Physics-only world models cannot predict people, and the fix costs six pipeline stages

The systems the paper indicts are generative world models: Sora, Genie 3, JEPA, Marble ^[1]. The systems it measures are eight language models from OpenAI and Anthropic ^[9]. That substitution deserves attention before the headline numbers do. Menti-Bench consists of 320 text descriptions, 100 picture stories and 28 sound-video clips ^[7], which puts 420 of its 448 scenes, 93.8 percent, outside anything a video model would consume ^[1]. The demonstrated result, then, is that language models choose better actions when forced to write a belief state down first. Whether a rendering model gains the same from a bolted-on mental layer is a different experiment.

Inside the ablations sits a number the framing plays down. Removing the mental channel costs 12.1 F1 points on average; removing the physical channel costs 16.5 ^[12]. Physics is carrying 4.4 more points than mind on the authors' own benchmark ^[2], which is the correct result for a paper arguing that mental variables are a missing addition rather than a replacement. The figure with architectural teeth is 6.4, the loss when the physical and mental transitions are predicted independently instead of jointly ^[12]. The obvious build, a scene predictor running beside an intent classifier with a merge step at the end, is precisely the configuration that gives those points away.

The compute comparison rewards a close read. Direct answering scores 63.3, six-sample self-consistency 77.9, the full pipeline 87.9 ^[10], so the scaffold buys 10.0 points past what majority voting already extracts ^[4]. GPT-4.1 inside the pipeline reaches 84.9 and beats GPT-5.6-Sol answering directly with self-consistency at 83.6 ^[11], a margin of 1.3 ^[6]. Thin, but the ordering holds, and weaker base models gain more from the explicit structure than stronger ones ^[13], which is what you would expect if the scaffold supplies something the model was otherwise improvising.

The gain also concentrates where the benchmark is thickest: 26.4 points on interpersonal scenes against 14.0 on object-focused ones ^[13], nearly double ^[7], with 78 percent of scenes containing at least two characters ^[8]. The benchmark was built to have the property the framework exploits. That is normal for a paper introducing both, and a reason to want someone else's scenes.

Humans score 98.5 against the pipeline's 87.9 ^[10], a gap of 10.6 points ^[5] on six-option questions with a human-written reference solution ^[7]. For a service robot that gap is the whole job. The authors' own illustration is a cup slid across a table, which can be an apology, a deception or an act of care, and only the world model holds the variables that tell them apart ^[4]. To their credit they do not claim to simulate anything: mental states here are hypotheses inferred from behaviour and context, not measurements, and they ask that systems built this way keep their uncertainty and assumptions visible ^[5]. The pipeline at least writes a machine-readable intermediate at every stage, so a wrong answer can be traced to the step that produced it ^[6].