[论文] JEPA-Anything: Learning Predictive Models across Different Worlds
研究领域: NLP 作者: Taoyong Cui, Zhongyao Wang, Xinyue Xu, Weiyang Liu, Zhaochen Yu, Yuying Zhang, Qiang Gao, Mengyue Yang, Wanli Ouyang, Pheng Ann Heng, Yingcheng W…
论文概要
研究领域: NLP 作者: Taoyong Cui, Zhongyao Wang, Xinyue Xu, Weiyang Liu, Zhaochen Yu, Yuying Zhang, Qiang Gao, Mengyue Yang, Wanli Ouyang, Pheng Ann Heng, Yingcheng Wu, Zhenfei Yin, Ling Yang 发布时间: 2026-09-17 arXiv: 2609.20800
中文摘要
世界建模使智能能够预见后果、指导干预并从交互中学习。然而预测模型仍是领域特定的:一个共同的学习原则能否支持截然不同系统的世界建模?我们引入 JEPA-Anything——一个基于正交预测分解(OPF)的领域无关框架。作为联合嵌入预测架构(JEPA)的扩展,OPF 将潜在目标分解为互补因子,通过专用路径学习它们,并在共享的预测设计中重新组合。我们在七个领域评估 JEPA-Anything:视觉、生物学、临床轨迹、控制、分子动力学、物理场和天气。实验涵盖表征学习、干预预测、分布外泛化和长周期动力学,包括 10 个匹配的动力学任务、超过 1,000 个临床事件的预测以及跨四个系统的 100 步分子 rollout。与匹配的 JEPA 基线相比,JEPA-Anything 在所有 10 个动力学任务上改进了报告指标,在 Interventional Pong 上将单一干预预测误差降低了 34.8%。在全部四个系统中,它取得了对比方法中最低的单步和 100 步分子误差。超越预测,一个由因子提名的生物学干预在细胞共培养、患者来源类器官、肿瘤碎片和小鼠中获得了实验支持;潜在轨道模式恢复了开普勒标度指数,拟合斜率为 -1.4991。这些结果支持跨异质世界的共同分解预测原则,将世界建模与干预和有实验依据的科学发现联系起来。代码:https://github.com/Gen-Verse/JEPA-Anything
原文摘要
World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically different systems? We introduce JEPA-Anything, a domain-agnostic framework based on orthogonal predictive factorization (OPF). Extending joint-embedding predictive architectures, OPF decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. We evaluate JEPA-Anything across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. Experiments span representation learning, intervention prediction, out-of-distribution gen...
*自动采集于 2026-09-19*
#论文 #arXiv #NLP #小凯