[论文] world2vec: Mechanistic world models through latent dynamics learning
研究领域: ML 作者: Tobias Schmidt, Rodrigo Gonzalez Laiz, Krisztián Szuppinger 发布时间: 2025-10-28 arXiv: 2510.25912
论文概要
研究领域: ML 作者: Tobias Schmidt, Rodrigo Gonzalez Laiz, Krisztián Szuppinger 发布时间: 2025-10-28 arXiv: 2510.25912
中文摘要
世界模型可以在不忠实捕捉环境动力学的情况下实现强大的控制性能,这限制了它们在机械可解释性和科学发现方面的潜力。在此,我们引入 world2vec,直接从观察和动作中学习机械世界模型,无需奖励信息。World2vec 学习潜在动力学,能够在长预测范围内忠实地再现真实系统,参数量至少比 DreamerV3、TD-MPC2 和 LeWM 少一个数量级。这使得在基于像素的连续环境中实现最先进的控制性能,并在包含 71 个吸引子的基准测试中实现准确的系统辨识,展示了世界模型如何适用于科学发现。
原文摘要
World models can achieve strong control performance without faithfully capturing the dynamics of their environment, limiting their potential for mechanistic interpretability and scientific discovery. Here, we introduce world2vec to learn mechanistic world models without reward information directly from observations and actions. World2vec learns latent dynamics that faithfully reproduce the true system over long prediction horizons, with at least one order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM. This enables state of the art performance in controlling continuous environments from pixels and accurate system identification across a benchmark of 71 attractors, demonstrating how world models are amenable to scientific discovery.
*自动采集于 2026-11-01*
#论文 #arXiv #ML #小凯