How many exposures does a world model need? WARC learns unseen game mechanics in ~80 exposures
third-intelligence · hf · 2026-11-10
Researchers from Third Intelligence and the University of Tokyo ask a fundamental question: how many exposures does a world model need to learn unseen environment mechanics? They study this on ARC-AGI-3, releasing MechEval, which labels 69 recurring mechanics across 20 games and scores predictions per exposure.
- They build WARC, a 35M-parameter vision transformer world model pretrained on other games that updates all parameters after each transition (test-time training).
- On held-out games, WARC predicts 4 out of 5 mechanics with ≥80% accuracy on affected cells; half reach that within 80 exposures in a single playthrough—only 1 in 7 does without updates.
- In closed-loop play, WARC-Agent (a fixed graph explorer + WARC) triples the explorer's official efficiency score from 0.15 to 0.46, mostly via test-time training; their Kaggle entry scores 0.20 on hidden games.
More from AGI Musings
- 1985 Whole Earth Review cover: "Computers as Poison" mirrors today's AI debates — beffjezos · 2026-11-10
- A decade of Cambridge students now run FAR AI, METR, UK AISI and GovAI — S_OhEigeartaigh · 2026-11-10
- Comic argues AI scientists will write the next chapter of science, not humans — YogeshMalik · 2026-11-10
- "Alignment" is semantically overloaded — taboo the word to argue clearly — Sauers_ · 2026-11-10
- Creator vows to quit AI-generated content: 'AI really feels like AI, not human' — Abhishekcur · 2026-11-10
- doty clarifies the scaling debate: the point is against hardcoding human priors, not abandoning algorithms — dotey · 2026-11-10