Emotion as a causal variable: AffectWorld scores five world models, and counterfactual sensitivity tops out at 0.070

World-model evaluation gains an emotion causal chain: given a history video, a model must generate a future that is physically plausible and in which emotion drives a person's actions and reshapes the social state. Earlier world-model benchmarks mainly assessed physical plausibility and perceptual quality, with variables that are physical, geometric, or task-oriented, so emotion-driven world evolution was largely untested. AffectWorld uses 1,498 everyday video samples from short-video platforms, annotated with event context, plausible emotions, action tendencies, and atomic world-state constraints, and splits the chain into World-State Consistency, Emotion-Action Consistency, and Causal Consistency.

The evaluation adds a counterfactual branch: the scene and event are held fixed while only the target agent's emotion is replaced, testing whether the generated future changes accordingly. Because one emotion can lead to several equally plausible futures, the annotation does not fix a single target world state but instead marks required, prohibited, and directional constraints. The counterfactual subset contains 233 samples, each sharing one history observation and diverging only in emotion and future evolution; the five models score 0.027–0.070 on counterfactual emotion sensitivity, with HunyuanVideo-1.5 at 0.0701, 29% above Wan2.1 and 164% above CogVideoX-5B.

Scoring combines automatic perception metrics with vision-language and language model semantic judgments, calibrated against human ratings. Task 1 takes the geometric mean of identity preservation, spatial structure stability, and temporal coherence, while Tasks 2 and 3 separate perception from judgment. A vision-language model first converts the generated video into a time-anchored description of observable cues, then a language model scores each candidate against a graded rubric, with each semantic judgment repeated three times at temperature 0.3 and the median retained; human calibration used 60 generated videos, and judge-human agreement was reported on a separate held-out set of 150 generated videos.

The benchmark targets everyday social scenes in short videos, covering 16 scenarios grouped into solo, dyadic, and group settings, with emotions and actions drawn from closed taxonomies of 6 emotion families and 7 action categories. It suits researchers working on world models and video generation who want to check whether a model treats emotion as an intervenable causal variable rather than only recognizing it. Scores are reported separately for the three tasks and their sub-dimensions, which helps locate the level at which a model falls short.

Emotions and social consequences are often only partially observable, so the annotations characterize plausible transitions rather than uniquely correct latent states or futures. Semantic evaluation relies partly on vision-language and language model judges and may inherit evaluator bias despite human calibration. The benchmark focuses on short everyday videos and may not fully represent diverse cultural contexts or long-term social dynamics.