When world models should update states to reduce loss
When Should a World Model Move? Loss-Conditioned State Execution
Artificial IntelligenceMachine Learning
Summary
World models predict future states to help machines make decisions, but not every predicted update improves outcomes. The authors introduce a way to decide when a model should update its current state or keep it, based on whether doing so will reduce the overall error or loss. Their method checks if proposed changes reliably lower loss across groups before applying them. Tests show this approach improves accuracy compared to always changing or always keeping the state unchanged.
What this means in practice
- •For forecasting teams: Improve when to update forecast states to reduce prediction errors across time series data like financial or sales data.
- •For inventory management teams: Decide state updates for supply and inventory systems more selectively to better balance predictive accuracy with update frequency.
Authors
Jintao Xu, Zhengyu Chen, Ben Zhang, Yongzhi Qi, Jianshen Zhang
Abstract
We introduce loss-conditioned state execution, a model-agnostic method that decides whether to execute a world model's fixed feasible proposal or retain the current state. Predictive informativeness alone, however, does not establish whether an update will reduce downstream loss. Occurrence ranking can approach perfection while persistence remains the unique absolute-loss Bayes action. Two transition laws can also share occurrence information and conditional variance yet require opposite absolute-loss decisions. We formalize state movability as the existence of a loss-reducing feasible correction and distinguish it from the benefit of a particular proposal. Our method constructs a loss-specific feasible proposal from a predictive distribution and evaluates its groupwise bounded-loss gain over persistence on independent calibration units. The proposal is executed only in groups with a positive simultaneous lower confidence bound. For fixed proposals and groups with bounded unit losses, we prove that every accepted group has lower expected loss than persistence with high probability when calibration units are i.i.d. draws from the target population. Experiments on public forecasting and action-conditioned dynamics benchmarks show supported updates and a trade-off between certification and coverage. On 28,684 held-out M4 Monthly series, the method executes the proposal for 14.0% of series and achieves bounded loss 0.588, compared with 0.599 for persistence and 0.621 for always executing the proposal. The paired 95% bootstrap intervals for both comparisons lie below zero. In constrained forecasting of six unhealthy-inventory types from JD$\mbox{.}$com, a leading e-retailer in China, strong occurrence-ranking signal coexists with a loss-based preference for persistence, illustrating why event predictability and state execution must be evaluated separately.