About the collection

Where world models meet action

The collection connects foundational work, vision-language-action models, World-Action Models, datasets, evaluation metrics, benchmarks or simulators, core components, and related resources. These resources support the study of prediction, planning, and control.

Browse by major category and subcategory, compare model approaches in the quadrant view, or search for an idea or author. Each entry links to its original research source, with PDF, DOI, code, and project links when recorded.

Three ways into the collection

The 8 major categories organize entries by their main contribution. Subcategories describe more specific roles and connections. 3 entries have no major category recorded and remain in the library.

The quadrant view crosses architecture—One Model or Dual-system—with prediction paradigm—joint prediction or inverse dynamics (IDM). 149 entries are recorded in Q1–Q4. The remaining states—Outside quadrants, Not applicable, Pending verification—are kept separate. Joint training alone does not establish a One Model architecture.

The 9 original research topics remain available as a separate thematic view. 390 entries have no original Primary Category recorded; this does not mean their major category, subcategories, or quadrant are missing.

Components of WAMs contains core encoders, pretrained backbones, tokenizers, and canonical action heads, such as VAE, CLIP, DINO, Wan, and Cosmos. Using a component does not make a complete WAM or a task-specific training method a component. Foundational work is limited to historically foundational research first published before 2026. Related resources collects relevant surveys, runtimes, supporting representations, and adjacent research. Dedicated neural simulators belong in Benchmarks & simulators. Explicit category reviews draw on the reading evidence and remain authoritative during synchronization. Source-review status and architecture or prediction-paradigm judgments remain separate; unsupported assignments stay unresolved. Missing abstracts and dates remain unfilled; recorded BibTeX can be copied from a detail page. Submission dates and publication years are shown separately, and year filters use only recorded years.

Help the collection grow

Found a relevant paper or a detail that needs updating? Open an issue on GitHub. A useful submission includes the title, authors, source link, and suggested major category or subcategories when known.

For classification corrections, include the paper link, the field to update, and supporting research sources. Maintainers record accepted category reviews locally or update the Notion source before exporting it. Accepted updates appear in the catalog and project website.