Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model. By Haoyu Zhao and twelve others, from the National University of Singapore, Fudan University, and one other institution. Published on the sixteenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Omni-modal generative models aim to handle text, images, video and audio in one unified model. MiniMax-H3 is an example, combining multimodal context understanding with joint audio-visual generation in a shared latent framework. That raises a question: does multimodal alignment improve a model's reasoning about the world, and what new kinds of evaluation do omni-modal inputs make possible? Existing evaluation frameworks for video generation and world models tend to be limited in input modalities, and use prompts that closely match the target video. Such prompts leave little for a model to infer. If the prompt already describes the video to be made, a model can succeed by following the description, and the test says little about whether it understood anything.

The work introduces a comprehensive evaluation framework organised around four complementary dimensions of physical world reasoning. It is designed specifically to exploit the multimodal inputs of an omni-model, with novel tasks that require the model to integrate complementary information across modalities. The four dimensions of physical world reasoning organise what is tested. The multimodal inputs determine how it is tested: evidence is spread across channels, so understanding has to come from combining them.

Four input scenarios are used: implicit prompts paired with multiple frames, audio with an image, prefix videos, and audio with video. The key design constraint is that every single modality provides only partial evidence about the underlying event. The model has to reason jointly over the complementary cues to infer hidden event states and future dynamics. No one input is enough by itself. Each task becomes a small puzzle whose pieces arrive through different senses. Frames, a prefix of video, an image or a sound each carry a fragment, and the event has to be assembled from the fragments before the model can say where things go next.

Across five hundred and seventeen evaluation instances, MiniMax-H3 achieves an overall success rate of forty-one point nine seven percent. Video-based Decision Reasoning gives its highest success rate, fifty-six percent. Audio-based Disambiguation Reasoning is the weakest, at only twenty-seven point four percent. Put plainly, the model succeeds on somewhat fewer than half of the instances overall, a little over half when reasoning from video to a decision, and only about a quarter of the time when audio has to resolve what is going on.

The spread between those two points to where the difficulty lies. The model does best when the decisive evidence is visual, and worst when audio has to settle an ambiguity.

The authors conclude that effective multimodal integration remains the key to fully exploiting diverse input modalities. The framework's contribution is to make that integration unavoidable. Each task is built so that the answer only exists when the pieces are put together. The project is available on GitHub.