VIScore predicts whether a latent world model can plan well — without running the planner

DrMorganLevine · x · 2026-08-17

Problem: MPC-driven planning with latent world models involves a vision encoder, a predictor, a planner, and a success tolerance — yet prior metrics (straightness, loss, physical-state probing, empowerment) only examine one or two modules, leaving them disconnected from actual planning success.

Method: The paper (arXiv:2608.11174, by Haiyu Wu, Randall Balestriero, and Morgan Levine) proposes VIScore (Veracity-Influence-Sobriety), a latent-space metric that quantifies the reachability and capacity of the predictor given the encoded features, plus the hallucination of the search-based planner — covering encoder, predictor, and planner together. VIScore is interpretable, diagnostic, and achieves a strong Spearman correlation with planning success, outperforming prior metrics.

Additional finding: Comparing two regularization losses with the same distribution target, SIGReg benefits self-supervised learning but not planning, while VISReg improves planning success on out-of-domain (OOD) datasets. Code is open-sourced on GitHub.

Original post →

More from Research

Research channel →