Tool
Compare models on the register
Short answer: This is a side-by-side comparator for the register of world action models. It currently covers 17 models, classified by paradigm — inverse dynamics, joint prediction, representation only, latent actions, and hybrid. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and each model’s paper.
How to use: Pick 2–4 models below. The selection is encoded in the URL, so you can share it (e.g. /compare?m=fast-wam,cosmos-policy). What this does not show: any figure not stored in this repository — numbers are only shown where the model card cites them. Download the same register rows as JSON or CSV.
Register fields
These fields are taken directly from the register row in lib/models.ts. Where a value is unconfirmed, it shows “not stated in register”. Each source links to the model's paper or official page.
| Field | Fast-WAM | NVIDIA/Cosmos Policy |
|---|---|---|
| Lab | not stated in register | NVIDIA |
| Year | 2026 | 2026 |
| Paradigm | representation only | joint prediction |
| Backbone | Wan 2.2 (5B) | Cosmos |
| Source | https://arxiv.org/abs/2603.16666 | https://arxiv.org/abs/2601.16163 |
| Notes | Skips video generation at inference. Fig. 4 on one RTX 5090D towel-fold: Fast-WAM 190 ms, Fast-WAM-Joint 580 ms, Fast-WAM-IDM 810 ms, Pi-0.5 180 ms. | Encodes actions as extra latent video frames. |
Sourced claims
Fast-WAM
- The paper asks whether WAM gains come from video modeling during training or from imagining pixels at inference. Fast-WAM keeps the training objective and skips future prediction at test time. Yuan, Dong, Liu & Zhao, 2026
- The architecture is a Mixture-of-Transformers: a Wan2.2-5B video DiT plus a 1-billion-parameter action expert (about 6 billion parameters in total) with shared attention. At inference the video branch encodes the observation in a single forward pass. Yuan, Dong, Liu & Zhao, 2026
- Controlled variants in the paper stay close to imagine-then-execute WAMs when video co-training is kept, while removing that co-training causes a larger drop. Yuan, Dong, Liu & Zhao, 2026
NVIDIA/Cosmos Policy
- The method adapts Cosmos-Predict2-2B through a single stage of post-training on robot demonstrations, with no architectural modifications. Actions, proprioception, future images and state values are encoded as latent frames inside the video diffusion process. Kim et al., 2026
- At test time it can act as a direct policy or plan with best-of-N sampling: generate candidate action chunks, imagine future states, rank them by predicted value, and execute the highest-value chunk. Kim et al., 2026
Limits the authors state
Fast-WAM
—
NVIDIA/Cosmos Policy
- The authors report that model-based planning is slow (around 5 seconds per action chunk), which they say may limit dynamic tasks; that effective planning needs substantial rollout data beyond the demonstration distribution; and that they only studied best-of-N search with one layer in the search tree. Kim et al., 2026
Pick models
FAQ
What is being compared on this page?
Only fields stored in the register (name, lab, year, paradigm, backbone, notes, source URL) and the sourced claims and stated limitations from each model’s card. Where a field is unconfirmed in the register it shows “not stated in register”.
Where do the claims come from?
From each model’s paper or official page, cited on its model card. Classification follows Moritz Reuss’s field map on the NVIDIA Technical Blog (June 2026) and each model’s own paper.
What is not shown here?
Figures not stored in this repository — including latency or benchmark tables that are not in the model card — do not appear here. This page does not add facts beyond what the register and model cards already record.