MiniMax-H3 accepts every modality but reasons across them less than half the time

Accepting multiple modalities as input is not the same as reasoning across them. That distinction drives this evaluation of MiniMax-H3, an omni-modal generative model that can ingest text, images, video, and audio simultaneously and produce joint audio-visual output. The central question is whether richer input actually produces richer inference — or whether the model simply ignores non-dominant evidence and leans on textual priors.

The framework built here is deliberately adversarial to that shortcut. Rather than writing prompts that describe the target event, the researchers strip the critical information out of the text and distribute it across modalities. A static image might show three cups of different materials; only the sound of one falling tells you which cup moved. A prefix video shows a ball rolling into the road; only anticipatory reasoning about what follows next — a child, a vehicle stopping — constitutes a correct continuation. The generated video becomes the behavioral readout: if the output is consistent with the multimodal evidence, the model reasoned correctly; if it merely looks plausible, it did not.

Across 517 expert-verified instances spanning four scenarios and 29 subcategories, MiniMax-H3 succeeds on roughly four in ten cases overall. Video-based Decision Reasoning is the strongest scenario at 56%, while Audio-based Disambiguation Reasoning bottoms out at just over a quarter of cases. The gap between those two — nearly 29 percentage points — is the paper's most telling number. The model handles temporal continuation from video reasonably well but struggles badly when acoustic evidence must determine which visual event occurred. Machinery sounds, contact sounds, and alert sounds are particularly difficult; the model can generate a visually coherent scene while completely ignoring what the audio implied.

The failure taxonomy is precise and useful. Four recurring patterns emerge: grounding the right event to the wrong entity, generating the relevant objects without instantiating the interaction, violating physical constraints like contact dynamics or valid state transitions, and failing to preserve object states across frames. These are not perceptual failures — the outputs often look fine. They are reasoning failures, and the generation-based evaluation makes them visible in a way that multiple-choice benchmarks cannot.

The honest limitation the authors name is that a failed output could stem from incorrect input understanding, weak cross-modal integration, or generation errors downstream — and the current design cannot separate those causes. Ablations that remove individual modalities or swap audio while holding visuals fixed would help, and the paper flags this as future work.

What the framework contributes is a cleaner question than most video-generation benchmarks ask. Existing benchmarks test whether a model generates what it is told. This one tests whether a model can determine what it should generate. That reframing matters for anyone building systems where the model must act on partial, distributed evidence rather than a fully specified instruction — which is most real deployment scenarios.