VPFlow, a new preprint on arXiv, addresses offline goal-conditioned reinforcement learning, where an agent learns to reach arbitrary goals from arbitrary starting states without relying on dense reward signals. The abstract notes that model-based methods offer a path to solving complex tasks by scaling computation at test time.
The proposed approach, value-predictive planning, is aimed at settings where the world model is imperfect. The title indicates that the method integrates value prediction into the planning process, though the abstract does not provide implementation details.
The work is positioned as a way to make offline goal reaching more robust when the available world model is inaccurate. As with many preprints, the claims are not yet peer-reviewed.