All four papers covered here landed on arXiv within 36 hours of each other last week. Nobody coordinated this. Nobody announced a shift. This is just the point where the entire field quietly abandoned the failed premise of pixel space world models.

For three years every major world model paper trained on video. They predicted pixels. They learned physics as an emergent side effect of guessing the next frame. They failed. Every single one drifted, broke momentum, leaked energy, produced physically impossible motion when rolled out more than 10 steps.

That era is over. Every work here operates directly on physical state. On vertices. On coordinates. On material properties. On forcing functions. None of them use pixels as the primary representation. None of them require hard coded inductive biases for physics. All of them work.

The week world models left pixel space

This is a phase shift, not incremental improvement. Up until this month, every production world model you had heard of: DreamerV3, Genie, Google World Model, VideoPoet, all operated in view space. All predicted pixels. All of them had fundamental failure modes that could not be fixed with more parameters or more training data.

You cannot learn consistent physics from pixels. The projection from 3d world state to 2d pixels destroys information. It introduces ambiguities that can never be resolved. No amount of training will ever fix that. It is a representational dead end.

This cluster of papers is the first coherent demonstration that we do not need to accept that tradeoff. We can run diffusion, transformers, autoregressive modelling directly in unstructured world coordinate space. And it works better. Not a little better. 20-40% better on every physical consistency metric.

What changed: nobody is enforcing physics anymore

This is the most counterintuitive result across all four papers. For the last decade every neural physics paper opened with a paragraph about how you must inject inductive biases. You must hardcode conservation of momentum. You must enforce rigidity constraints. You must regularise for energy. You must bake Newton's laws into the loss function.

Every single one of these papers throws all of that out.

They do not enforce anything. They do not add physics terms to the loss. They do not hardcode constraints. They just run standard transformer attention and standard diffusion denoising directly on raw coordinate values. And they preserve momentum. And they preserve rigidity. And they conserve energy. Better than every model that had those rules baked in.

That is the single most important result published this entire year. Nobody expected this.

PhysiFormer: diffusion directly on vertex coordinates

PhysiFormer is the cleanest demonstration of this principle. There are no tricks here. There is no clever latent space. There is no graph convolution.

Input is N objects, each with V vertices. Each vertex is just 6 raw 32 bit floats: x y z position, vx vy vz velocity. That is the entire input. No embeddings. No object tokens. No boundary detection.

They flatten that entire list of vertices, throw it into a diffusion transformer, and train it to denoise future trajectories. That is it.

It learns rigid bodies. It learns elastic deformation. It learns collisions. It learns friction. Nobody told it any of those things exist. It just sees 100,000 simulated trajectories and extracts the rules.

The efficiency trick: factorised attention over three axes

This is the detail that makes this architecture scale to real world use. Standard self attention over N vertices across T timesteps would be O((N*T)^2). That dies completely at 2 objects and 20 timesteps.

PhysiFormer factorises attention independently along three separate axes. It runs one attention pass over time for each vertex. One pass over space for each timestep. One pass over objects for each position. Total complexity O(NT(N + T + O)).

This is not an approximation. This is an explicit decomposition. And it works. It maintains full permutation invariance over objects. It never needs explicit object boundaries. It will correctly handle 1 object or 12 objects exactly the same way, no retraining required.

At release this model runs 8 objects, 100 vertices each, 200 timesteps, on a single A100. That is enough for most robotics manipulation tasks today.

SubdivAR: mesh generation stopped being a graphics problem

SubdivAR is the sleeper paper here. Everyone will ignore it for six months then realise this is how every 3D asset will be generated by 2027.

Mesh subdivision was always a graphics operation. You had a coarse mesh, you ran Catmull-Clark, you got a smooth surface. All neural attempts up until this point tried to predict local vertex offsets. All of them broke on sharp edges, thin geometry and open surfaces.

SubdivAR reframes this as an autoregressive world model problem. Each subdivision level is just the next timestep. The model predicts the next scale of geometry exactly the same way PhysiFormer predicts the next position of a vertex.

They built a new curated dataset of 40,000 high quality meshes with full multi level subdivision supervision. It outperforms every existing neural subdivision method by 14-19%. It works on every edge case every prior method failed on.

ICMPG: closing the loop between LLMs and physics

This paper solves the problem everyone has been complaining about with LLM motion generation. LLMs write great motion plans. None of them can stand up. None of them know that you cannot put your foot through the floor.

ICMPG does not try to teach the LLM physics. It does not fine tune it. It does not inject physical embeddings. It runs a very simple closed loop at inference time:

1. Ask the LLM for 8 candidate motion sequences
2. Run all 8 through a lightweight 100 step physics simulator
3. Score each one for physical validity and semantic match
4. Feed the best one back into the LLM context window for the next step

That is it. No new training. No custom model. Just run model predictive control using the LLM as a proposal generator.

It beats every existing text to motion baseline on both physical plausibility and semantic fidelity. And you can drop any LLM into it. Any motion model. Any physics engine.

EO-WM: world models don't just run in lab sims

This is the one that should make everyone realise this is not just for robotics and games. EO-WM is a world model for the entire surface of the Earth.

It does not predict satellite pixels. It predicts physical state: vegetation stress, soil moisture, surface temperature. It takes weather forcing as an input. It correctly predicts that a 2 week heatwave will kill grass. It correctly predicts that the same temperature spread over 2 months will do nothing.

Prior models got this wrong every single time. They did not accumulate stress over time. They just mapped weather pixels to land pixels.

EO-WM separates baseline climate from anomaly forcing. It accumulates physical stress over time. It reduces NDVI decline prediction error by 5.6% and directional accuracy by 7.8%. That is an enormous improvement for this domain. That difference is enough to change agricultural insurance payouts.

Benchmark breakdown: what the numbers actually mean

Model	Domain	Primary Representation	Relative Improvement Over SOTA	No Hardcoded Physics	Open Source Release
PhysiFormer	3D Mechanics	World Vertices	+31% trajectory accuracy	✅	2026 Q3
SubdivAR	Mesh Geometry	Mesh Vertices	+18.8% Hausdorff Distance	✅	2026 Q3
ICMPG	Human Motion	Joint Kinematics	+27% physical plausibility	✅	2026 Q4
EO-WM	Earth System	Surface State	+7.8% directional hit rate	✅	2026 Q3

Note that PhysiFormer comes very close to a hard coded physics simulator, despite having zero knowledge of physics written into it. It is not perfect. But it is good enough for almost every real world use case.

The unstated tradeoff everyone is ignoring

None of these papers talk about failure modes. None of them talk about what happens when you roll them out 1000 steps.

All of them are probabilistic. All of them will drift. They drift slower than pixel models. They drift much slower. But they still drift.

This is not a bug. This is the tradeoff. When you remove hard constraints you get better average behaviour and graceful degradation. But you give up guarantees. You will never get perfect conservation. You will never get zero drift.

For robotics, for graphics, for earth forecasting this is absolutely the correct tradeoff. For engineering simulation it is not. Nobody is talking about this yet.

What comes next

This stack is complete. We now have all the components:

• We can simulate rigid and soft bodies
• We can generate high resolution geometry
• We can connect this to natural language
• We can run this at planetary scale

Over the next 12 months every major player will ship a world model built exactly this way. None of them will use pixels. None of them will hardcode physics.

Nobody won an argument. Nobody proved anyone wrong. The field just moved. One week everyone was building pixel world models. The next week nobody was. That is how progress happens.