PhAI Labs introduces JEPA-Anything framework for cross-domain world models
Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford, and Princeton released JEPA-Anything, a domain-agnostic framework that applies a single learning recipe across seven fields: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. The method extends joint-embedding predictive architectures (JEPAs) with Orthogonal Predictive Factorization (OPF),…
Key points
- JEPA-Anything applies one shared learning recipe across seven domains, improving stability via orthogonal predictive factorization
- Outperformed matched JEPA baselines on 10 dynamics tasks, including a 34.83% error reduction in Interventional Pong
- Core code is Apache-2.0 licensed; research checkpoints available on Hugging Face, with wet-lab validation for cancer intervention
JEPA-Anything outperformed matched JEPA baselines on 10 dynamics tasks, including a 34.83% reduction in single-intervention error on Interventional Pong. In scientific analysis, it identified IL-18 plus CD73 blockade as a potential cancer intervention, validated in wet-lab tests. The core code is licensed under Apache-2.0, with research checkpoints available on Hugging Face. Results vary by task—planning gains favor Walker2d and HalfCheetah but lag in Hopper—while rollout errors dropped ~44.7% on APEBench Burgers.
Beyond Domain-Specific World Models: JEPA-Anything Uses 1 Recipe for 7 Fields
MarkTechPost · 6 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic model refutes 3SUM and APSP hypotheses with subquadratic algorithms · 1 src
- Llama.cpp adds MTP decoding for Qwen4Exp and GLM-5.3-Flash hybrid model · 2 src
- GitHub releases ReviewBench to evaluate AI code reviews across 219 pull requests · 2 src
- Anthropic study: task understanding beats job title for AI success · 1 src
- Researchers test fixed token codes for language models at 100B-token scale · 1 src
Comments
via GitHub Discussions