About World Models Radar

Two tiers: Editor picks are chosen by a human. Every other paper is summarized automatically by a language model and labelled “auto-summary”; nobody has reviewed it.

What this is

World Models Radar filters the daily arXiv flow down to papers that meaningfully advance world models: latent dynamics (Dreamer/RSSM), JEPA-style predictive models, generative video world models, and action-conditioned models for control. Every paper links to its arXiv abstract page.

How papers are selected and scored

An automated pipeline collects new papers from arXiv categories cs.LG, cs.CV, cs.RO and cs.AI, keeps those that match world-model topics, and stores them as logged. Logged papers get a composite score from five components: novelty (30%), state-of-the-art claims (25%), released artifacts such as code or weights (20%), community signals (15%) and author track record (10%). A score that ignores the two social signals is also computed so that recent sleeper results are not buried. The weights are provisional and may change.

Every paper that passes the world-model filter is published automatically with a machine-written summary. An editor then promotes some to significant (shown as “Editor pick”) or deep dive. Editor picks are featured on the home page, appear in the main RSS feed and the sitemap, and can be found by search engines. Auto-summarized papers are listed in the digest, tracks, search and the “all papers” feed, but are marked noindex so search engines skip them.

What we verify

Summaries are drafted by a language model from the paper text; editor picks may be edited by a human, the rest are not. Because a language model can misread a paper, always check the paper itself. Benchmark numbers are machine-extracted first and shown only after the editor has verified them against the paper; anything unverified is never displayed, on any paper.

What is excluded

Papers that do not pass the world-model relevance filter are never listed. The editor can reject a paper that slipped through, which removes it from the site.

Corrections

If you spot an error, email jtscientific2016@gmail.com. Corrections are applied to the paper page.

Licensing and attribution

Thank you to arXiv for use of its open access interoperability. We link to papers rather than reproducing their figures. Metadata comes from arXiv; our summaries and notes are original editorial text.