Reproduce “A Control Theory of Predictability in Latent World Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanzhe You, Yonggang Zhang, Maohao Ran, Zhiqin Yang, Zhenyuan Zhang, Wei Xue, Jun Song, Xinmei Tian and 1 more
A control theory for latent world models establishes that minimizing prediction error on observed data often fails to optimize control performance due to off-manifold state queries by planners. This work redefines the training objective to minimize the discrepancy between predicted and true plan-costs, quantifying contributions from on-manifold residual and off-manifold extrapolation, with empirical support showing weak correlation between prediction error and control success.
Key insights
- Prediction error averaged over the training data fundamentally decouples from control performance when a planner's actions lead to off-manifold states.
- The critical errors for control suboptimality are one-sided underestimates ("holes") in predicted costs, not symmetrical prediction errors.
- The discrepancy in plan-costs is seeded by "realizable nonlinearity" in the predictor and amplified by off-manifold excursions and the non-normality of the latent dynamics operator.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.