ClosurePairs: New Protocol Identifies Stochastic World Model Ambiguities
A recent publication on arXiv presents ClosurePairs, a novel approach for assessing stochastic models in practical environments. The research indicates that conventional transition data cannot distinguish between state aliasing and process noise in future conditional distributions, even with advanced probabilistic models. ClosurePairs utilizes compatible microstates and introduces external disturbances, employing a two-way variance decomposition to dissect state aliasing, process noise, and their intricate relationships. Furthermore, the study proposes an independent-repeat method for cases where disturbances can't be reused. The protocol, tested across 18 nonlinear Langevin scenarios, reportedly reduces alias-fraction error by nearly 16 times while maintaining test NLL.
Key facts
- Paper arXiv:2608.00591 introduces ClosurePairs protocol.
- ClosurePairs is an interventional evaluation protocol for stochastic world models.
- The paper proves non-identifiability of state aliasing vs. process noise from transition data.
- ClosurePairs uses compatible microstates and repeated exogenous disturbances.
- Two-way variance decomposition identifies state aliasing, process noise, and interaction.
- Independent-repeat variant handles non-reusable disturbances.
- Paired supervision reduces alias-fraction error 15.96-fold at identical test NLL.
- Tested on 18 nonlinear Langevin conditions.
Entities
Institutions
- arXiv