ProWAM Reports a Record on LIBERO-Plus (85.8%). What It Still Can't Do (2610.02508)

ProWAM, from researchers at Shanghai Jiao Tong University, SII, Meta and Imperial College London, reports 85.8% on LIBERO-Plus, a zero-shot resilience benchmark, which the authors call a new state of the art. The next best model in their own table, Image-WAM, scores 83.1%, so the margin is 2.7 points. The paper is arXiv 2610.02508, submitted 1 October 2026, and every number below is the authors' claim unless stated otherwise.

What is ProWAM and who wrote it?

ProWAM is a "progressive world action model" described in World Action Modeling with Progressive Visual Planning. The authors are Fei Zhang, Zhaochong An, Duncan Frost, Yikai Wang, Pengfei Liu, Ya Zhang, Michal Drozdzal and Amir Bar. Per the arXiv HTML version, Zhang and Liu are at Shanghai Jiao Tong University and SII, An, Frost, Wang and Drozdzal are at Meta, and Bar is at Imperial College London.

So this is a joint academic and Meta paper, not a Meta lab release. The arXiv listing links a project page and no other artifact.

What problem does this paper solve?

World action models (WAMs) predict future visual dynamics and actions together from an observation and an instruction. The abstract names two weaknesses in existing designs. Generating dense video rollouts is slow, which hurts long-horizon tasks. Methods that predict only a single future frame skip the question of how to make progress toward the goal.

ProWAM sits between the two: predict a few future frames, ordered by task progress, and use them to guide the actions.

How does ProWAM work?

The method section builds on the Wan2.2-TI2V-5B video backbone with flow matching. A vision-action mixture-of-transformers pairs a video generation expert with a lightweight diffusion-based action expert, and the two share attention heads.

The key idea is the sub-goal. Each sub-goal is a sparse future frame, encoded as a latent slot and tagged with a normalized progress coordinate r between 0 and 1. At r=0 the frame is the current observation, and at r=1 it is the trajectory endpoint. The coordinate enters the network through a sinusoidal embedding and an MLP that produces an adaLN modulation offset.

Other design details the paper states:

• During training the number of sub-goals k is sampled uniformly from 1 to K_max. Predicting only the final goal is the special case k=1 at r=1.
• At inference, sub-goals are re-anchored to the latest observation at every replanning step.
• Video tokens do not attend to action tokens, so the coupling is one-directional: actions read the sub-goals, not the reverse.
• Training is two-stage. Stage 1 is video-only pretraining on action-free video. Stage 2 is joint video and action fine-tuning on robot demonstrations.
• At inference, one video-expert pass caches the sub-goal key-value features and the action head reuses them across its denoising steps. The paper says this cuts video-expert passes from T to 1, roughly 10x fewer.

The authors argue the design scales because sub-goal prediction can be learned from action-free video, which is far more plentiful than robot data. They compare two variants: ProWAM-Base starts from Wan directly, and ProWAM starts from a video-only model pretrained on large-scale robot and egocentric data.

What are the results?

The headline numbers come from two tables. In the table below, "Plus" is LIBERO-Plus, which the paper uses strictly for zero-shot evaluation. All values are success rates in percent, as reported by the authors.

On in-distribution LIBERO, the field is saturated: ProWAM at 98.8 is below both Being-H0.7 and its own Base variant at 99.2. The separation appears only on LIBERO-Plus.

The RoboTwin table shows the same pattern more sharply. "Random" is the out-of-distribution setting.

The abstract's "relative gains of up to +35.9%" refers to this table: 75.7 against MV-WAM's 55.7 is 20.0 points, or 35.9% relative (our calculation). On LIBERO-Plus the same arithmetic gives 2.7 points over Image-WAM, about 3.2% relative.

The abstract adds two more results. On RoboCasa365, ProWAM scores 48.1% overall and 18.2% on the Composite-Unseen split, "ranking 4th overall". In zero-shot real-world tests in novel scenes it reaches 70.0%, against 55.0% for the strongest baseline, a 15-point gap. HyperAI's summary lists four real-world tasks and says ProWAM has fewer parameters than DreamZero, the 14B-parameter DreamZero.

Does the evidence hold up?

Mostly as a directional result, with four caveats.

The LIBERO-Plus margin is small and aggregate-only. A 2.7-point lead over Image-WAM is a single number. LIBERO-Plus spans seven perturbation dimensions: object layout, camera viewpoint, robot initial state, language, lighting, background texture and sensor noise. The main text gives no per-dimension breakdown and no evaluation size, so a reader cannot tell which perturbations ProWAM handles better, or whether the lead is within run-to-run noise.

The ablation points to pretraining. ProWAM-Base scores 80.3 on LIBERO-Plus, 5.5 points below ProWAM, which the authors attribute to missing sub-goal pretraining. Base is also ahead on LIBERO (99.2 vs 98.8) and on RoboTwin clean (87.6 vs 85.8). The LIBERO-Plus gain comes with the pretraining stage, which means part of the record reflects extra data, not only the sub-goal design. The paper does not separate those two effects in the tables above.

Code and weights are not confirmed. The arXiv page links only a project page, and the paper is licensed CC BY-NC-ND 4.0. Without released weights, nobody outside the team can rerun the LIBERO-Plus figure.

The real-world result is thin. A 15-point gap in a handful of tasks is easy to move with a few trials. The abstract reports only the aggregate. Treat 70.0% against 55.0% as suggestive until per-task counts are checked in the paper.

How does ProWAM compare with related work?

The related work splits into three camps, based on the paper's own baselines.

Fast-WAM is the useful contrast. Its authors argue that world action models do not need test-time future imagination and report competitive LIBERO and RoboTwin results. ProWAM's tables show that the no-imagination approach collapses out of distribution. A later paper, LAWA (arXiv 2608.24882), also reports lower generalization for Fast-WAM than for future-aware models, especially with scarce demonstrations. Two independent groups pointing the same direction is more convincing than one table, though both are preprints.

What does a working researcher take from this?

Three practical points:

1. Report out-of-distribution numbers. In-distribution LIBERO spans 85.5 to 99.2 here and ranks models poorly. LIBERO-Plus and randomized RoboTwin separate them.
2. Sparse foresight is a cheap middle ground. The paper's claim is that a few progress-indexed frames give much of the benefit of dense rollouts at a fraction of the video-model cost. That is worth testing on your own backbone.
3. Action-free video is the lever. If the pretraining effect holds, the cheapest resilience gain is more video, not more teleoperation.

Teams collecting robot demonstrations or evaluation data should note that the paper's real-world claim rests on zero-shot DROID-style transfer, which is a data-collection question as much as a modeling one. Disclosure: Humanbit works in data labelling and evals for physical AI, so we have an interest in that question.

What's still unclear?

• Which perturbations drive the LIBERO-Plus score? Without per-dimension numbers, the 85.8% cannot be tied to camera, layout or lighting resilience.
• How much is the sub-goal design and how much is pretraining? The Base ablation suggests pretraining matters, but a fair comparison would pretrain the baselines on the same video.
• Does it replicate? No code or weights are confirmed, and results are the authors' claims until someone reproduces them.
• Does it scale? The RoboCasa365 result is 4th overall and only 18.2% on Composite-Unseen, which suggests long, unseen compositions remain hard.
• Is the real-world gain larger than noise? The abstract gives an aggregate only.

We read the abstract and the HTML full text of the arXiv paper, including the method section and the two main results tables. The claims above are limited to those parts.