DSReg Makes JEPA World Model Latents Interpretable and Recoverable
Researchers from UIUC, Cold Spring Harbor Laboratory, Brown University, and MPI propose DSReg (Dependency-Sparsity Regularization), a method that can be applied to trained JEPA representations to align each dimension with a real world latent variable, such as sun, moon, temperature, or wind, without reconstruction, decoders, or labels. The team provides a theoretical framework yielding the first fully identifiable JEPA, and DSReg can reuse existing LeJEPA checkpoints with performance comparable to joint training.
Coverage timeline
机器之心机器之心
JEPA 能不能真正理解观测背后的生成机制？ 传统的 JEPA 只能保证学到的隐表征不坍缩，但却完全没有办法对应到真实生成观测数据的那些隐变量。 来自 UIUC、冷泉港实验室、布朗大学、马普所的研究者给出了肯定的回答。他们提出的 DSReg（Dependency-Sparsity Regularization）可以直接加在已经训练好的 JEPA 上，让学到的每一维都对应一个真实的世界变量，比如场景里的太阳、月亮、温度和风。整个过程不需要重建观测，不需要解码器，也不需要任何标签。 研究团队同时为 DSReg 建立了完整的理论框架，由此得到首个完全可识别的 JEPA。DSReg 可以应用在预训练表征上，直接复用现有的 LeJEPA checkpoint，达到和联合训练相似的效果。 论文标题：DSReg: Provably Recovering Individual World Latents without Reconstruction 论文链接：https://arxiv.org/abs/2610.09457 项目主页：https://dsreg.github.io/ 代码：https://github.com/kunwuz/dsreg 预测得准，不等于理解得对 JEPA 不去重建像素，而是在表征空间里预测未来。它因此省掉了昂贵的解码器，也成了构建世界模型的热门路线。但少了重建，也就少了一个把表征和真实世界对齐的 “锚”。 在一定假设下，LeJEPA 学到的表征和真实世界状态之间只差一个未知的线性变换。这个变换让模型完全无法理解真实的生成机制，因为表征的每一维往往是好几个未知物理因素的混合。 这些未知的混合有时候对预测并没有影响，但却完全阻碍了模型对环境的理解。比如机器人要把杯子往左推，同时让旁边的刀保持不动。如果某个学到的变量同时混着杯子和刀的位置，调这个变量就不再只是移动杯子，而是也可能把刀一起带动，给用户带来危险。很多时候我们不仅需要让机器看上去做得对，同时也要保证它们的认知没有偏差，不然永远会有隐藏的风险。 DSReg：给预训练表征加一个旋转 DSReg 的思路很直接。LeJEPA 学到的表征经过白化后，和真实变量之间只差一个旋转，而旋转不改变 JEPA 的训练损失。所以只要找到正确的那个旋转，就能把变量拆开，同时不损失任何预测能力。 但是如何找到正确的旋转？
Related stories
Fudan, Tencent Hunyuan, Tsinghua Release ExplorationBench to Test AI Rule Discovery
Fudan University, Tencent Hunyuan, and Tsinghua University released ExplorationBench, a benchmark measuring how AI systems discover rules in unfamiliar environments. It includes two altered 'alien worlds'—the programming language AlienCode and the logic proof system AlienLogic—where models must learn rules through trial and error. Results showed that without environment feedback, scores remained low (max 11.0%), while autonomous exploration led to high scores (up to 89.0%).
2 reports · 1 sources
Tadpole: Autoencoder Foundation Model for 3D PDEs with Online Learning
Researchers propose Tadpole, an autoencoder-based foundation model for 3D partial differential equations (PDEs), which learns to compress physical states before predicting evolution, with online learning capabilities. The approach differs from typical predictive models by first learning the spatial structure of physical fields, aiming to improve transferability across different physical systems.
WorldMemArena benchmark evaluates multimodal agent memory via action-world interaction
Researchers from UCSB, Stanford, ETH Zurich, Johns Hopkins, CMU, and J.P. Morgan Chase released WorldMemArena, a multimodal agent memory benchmark accepted at NeurIPS 2026. It evaluates memory through an observation-action-environment feedback loop, breaking long-term memory into write, maintain, retrieve, and use stages to pinpoint where agents fail. The benchmark includes multimodal cues, memory lifecycle tracking, and world changes after actions, addressing limitations of static recall or final QA evaluations.
UCSB and Microsoft Propose ScholarEvolve for AI Agents to Read Papers and Self-Modify Code
Researchers from UC Santa Barbara and Microsoft introduced ScholarEvolve, a method that automates the research process of reading literature, implementing algorithms, and evaluating results to enable AI agents to evolve their harness code without modifying underlying model parameters. The approach addresses the limitation of prior coding-agent-based harness optimization, which relied on the agent's existing knowledge and observed failures, by allowing agents to draw on the latest academic findings. The paper and code are publicly available.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe