <ref:2605.30346#pg2>.

Meng: From an engineering viewpoint, this means our focus needs to be on designing architectures that explicitly encode causal dependencies rather than relying solely on implicit statistical temporal patterns.

Tom: It’s a significant piece of research because it validates that scaling parameters can help, but it doesn't magically solve the problem of true causal understanding by itself <ref:2605.30346#pg2>.

Jane: This paper sets a benchmark for future development in ensuring that video generation models are capable of modeling the world in a more meaningful way than just tracking temporal sequences <ref:2605.30346#pg1>.

Conclusion: Tom: So, we've been diving deep into YoCausal, and now it's time to talk about what this whole project really means for us as creators and developers out there on the airwaves. Jane, you started by explaining how this benchmark tests causality versus just temporal patterns.

Jane: Exactly. The title itself is really telling, "YoCausal: How Far is Video Generation from World Model? A Causality Perspective," because it’s asking if these models are truly building a world model or if they're just tracking what happens next in a video sequence.

Lu: I think the authors really nailed the core idea by using those reversed videos as natural counterfactual samples, which is super clever for testing something this fundamental about how machines process time and cause and effect.

Meng: From an engineering standpoint, I'm focused on what this means for deployment; if we can finally measure a model's grasp of causality beyond simple correlation, that tells us we might be closer to building truly intelligent systems for complex simulations.

Lalam: If these models start understanding true causal reasoning, the implications for AI culture are huge because it moves us past simple pattern matching toward genuine world comprehension, which is a massive step in how we design interactive experiences.

Tom: That’s a big picture way to look at it, Lalam; I'm just trying to get the specifics down. So when we look at the authors and their work, what’s the main thing they want us to take away from this whole effort?

Jane: The central message is that perceiving how time flows doesn't automatically mean you understand cause and effect, which is a really important distinction for anyone working with these generative models.

Lu: They established a rigorous two-level measurement system, RSI and CCI, which gives us concrete metrics to see exactly where we stand compared to human performance in this area.

Meng: And the results show that scaling up the model parameters does help improve this causal understanding, which is a practical thing for us to keep in mind when we're thinking about future architecture designs.

Lalam: It shows that the path forward involves designing systems where causal dependencies are explicitly encoded, not just implicitly learned through massive amounts of data.

Tom: So, to wrap up this part of our discussion, YoCausal gives us a clear yardstick for evaluating whether video AI is getting smarter in its reasoning about the world or if it's still stuck in pattern recognition.

Jane: It really sets a high bar for what we expect from these powerful new generation models moving forward.

Tom: And that leads us right into how this impacts our daily work with these systems and where we go next to improve them.

Yu-Zhe Xie, Yu-Hsuan Li, Jie-Ying Lee, Kaipeng Zhang, Yu-Lun Liu

National Yang Ming Chiao Tung University · Shanda AI Research Tokyo

cs.CV

Submitted: 2026-05-28

Updated: 2026-10-03

Comments: Project page: https://www.youzhexie.me/papers/YoCausal/

Code: https://github.com/genmoai/models

Project page: https://www.youzhexie.me/papers/YoCausal/index.html

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 81/100

The gist: Video generation models are being evaluated for their capacity to understand causality, and this paper introduces YoCausal, a novel two-level benchmark designed to rigorously test whether current

Key concepts

Reverse Surprise Index (RSI)
This metric quantifies how well a model perceives the direction of time by comparing denoising losses between forward and reversed video sequences. A higher RSI score suggests the model assigns lower loss to the reversed video, indicating a stronger perception of causality.
Causality Cognition Index (CCI)
This index measures true causal reasoning by subtracting the RSI score of non-causal videos from that of causal videos. It distinguishes whether a model understands cause-and-effect relationships beyond just recognizing temporal patterns in the data.
Violation of Expectation (VoE) Paradigm
Inspired by cognitive science, this paradigm uses temporally reversed real-world videos as natural counterfactual samples. This setup allows researchers to rigorously test if a model's understanding goes beyond simple statistical temporal patterns and into genuine causal cognition.

Terminology

Summary

Video generation models are being evaluated for their capacity to understand causality, and this paper introduces YoCausal, a novel two-level benchmark designed to rigorously test whether current video diffusion models truly grasp causal reasoning or merely overfit to statistical temporal patterns.

The gist

Perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition.

How it works

YoCausal is a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science, utilizing temporally reversed real-world videos as natural counterfactual samples at zero cost. Level 1 introduces the Reverse Surprise Index (RSI), which quantifies arrow-of-time perception via denoising loss. This metric is calculated by measuring the proportion of videos for which the model assigns a lower denoising loss to the reversed video compared to the forward version, formalized as:

RSI(D) = 1/D ∑Di∈D (1/Di) ∑x∈Di (1/Ldenoise(θ; x r) > Ldenoise(θ; x f)). Higher RSI values indicate a stronger perception of the arrow of time and causality.

How it works

Level 2 introduces the Causality Cognition Index (CCI), which disentangles genuine causal reasoning from temporal bias. This is achieved by partitioning the dataset into a causal subset Dc and a non-causal subset Dnc based on whether obvious cause-effect interactions are present, determined by an advanced Vision-Language Model (VLM). The CCI is defined as:

CCI(D) = RSI(Dc) − RSI(Dnc). A higher CCI indicates that the model captures reversed causality cues beyond simple statistical temporal patterns. This partitioning leverages a VLM to automate the dataset splitting, ensuring scalability and reliability, as VLM-based labeling correlates strongly with human annotations.

How it works

The evaluation process involves several key steps:

  1. Constructing an arbitrarily extensible dataset D = ∪ Di, where subsets include General (unconstrained daily-life events), Physics (mechanics, optics), Human Action, and Animal Action.

  2. Level 1 Metric Calculation: Compute the Reverse Surprise Index (RSI) for each subset based on denoising losses of forward and reversed sequences.

  3. Level 2 Stratification: Use a VLM to classify videos into causal (Dc) or non-causal (Dnc) subsets, then compute the Causality Cognition Index (CCI).

  4. Aggregate Ranking: An aggregate causality score is derived by summing each model’s RSI and CCI ranks, with ties broken by RSI rank.

Key Findings

Comprehensive evaluation across 13 state-of-the-art VDMs reveals four key insights: (1) while advanced models perceive the arrow of time and some exhibit preliminary causality cognition, a significant gap persists relative to human model. (2) perceiving the arrow of time is not equivalent to understanding causality. (3) causal cognition correlates partially with intuitive physics but not with aesthetic quality, validating the benchmark’s unique focus. (4) scaling parameters and advancing architectures improve causal cognition, indicating that scaling laws extend to this higher-order reasoning. The analysis further confirms that models ranking high on RSI—such as LTX-Video-2B/13B and HunyuanVideo—also tend to rank highly on CCI, though the two metrics are distinct, demonstrating that a model’s performance requires strong scores on both for robust causal understanding.

Limitations

The paper notes limitations, including the struggle with temporally symmetric events (e.g., Newton’s cradle), where forward and reversed sequences are visually near-identical, which renders RSI ineffective for those cases. Furthermore, computing denoising losses requires access to model weights, which limits external evaluation of closed-source models. The authors argue that while implicit causality is beyond the current VLM partitioning, the benchmark covers the dominant evaluation regime relevant to deployment in robotic manipulation simulation and interactive game engines where visually explicit causality is present. Additionally, a moderate correlation (τ = 0.3333) with human preference suggests that while the benchmark captures causal understanding, it does not fully capture human causal preference due to potential conflation of visual quality with correctness. The study also shows that models ranking high on RSI but low on CCI only perceive the statistical arrow of time, confirming the necessity of both metrics for a holistic view.

Scaling Law and Generational Evolution

The research demonstrates that scaling laws are relevant to causal cognition, as both release date (r = 0.596) and model parameters (r = 0.688) exhibit significant correlations with the aggregate ranking, indicating that increasing parameter count enhances causal understanding.

Improvements for AI systems

Based on the YoCausal paper, here are specific, actionable improvements for AI video generation systems and how those improvements would enhance their capabilities:

  1. Improve Causal Reasoning in Video Generation Models (VDMs) by implementing a two-level evaluation framework (RSI + CCI).

  2. Implement a Reverse Surprise Index (RSI) to quantify the model's perception of the arrow of time by measuring denoising loss differences between forward and temporally reversed videos.

  3. Implement a Causality Cognition Index (CCI) that uses a Vision-Language Model (VLM) to partition datasets into causal and non-causal subsets, thereby disentangling genuine causal reasoning from statistical temporal patterns.

  4. Enable arbitrary scalability of the benchmark by utilizing real-world videos at zero cost via temporal reversal, eliminating the sim-to-real gap inherent in synthetic benchmarks.

  5. Develop a system capable of assessing whether model performance is driven by internal causal understanding (CCI > 0) versus mere arrow-of-time perception (RSI > 0 but CCI near zero).

  6. Ensure that scaling up model parameters and architectural evolution leads to measurable improvements in causal cognition, validating scaling laws for higher-order reasoning.

This improved AI system would be capable of:

  1. Perform more reliable world modeling by understanding the why behind events, not just the what.

  2. Generate videos that exhibit logical cause-and-effect sequences (e.g., correctly simulating a pencil leaving a mark, or a button press turning on a light) rather than merely producing visually smooth, statistically plausible temporal sequences.

  3. Be robust against temporal bias, meaning it won't incorrectly assume causality just because the data was presented in chronological order.

  4. Be more reliable when deployed for tasks requiring physical interaction simulation (like robotic manipulation or autonomous driving), as it can distinguish between physically constrained actions and non-causal visual artifacts.

  5. Provide a quantitative diagnostic tool to tell developers precisely where their model fails—whether the failure is in perceiving time or in understanding causality, guiding targeted architectural improvements.

Abstract

As video diffusion models (VDMs) advance toward world models, a key question arises: do they truly understand causality, or merely overfit to statistical temporal patterns? Existing benchmarks mostly rely on synthetic data, limiting real-world generalization due to the sim-to-real gap. We present YoCausal, a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science. By temporally reversing real-world videos at zero cost as natural counterfactual samples, YoCausal establishes an arbitrarily extensible evaluation protocol. Level 1 introduces the Reverse Surprise Index (RSI), quantifying arrow-of-time perception via denoising loss. Level 2 introduces the Causality Cognition Index (CCI), which leverages a VLM to stratify datasets into causal and non-causal subsets, disentangling genuine causal reasoning from temporal bias. Evaluation of 13 state-of-the-art VDMs reveals that perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition. Project page: https://www.youzhexie.me/papers/YoCausal/

Sources

Related papers