Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model
Haoyu Zhao, Zihao Zhao, Tianyu Deng, Ziqin Xu, Zihao Zhang, Xudong Wang, Jinxiang Guo, Chen Gao, Ziyi Ye, Yeying Jin, Jiaxi Gu, Zuxuan Wu, Shuicheng Yan
cs.CV
2026-09-16
MiniMax-H3 scores 41.97% on 517 implicit tasks that hide answers in image, video, and audio. Video continuation leads at 56%; audio disambiguation is last at 27.4%.
Most video-generation benchmarks write the answer into the prompt. VBench and TC-Bench pair text with an image and name the target event. VideoPhy and WorldModelBench add physical-commonsense checks, but the event to render is still stated in language. A model can pass by painting what it was told.
Omni-modal generators accept text, images, video, and audio in one stack, so they could in principle see different slices of the same event. Accepting four modalities is not the same as aligning them. Researchers at NUS, Fudan, and Tencent treat MiniMax-H3 as a testbed for a harder question: if the critical semantics are withheld from the prompt and scattered across observations, can the model infer what it should generate?
The protocol is implicit generation. The text names the task type and omits the target event. Each modality supplies only partial evidence; the intended event is recoverable only by jointly reading the inputs. The generated video is a behavioral readout: success means the clip satisfies the semantic constraints supported by the observations, not that it matches a single reference video.
Four scenarios, 517 items, 29 subcategories:
Visual sources include HiFi-UMI-2K, VISTA-UMI-5K, and LLaVA-Video-178K; audio comes from FSD50K and ESC-50, plus synthetic clips from ChatGPT Voice and Seedance 2.0. Prompts are drafted with ChatGPT and Qwen3, then run through a 10-loop expert review on scene complexity, inferential richness, condition alignment, and prompt implicitness. Three experts score each generated video independently and then cross-check. A pretty clip that misses the task condition is a failure.
Overall success on 517 items is 41.97%. The four scenarios split hard.
| Task | N | Success rate |
| VDR video continuation | 100 | 56.00% |
| AVIR audiovisual | 71 | 47.89% |
| MSR multi-view spatial | 200 | 43.50% |
| ADR audio disambiguation | 146 | 27.40% |
Inside VDR, traffic is 9/9, cartoons 20.00%, puzzles 16.67%. Inside MSR, threading is 75.00% and pouring 58.30%, while cleaning, articulation, and transport sit around 33% to 35%. Inside ADR, nature sounds reach 54.50%; machinery is 17.60%. The VDR-ADR gap is 28.60 points, the largest spread among the four.
Failures cluster into four patterns: the cue is grounded on the wrong object or action; relevant objects appear but the implied interaction never happens; contact, geometry, or state transitions violate physics; object states established in the prefix are dropped later. The paper reads intermediate generated frames as a Chain-of-Frames, a visual analogue of a reasoning trace, which only helps if those frames are actually reasoning.
The suite splits "the model accepts multimodal inputs" from "the model uses them." For people training Omni-Models, the stubborn gap is acoustic grounding and cross-view spatial binding, not short-horizon visual continuation. For people hoping MiniMax-H3 can stand in as a world model or an embodied rehearsal engine, it currently behaves more like a visual continuer than an audio disambiguator.
No other Omni-Model is scored, so the numbers describe this one checkpoint. The item design is reusable: take the answer out of the prompt and see whether generation still matches the observations.
Only MiniMax-H3 is tested. The four scenarios differ in data, prompts, and generation targets, so VDR beating ADR is not a clean video-versus-audio comparison.
A failed clip could come from misunderstanding the inputs, weak cross-modal binding, or the video decoder drawing the wrong pixels. The paper flags the missing ablation: drop one modality, or swap the audio while freezing the visual input. Scoring is three experts with no automatic metric, which will not scale cheaply. Some conditions are synthetic; real acoustics and multi-view noise may not match. Traffic has 9 items and cartoons 5, so those subcategory rates will move a lot on a rerun.