JEPA-Anything：用一套配方为 7 个领域构建世界模型
Beyond Domain-Specific World Models: JEPA-Anything Uses 1 Recipe for 7 Fields - MarkTechPost
PhAI Labs、CUHK、复旦、斯坦福、牛津与普林斯顿的研究者发布 JEPA-Anything，用一套共享学习配方替代逐领域设计预测模型，通过正交预测分解（OPF）把宽度 d 的 JEPA 目标拆成 K 个正交子空间并各配一个预测器。它在 10 个动力学任务上全部优于匹配的 JEPA 基线，Interventional Pong 单干预误差下降 34.83%，APEBench Burgers 6 步 rollout 误差降约 44.7%。核心代码以 Apache-2.0 开源，研究检查点已放在 Hugging Face。
正文
当前显示原文，中文全文译文尚未就绪。
Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford and Princeton have released JEPA-Anything, a domain-agnostic framework for building world models. Instead of designing a new predictive model for each field, it applies one shared learning recipe to very different systems. It extends joint-embedding predictive architectures (JEPAs) with a method called Orthogonal Predictive Factorization (OPF). The research team tested it across 7 domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather.
What problem does JEPA-Anything solve?
A standard JEPA, such as I-JEPA or V-JEPA 2, uses a context encoder, an EMA target encoder and one predictor. The predictor outputs one monolithic target embedding. The research team call this a capacity-allocation problem: high-variance structure dominates, and weaker modes get conflicting gradients.
How does Orthogonal Predictive Factorization work?
OPF splits the latent target of width d into K learned subspaces of width r, with d = K × r. Most experiments use K = 4. Each factor gets a dedicated predictor. The factor predictions are then recombined through the Moore-Penrose pseudoinverse of the projector matrix. The result is 1 complete latent state for decoding, planning or rollout.
Three regularizers keep the factors useful:
Orthogonality loss: keeps columns within each projector orthonormal and different projectors in non-overlapping subspaces; Factor-activity loss: a hinge on per-coordinate standard deviation so no factor goes dead; Encoder-variance loss: sends a direct anti-collapse signal to the online encoder
The OPF loss is simply added to each domain’s original training loss. Domain adapters handle tokenization and encoders; the core library exposes the shared core as OrthogonalFactorProjection.
Orthogonality matters for stable synthesis. On CITRIS Interventional Pong, a capacity-matched unconstrained multi-head model had a condition number of 438.52. The orthogonal version reached 1.00005, with cross-factor overlap near zero.
What results does the research report?
Group I, terminal readout: On single-cell data, zero-shot PBMC clustering (AvgBIO) rose to 0.7752 versus 0.7194 for Cell-JEPA. Norman perturbation Pearson rose from 0.787 to 0.814. For forecasting over 1,000 clinical events on UK Biobank data, mean PRAUC was 0.718 versus 0.711 for the matched standard JEPA.
Group II, latent world dynamics: On Interventional Pong, single-intervention MSE fell 34.83%. Unseen combined interventions improved 12.90%, and 6-step free rollout improved 8.58%. JEPA-Anything improved reported metrics on all 10 matched dynamics tasks. Benchmarks include CausalWorld, DeepMind Control, PDEBench and WeatherBench2. On APEBench Burgers, 6-step rollout error dropped about 44.7%, improving in every seed. For 100-step molecular rollouts with a TrajCast-style backbone, it posted the lowest MAE and RMSD on water, quartz, paracetamol and benzene.
Planning results are mixed. With parameters matched within 0.3%, JEPA-Anything improved CEM return on Walker2d and HalfCheetah. Hopper favored standard JEPA.
Group III, scientific analysis: Factor analysis nominated IL-18 plus CD73 blockade as a cancer intervention. Wet-lab tests supported it in co-cultures, patient-derived organoids, tumor fragments and mice. Latent orbital modes also recovered Kepler’s law with a fitted slope of −1.4991 against the theoretical −1.5.
How does JEPA-Anything compare with other world models?
Sources: each project’s GitHub README, linked in the header row. JEPA-Anything checkpoint status from its Hugging Face card. Verified October 5, 2026.
Key Takeaways
OPF splits 1 JEPA target into K orthogonal factors with dedicated predictors; It beat matched JEPA baselines on all 10 dynamics tasks; Interventional Pong single-intervention error fell 34.83%; Planning gains are environment-dependent; Hopper favored standard JEPA; Core code is Apache-2.0; research checkpoints sit on Hugging Face
Asif Razzaq
Asif RazzaqMeet Together Link: A Free CLI That Runs Open Models Like Kimi K3 and GLM 5.3 Inside Claude Code, Codex, and OpenCode; Asif RazzaqReflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads; Asif RazzaqYandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade; Asif RazzaqThe Story of Qwen: Alibaba’s AI Models From 7B to 2.4T