world2vec：用潜在动力学学习构建机理世界模型
world2vec: Mechanistic world models through latent dynamics learning
一篇有意思的论文：world2vec 不靠奖励就能学出忠实的潜在动力学，参数量比 DreamerV3 少一个量级，还能做系统辨识，做可解释性和科学发现方向的可以看看。
论文 world2vec 提出在无奖励信号的情况下，仅从观测和动作序列中学习机理性的世界模型。其潜在动力学能在长预测时域上忠实复现真实系统，参数量比 DreamerV3、TD-MPC2 和 LeWM 少至少一个数量级。在从像素控制连续环境的任务上取得 SOTA 表现，并在包含 71 个吸引子的基准上完成准确的系统辨识。作者认为这类方法可用于科学发现场景。
world2vec: Mechanistic world models through latent dynamics learning
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.