PROWL: Prioritized Regret-Driven Optimization for World Model Learning
The paper introduces PROWL, a method for enhancing world model learning through prioritized regret-driven optimization. It focuses on improving the robustness of action-conditioned video world models by actively eliciting model failures. The proposed approach demonstrates that selectively generating informative training data can significantly enhance model performance.
- ▪PROWL employs a KL-constrained adversarial curriculum to train policies that expose high-error trajectories.
- ▪The method includes a Prioritized Adversarial Trajectory buffer that focuses on unresolved failure modes.
- ▪Results indicate that PROWL improves robustness over models trained solely on passive data.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.18803 (cs) [Submitted on 11 May 2026] Title:PROWL: Prioritized Regret-Driven Optimization for World Model Learning Authors:Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham, Jonathan Sadeghi, Jeffrey Hawke, Jack Parker-Holder, Ilija Bogunovic View a PDF of the paper titled PROWL: Prioritized Regret-Driven Optimization for World Model Learning, by Ahmet H. G\"uzel and 6 other authors View PDF HTML (experimental) Abstract:Modern action-conditioned video world models achieve strong short-horizon visual realism, yet remain unreliable on rare, interaction-critical transitions that dominate downstream planning and policy performance.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
