Tsinghua Study: Robot World Models Only Need the First Denoising Step (arXiv 2609.34981)

A Tsinghua-led team reports that when a robot policy borrows features from a video generator, almost all the generalization benefit comes from the first denoising pass, not from generating clean future frames. In arXiv 2609.34981, their Simple-WAM scores 79.5 on the LIBERO-Plus perturbation benchmark, against 67.7 for an explicit world action model and 53.8 for a latent one, while running 3.8 times faster than the explicit design. All numbers are the authors' claims from simulation benchmarks and have not been independently replicated.

The paper is "What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling" by Renping Zhou, Zanlin Ni and nine co-authors, with Gao Huang as corresponding author. The affiliations are Leap Lab at Tsinghua University, the University of Science and Technology of China and the Beijing Institute of Technology. It was submitted on 28 September 2026 and revised on 29 September 2026 (arXiv).

What problem does this paper solve?

It asks whether a world action model (WAM) must generate future video at inference time to keep the benefits that motivated WAMs in the first place. WAMs predict the future alongside actions during training, but video denoising is expensive, so it is disputed whether the future must still be generated when the robot is running (abstract).

Two camps exist. Explicit WAMs denoise the future into clean frames with every action chunk. Latent WAMs drop the future entirely at inference to run faster. The authors find that latent WAMs match explicit ones on in-distribution tasks but lose the generalization that justified WAMs.

How does Simple-WAM work?

Simple-WAM keeps the future video tokens as pure Gaussian noise and runs the video expert once, so a single feature conditions all action denoising steps (HTML v2). The training objective is unchanged: L = L_act + λ·L_vid with λ = 3.0.

The setup, as described in the paper:

• Backbone: a pretrained Wan2.2-5B video diffusion transformer, reusing its text encoder and VAE, plus a separate action expert.
• Chunking: action chunks of H = 32, with 9 video frames per chunk and K = 10 action denoising steps.
• Training data: benchmark demonstrations only. The paper marks Simple-WAM as having no embodied pretraining.
• Noise schedule: with probability p = 0.5 the future tokens are fully noised (τ = 1). Otherwise τ follows the explicit paradigm's schedule.

The three paradigms differ in inference cost. Using the paper's notation, explicit costs K·C_vid + K·C_act, latent costs C_cur + K·C_act (one pass on the current frame only), and Simple-WAM costs C_vid + K·C_act.

What is the first-denoising-step finding?

The authors take an explicit-trained model, run its video expert for only the first few of K = 10 steps, freeze that feature, and measure the result without retraining. On LIBERO-Spatial, one pass (where the future tokens are still pure noise) already beats the latent model by 26.44, 8.2, 13.4 and 89.8 points across four settings (HTML v2).

The remaining nine passes add at most 1.8 points in three settings and subtract 1.62 in the fourth. The authors' reading is that the benefit comes from "preparing the future, not generating it": the pass forms a representation the action expert can use, and extra fidelity in the generated video does not translate into better actions.

What are the results?

Under a matched backbone, data and budget, the explicit model beats the latent one on every generalization axis while the two are close in distribution (HTML v2, Table 1).

The last row is the striking one. A latent model cannot use action-free video from a held-out task, because it never generates or conditions on the future, so the video data does nothing for it.

Table 2 adds the baselines. These are all lab-reported numbers, with no embodied pretraining for the three WAM rows except where noted.

Speed is the other half of the claim. Simple-WAM runs "at 3.8× the speed of the explicit paradigm and within 1.2× the latency of the latent one", and the project page gives 74.7 ms per action chunk (project page). Latency was measured on a single RTX 5090.

Does the evidence hold up?

Partly. The controlled comparison is the paper's strongest asset: explicit and latent differ only in an attention mask that decides whether action tokens may attend to future video tokens, with architecture and hyperparameters following Fast-WAM (HTML v2). That isolates the variable the authors care about.

Several caveats apply.

• Not a clean sweep. The paper says Simple-WAM leads explicit on 7 of 8 measurements. FastWAM-Joint is higher on RoboTwin task generalization with video (47.3 against 44.5), and Lingbot-VA beats Simple-WAM on the Init (83.0 against 82.1) and Cam (86.4 against 55.6) perturbation factors.
• Pretrained policies still win on perturbation. π0.5, which uses embodied pretraining, scores 84.4 on the LIBERO-Plus average against 79.5.
• Evaluation size is thin in the text. The paper's text gives no trial counts for the simulation benchmarks, so it is unclear how much a point of success rate is worth. Real-world evaluations use 30 trials each, so a single trial moves a score by roughly 3 points.
• Small-data regime. Training uses benchmark demonstrations only. Task-generalization scores on held-out tasks without video are low for every method, with Lingbot-VA's 15.2 on LIBERO the highest.
• One backbone. Everything rests on a 5B video model. Whether the first-step effect holds at other scales or with embodied pretraining is not tested in the paper.
• Self-reported speed. Latency comes from the authors' own measurement on one GPU.

How does it compare with related work?

The study sits in a short lineage of papers that question test-time future generation. Fast-WAM is the latent baseline here: it drops future tokens at inference but keeps them in training. Simple-WAM argues that this drops too much, because the action expert loses the future-conditioned representation entirely.

The authors also name Faster-WAM (Zhao et al., 2026) as concurrent work that likewise stresses inference-time future conditioning to hold up under perturbation. They say they differ by using a broader three-axis protocol (perturbation, data efficiency, task generalization) and controlled experiments (HTML v2). Evaluation relies on LIBERO, LIBERO-Plus and RoboTwin 2.0.

What does a working researcher take from this?

Three practical points follow, if the results replicate.

1. Do not pay for full video denoising at deployment by default. The paper's evidence is that one pass over noise captures nearly all of the gain, at 74.7 ms per chunk rather than a multi-step cost.
2. Keep the future tokens in the input. Dropping them cost 14 points on LIBERO-Plus in the matched comparison, and eliminated the value of action-free video.
3. Evaluate beyond in-distribution success. In distribution, explicit and latent differ by 0.9 points. Perturbation and low-data splits are where the difference shows.

If you build on the released code, check the repository linked from the project page (LeapLabTHU/Simple-WAM, with weights on Hugging Face).

What's still unclear?

Several questions remain open.

• Why does noise work? The authors speak of "misalignment" between video fidelity and what the action expert needs, but a mechanistic account of what a pass over pure noise encodes is not established by the paper's evidence.
• Does it scale? The 5B backbone and absence of embodied pretraining leave open whether the gap between paradigms shrinks or grows with scale.
• Does it transfer to real robots broadly? Real-world evaluation uses 30 trials per setting, which gives wide uncertainty.
• Independent replication. All numbers are lab-reported. Concurrent work such as Faster-WAM points the same direction, but it comes from a different group with its own protocol, so agreement on method is not yet agreement on numbers.

Paper Huddle covers physical AI data and evaluation, an area where Humanbit (data labelling and evals) operates, so read this analysis with that affiliation in mind.