YoCausal: Benchmark for Causal Video Generation

• YoCausal is a two-level benchmark for evaluating video generation by distinguishing between temporal regularity detection and genuine causal cognition using reversed real-world videos.
• It introduces the Reverse Surprise Index (RSI) and Causality Cognition Index (CCI) to measure arrow-of-time perception and causal reasoning, respectively, in video diffusion models.
• Empirical findings show that state-of-the-art models detect temporal order but still lag in human-level causal understanding, highlighting a significant gap in world-model evaluation.

to=container.exec code qq的天天中彩票 ＿影音先锋 code: {"cmd":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:YoCausal&start=0&max_results=3'\ntry:\n print(urllib.request.urlopen(url,timeout=20).read()[:300])\nexcept Exception as e:\n print('ERR',e)\nPY"]}】【：】【“】【stdout":"ERR timed out"}򐀨truncated 232 chars YoCausal is a two-level benchmark for evaluating video generation from a causality perspective. It is motivated by a specific question: as video diffusion models (VDMs) advance toward world models, do they truly understand causality, or merely overfit to statistical temporal patterns? The benchmark is inspired by the Violation of Expectation (VoE) paradigm from cognitive science and uses temporally reversed real-world videos as natural counterfactual samples, thereby establishing an arbitrarily extensible evaluation protocol (Xie et al., 28 May 2026).

1. Problem formulation and scope

YoCausal addresses a distinction that is often blurred in discussions of generative video systems: the distinction between modeling temporal regularities and modeling causality. Its motivating premise is that progress in video generation does not, by itself, establish progress toward world modeling in the stronger causal sense. The benchmark is therefore posed not as a fidelity or aesthetics test, but as an evaluation of whether a model’s internal regularities align with causal structure.

A central part of this framing is the critique of existing evaluation practice. Existing benchmarks are described as relying mostly on synthetic data, which limits real-world generalization because of the sim-to-real gap. YoCausal responds by shifting the evaluation basis to real-world videos rather than synthetic scenes alone. This suggests a deliberate attempt to test causal competence under naturalistic visual statistics rather than under narrowly scripted environments.

The benchmark’s orientation is diagnostic rather than interventionist: it evaluates causal cognition in trained models, rather than proposing a new training objective for causally grounded video generation. In that sense, YoCausal belongs to the methodology of model assessment and benchmarking rather than to architectural design.

2. Benchmark construction

The core construction principle is temporally simple but conceptually loaded: real-world videos are reversed in time and the reversed versions are treated as natural counterfactual samples. This reversal is said to occur at zero cost, and it makes the protocol arbitrarily extensible because any suitable real-world video can be converted into an evaluation pair without requiring manual synthesis of counterfactual scenes.

This design choice ties the benchmark directly to the VoE paradigm. In cognitive science, VoE-style evaluation probes whether an observer registers a mismatch between expected and observed event structure. YoCausal translates that logic into model evaluation by asking whether a VDM responds differently to forward and reversed event sequences. A plausible implication is that temporal reversal functions here as a stress test for learned assumptions about physical and causal order.

The use of real-world videos is not merely a dataset choice; it is the principal methodological claim of the benchmark. By grounding the protocol in real videos and their reversals, YoCausal attempts to preserve visual realism while perturbing temporal and causal plausibility. This makes the reversal operation the benchmark’s central intervention.

3. Level 1: Reverse Surprise Index

The first evaluation layer is the Reverse Surprise Index (RSI). RSI is introduced as a measure that quantifies arrow-of-time perception via denoising loss. In other words, the benchmark’s first question is whether a model can discriminate forward from reversed temporal order in a way reflected by its own generative objective.

Because RSI is defined through denoising loss, it is anchored in the internal scoring machinery of diffusion-based video generation rather than in an external classifier. This makes the measure intrinsic to the VDM under test. The benchmark therefore asks whether the model’s learned generative structure is sensitive to temporal asymmetry.

RSI is explicitly about arrow-of-time perception, not about causality per se. That separation is crucial. A model may detect that one temporal direction is statistically more plausible than the other without having a deeper representation of causes and effects. This suggests that RSI functions as a necessary but not sufficient indicator for the broader notion of causal cognition that the benchmark seeks to isolate.

4. Level 2: Causality Cognition Index

The second evaluation layer is the Causality Cognition Index (CCI). CCI is introduced as a measure that leverages a VLM to stratify datasets into causal and non-causal subsets, thereby disentangling genuine causal reasoning from temporal bias. If RSI tests generic temporal asymmetry sensitivity, CCI tests whether such sensitivity tracks causally meaningful structure rather than merely chronology.

The introduction of a VLM at this stage gives the benchmark a hybrid architecture: generative video models are evaluated, but a vision-LLM is used to organize the evaluation space. The role of the VLM is not to replace the VDM’s judgment, but to partition the data so that temporal bias and causal understanding are not conflated.

CCI therefore formalizes a methodological distinction between two abilities. The first is detecting directional temporal regularities. The second is allocating that sensitivity in a way that reflects causal content. This suggests that YoCausal treats causal cognition as a stricter criterion than arrow-of-time recognition alone.

A common misconception in model evaluation is that temporal plausibility and causal understanding are effectively the same phenomenon. The two-level design of YoCausal is expressly constructed to reject that equivalence.

5. Empirical findings

YoCausal reports an evaluation of 13 state-of-the-art VDMs. The principal empirical conclusion is that perceiving the arrow of time does not imply understanding causality, and that a significant gap persists relative to human-level causal cognition (Xie et al., 28 May 2026).

This claim gives the benchmark its main interpretive force. A model may score well on a temporal asymmetry criterion while still failing the stronger causal criterion. The result is not just comparative ranking among VDMs, but a conceptual separation between temporal competence and causal competence. In benchmark terms, RSI and CCI are not redundant; they probe different properties.

The reference to a gap relative to human-level causal cognition also positions YoCausal within the literature on model-to-human comparison. The benchmark is not satisfied with showing that models respond to temporal reversal; it asks whether they do so in a way commensurate with human causal judgment. That comparative framing aligns the benchmark with broader attempts to evaluate world-model claims against cognitive rather than merely statistical standards.

6. Significance within world-model evaluation

YoCausal is best understood as an evaluation protocol for the claim that advanced video generation constitutes a world model. Its underlying argument is that world-model status cannot be inferred from sample quality or temporal smoothness alone. Causal understanding must be evaluated separately, and it must be evaluated in a setting where temporal bias can be disentangled from causal reasoning.

Its methodological significance lies in three linked decisions: using real-world rather than predominantly synthetic data, constructing natural counterfactual samples by temporal reversal, and separating arrow-of-time perception from causal cognition through RSI and CCI. Together, these decisions define a benchmark whose target is not generic video realism but causal competence under realistic visual conditions.

A plausible implication is that YoCausal reorients the assessment of VDMs away from the question of whether they can continue a sequence plausibly and toward the question of whether their plausibility judgments are causally structured. On that reading, the benchmark does not deny that current VDMs learn powerful temporal statistics; it argues that such statistics remain insufficient evidence for world-model-level causal understanding.

In this sense, YoCausal marks a conceptual boundary within video-generation research. It proposes that chronology detection, temporal pattern modeling, and causal cognition should be treated as distinct evaluation targets, and that progress toward world models requires success on all three rather than on the first alone.