Awomo-v0.1 on Physics-IQ Verified Benchmark: A small step to physical intelligence
The real world is not made only of language and pixels. An intelligent system that can operate in real tasks and real environments must understand space, objects, motion, constraints, and causal change. For Physical AI, video generation is not merely about generating a visually realistic clip. More importantly, it is about enabling models to learn how the physical world evolves through dynamic visual sequences: how objects move, how actions lead to consequences, and how occlusion, collision, contact, gravity, and failure unfold over time.
Awomo is committed to building concept world models for the physical world, enabling AI to understand physical environments, act reliably, and continuously evolve. Along this direction, we view physical video generation foundation models as one of the key building blocks of Physical AI. They connect perception, prediction, and action. On one hand, they can learn physical laws from large-scale real-world and simulated data. On the other hand, they can provide higher-quality scene understanding, risk prediction, data generation, and policy validation capabilities for robotics and intelligent hardware.
The central challenge in this field is not simply improving visual realism, but enabling models to maintain physical consistency and temporal consistency in complex scenes. A video that looks continuous does not necessarily mean the model truly understands the consequences of actions. Therefore, we care more about whether a model can capture object relationships, state transitions, action constraints, and physical regularities in long-tail scenarios, and gradually turn these capabilities into measurable, iterative, and transferable foundations for Physical AI.
Observations on Physics-IQ Verified Benchmark
Recently, we conducted early validation experiments on the Physics-IQ Verified benchmark to evaluate physical video generation capabilities. Physics-IQ Verified is a benchmark designed to assess the physical understanding of generative video models, covering a range of physical phenomena such as collision, fluids, gravity, material properties, lighting and shadows, and magnetism. Compared with metrics that only evaluate visual quality or text alignment, Physics-IQ Verified focuses more directly on whether generated videos follow real physical laws.
This aligns closely with Awomo’s long-term direction. We do not view video generation models merely as content generation tools, but as a way to validate foundational capabilities for understanding the physical world. If a model can generate dynamic processes that more consistently follow physical laws, it often suggests stronger potential in representing space, objects, and motion. Physics-IQ Verified provides a useful external reference that allows us to observe progress in this direction in a more measurable way.
Based on Awomo’s understanding of Concept World Models, we explored several directions, including decomposing complex physical processes into clearer conceptual units, strengthening the relationships among objects, states, actions, and outcomes, and focusing on the key factors that truly influence physical evolution in a scene. We observed that, after adjustments along these directions, Awomo-v0.1 achieved a clear improvement on Physics-IQ Verified. Table 1 below compares public results from the Physics-IQ Verified leaderboard with our preliminary Awomo-v0.1 result. Compared with the latest SOTA, Cosmos3-Super-Image2Video, Awomo-v0.1 improves the score by approximately 5.6%. The public baseline results in the table are taken from the Physics-IQ Verified benchmark.
| Rank | Model | Physics-IQ Verified |
|---|---|---|
| Input type: I2V | ||
| 1 | Awomo-v0.1 (Ours) | 45.1 ± 0.6 |
| 2 | Cosmos3-Super-Image2Video | 39.5 ± 0.8 |
| 3 | Grok Imagine Video | 34.8 ± 0.6 |
| 4 | Magi-1 24B + GeoPhys (BoN) | 33.7 ± 1.4 |
| 5 | Hunyuan Video 1.5 | 33.4 ± 0.8 |
| 6 | Wan 2.2 | 32.2 ± 0.6 |
| 7 | Kandinsky-WM 1.0 | 30.8 ± 0.9 |
| 8 | Cosmos3-Nano | 30.3 ± 0.6 |
| 9 | Magi-1 24B | 30.2 ± 1.1 |
| 10 | Sora 2 | 26.5 ± 0.8 |
| 11 | P-Video | 25.3 ± 1.8 |
| Input type: multiframe (v2v) | ||
| 1 | Magi-1 24B + GeoPhys (BoN) | 58.2 ± 1.8 |
| 2 | Magi-1 24B | 48.4 ± 1.1 |
The result of Awomo-v0.1 is an early validation: when the generation process pays more attention to concept decomposition and key physical factors, physical video generation capability can show observable improvement. This also supports the direction Awomo has consistently pursued: the core of Physical AI is not merely generating the next frame, but understanding the structure and dynamics of the real world in latent space.
Toward Physical AI Foundation Models
Awomo-v0.1’s result on Physics-IQ Verified is only a starting point. Moving forward, we will continue along the direction of Concept World Models and build physical video generation foundation models that can understand space, objects, motion, and causal change, helping AI move from generation on screens toward understanding, judgment, and action in the real world.