Completion Aware Guidance Improves World Action Model Success Rates
October 2, 2026
Completion Aware Guidance (CAG) is a training-free sampling method designed to prevent task-incomplete imagination in World Action Models. CAG improved success rates from 64% to 70% on RoboTwin 2.0 and reduced task-incomplete prediction errors from 79% to 40% in simulations.
HOW THIS AFFECTS YOU
●
builderYou can implement this sampling method to improve robot control without retraining the underlying model.
●
researcherThis provides a path to mitigate the tendency of world models to favor plausible but incomplete local continuations.