VIScore: A New Metric for Diagnosing Planning Quality in Latent World Models

DrMorganLevine · x · 2026-08-13

Paper: VIScore: Diagnosing Planning-Relevant Quality in Latent World Models

Problem: The connection between latent space properties and successful planning remains unclear. The study compares SIGReg and VISReg regularization functions, finding that SIGReg benefits SSL but not planning, whereas VISReg improves out-of-domain (OOD) planning success.

Method: The authors propose VIScore, a metric quantifying the reachability and capacity of a predictor given encoded features, and the hallucination of searching-based planners.

Results: Covering the encoder, predictor, and planner, VIScore explains planning success rates better than existing metrics like straightness and physical-state probing, consistently achieving a Spearman correlation over 0.75.

Original post →

More from Research

Research channel →