• Survey Highlights
• Timeline At A Glance
• How To Use This List
• Render-and-Decode (58)
• Latent-Only (38)
• Video-Generation-Free (13)
• Contributing — add or update a WAM paper
• Citation
• Related Resources

• The rule is simple: a model is a WAM when its predicted future is used by the action path. The future may help produce, score, train, or check actions.
• The survey uses two views. One view asks what the method predicts. The other breaks each method into substrate, backbone, action coupling, and deployment style.
• The list has three families: Render-and-Decode (58), Latent-Only (38), and Video-Generation-Free (13).
• The trend is easy to see: newer methods often generate less of the future when control does not need full video. The ⚡ mark highlights papers with an explicit speed, memory, or data-efficiency move.
• The paper connects the list to broader questions about interaction, causality, persistence, physical plausibility, data, evaluation, and open challenges.

The figure counts papers in this README by first arXiv year. The 2026 count is year-to-date.

Papers are grouped by the survey's three families and listed by first arXiv month. The columns expose the main comparison points used in the survey.

These methods produce a visible or otherwise rendered future, such as RGB video, RGB-D, or multi-view frames. The action module then reads, tracks, or inverts that future.

These methods predict the future inside hidden states or compact signals instead of decoding full video. Actions use features, flow, masks, value maps, teacher embeddings, or similar carriers.

These methods do not put a video generator in the loop. The useful future signal comes from an LLM, VLM, diffusion policy, JEPA / DINO predictor, audio model, geometric model, or another non-video predictor.

Contributions are welcome. This is meant to be a living list maintained by the WAM community, and adding a paper usually takes one row.

We follow the survey's definition. A paper belongs here if it includes an explicit prediction about a future observation or future-derived signal (pixels, latents, features, flow, masks, affordance / value maps, audio, tokens, and so on) and uses that future to produce, score, train, or check actions.

Not included: generic world-model simulators with no action use, plain video generators, direct VLAs with no predicted future, and future heads that are discarded before action use.

1. Fork this repository and edit README.md.
2. Pick the table for the matching design philosophy (Render-and-Decode / Latent-Only / Video-Generation-Free).
3. Insert your paper in chronological order (by first arXiv month) using the row template below.
4. Open a pull request. A short note on why the paper is a WAM (which future, how it becomes action-facing) helps review.

Row template:

| YYYY-MM | [⚡ Full Paper Title (ShortName)](https://arxiv.org/abs/XXXX.XXXXX) | Backbone | Substrate | One-line TL;DR. |

Drop the ⚡ unless the paper makes an explicit efficiency move. Keep the TL;DR to one sentence, and prefer the substrate vocabulary already used in the tables (e.g. Pixel (decoded), Pixel (latent), Feature (tap), Geometric, Affordance).

If this list or the survey helps your research, please consider citing:

@article{202606.1403,
	doi = {10.20944/preprints202606.1403.v1},
	url = {https://www.preprints.org/manuscript/202606.1403/v1},
	year = 2026,
	month = {June},
	publisher = {Preprints},
	author = {Qiuhong Shen and Shihua Zhang and Yue Liao and Qi Li and Zhenxiong Tan and Shizun Wang and Shuicheng Yan and Xinchao Wang},
	title = {World Action Models: A Survey},
	journal = {Preprints}
}

• 🌐 Survey homepage & interactive paper explorer: https://world-action-models.github.io/
• 📝 Preprints record: https://www.preprints.org/manuscript/202606.1403/v1
• 📄 Survey PDF: https://world-action-models.github.io/wam-survey-nus.pdf

The curated content of this list is released under CC0-1.0. Linked papers remain under their respective authors' rights.