A field guide to world models that actually do things

What it does

This repository is a living literature review that accompanies the first systematic survey of World Action Models — systems that merge predictive world modeling with action generation for embodied AI. It catalogs papers into a strict taxonomy (Cascaded versus Joint architectures, with sub-tags for diffusion, autoregressive, geometric, and latent approaches) and provides a concise summary blog for each entry. There is also a benchmark leaderboard and a linked open-source Paper2Blog skill that generates the summaries.

The interesting bit

Instead of dumping links, the maintainers force a rigid, color-coded classification scheme onto a nascent field — tagging every paper by its architectural approach — which makes the repo feel more like an interactive field map than a static bibliography. The generated paper blogs are produced by an open-sourced pipeline, so you can inspect the summarization mechanics rather than treating them as black-box blurbs.

Key highlights

• Visual taxonomy roadmap splitting WAMs into Cascaded and Joint families with further diffusion and autoregressive subdivisions.
• Every paper entry carries architecture tags (e.g., Explicit, Implicit, Learned, Geometric) and links to a dedicated summary blog.
• Includes a benchmark leaderboard with performance trend visualizations.
• The Paper2Blog summarization skill used to generate reading blogs is itself open-sourced.
• Community-driven updates via issues and PRs; the maintainers explicitly commit to keeping the list current.

Caveats

• Some recent papers already carry a Needs-Taxonomy-Review badge, signaling that the classification framework is still shifting as the field defines itself.
• The repo is overwhelmingly a paper index; the open-sourced Paper2Blog skill is linked but not documented within the README itself.

Verdict

Worth bookmarking if you are doing literature reviews in embodied AI or trying to understand how world models are being repurposed for action generation; skip it if you are looking for runnable training code or off-the-shelf models.