UniW(A)M: A Unified Mixture-of-Experts Framework for World and World-Action Models

What it claims

A single diffusion Transformer handles five embodied prediction tasks, called "interfaces." They are distinguished only by video–action attention masks: one action-conditioned world model (AC-WM) and four world-action model (WAM) variants. Each interface is a composition of four atomic capabilities, and each capability has its own expert:

• VGM: action-free video prediction
• VLA: direct action generation
• WM: action-conditioned video prediction
• IDM: video-conditioned action generation

The paper's distinctive claim is that AC-WM co-training improves action generation on every action interface. This includes the two interfaces that never generate video at deployment. A further claim is that a capability-aligned routing scheme (CATR) keeps all four WAM interfaces effective under joint training.

Action-Centric and Action-Only skip future-video prediction at deployment. They are the "fast" interfaces.

How it works

Backbone. The dense FFN sublayers of a pretrained DiT are replaced with MoE blocks initialized from those sublayers. Each block has a shared expert plus the four capability experts.

CATR routing. Routing is per token group, not per token:

• Each video or action target group gets a prescribed eligible set. Video groups use shared + VGM/WM. Action groups use shared + VLA/IDM.
• A learned router produces mixture weights within that set only. Experts outside the set contribute nothing.
• Tokens in a group share the weights.

PAIR (stage II). With the backbone frozen, a classification head on frozen context features picks one of the four WAM interfaces per action chunk. Its training targets are built as follows:

1. Run all four interfaces on the same chunk.
2. Compute the MAE of each interface against the ground-truth actions, then min–max normalise the four errors.
3. Convert the normalised errors to a softmax target with temperature τ.
4. If any fast interface is within 1% of the minimum MAE, replace the target with a one-hot on that interface (the faster one if both qualify). Otherwise keep the soft target.
5. Train with cross-entropy.

So PAIR's training data must include ground-truth action chunks from the target task.

Evidence

• Real world: three manipulation tasks, one of them cloth folding. Average success is 88.3% against 76.7% for the strongest baseline. The excerpt does not name that baseline.
• WorldArena: UniW(A)M is first on all three action-conditioned world-modeling tasks. No scores are given in the excerpt.
• OOD with PAIR: average success rises from 63.3% to 76.7% with the backbone frozen. The excerpt does not say what configuration produced the 63.3%; it is presumably a fixed interface.
• Simulation: LIBERO and RoboTwin 2.0 were evaluated, but no numbers appear in the fetched text.

Limits

• The fetched text stops partway through the PAIR loss (Section 3.3). Results tables, ablations, model size, compute, training data scale and any code or weights release are all absent. Nothing here says the model is available to use.
• The appendices hold the conditional formulations, masks, expert assignments, cache handling and loss, and none of them were fetched.
• PAIR is not zero-shot. It needs labelled target-task chunks, and the excerpt does not quantify how many ("a small subset").
• The paper is under review at ICLR 2027 (arXiv v1, 13 Oct 2026). The headline gains rest on three real-world tasks.