PROWL: Prioritized Regret-Driven Optimization for World Model Learning

Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham, Jonathan Sadeghi, Jeffrey Hawke, Jack Parker-Holder, Ilija Bogunovic

arXiv:2605.18803 · 2026-05-20 공개 · arXiv · PDF

diffusion-models world-models reward-hacking out-of-distribution adversarial-training mine-rl prioritized-learning regret-driven-optimization

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. Because passive demonstration data systematically under-samples these high-impact regimes, improving robustness requires actively eliciting model failures rather than relying on their natural occurrence. We introduce a KL-constrained adversarial curriculum in which a policy is trained to expose high-error trajectories of a diffusion-based world model while remaining close to the behavior distribution. The world model is continuously fine-tuned on these adversarially discovered trajectories, yielding an adversarial training loop that converts rare failures into a stable, near-distribution training signal without drifting into out-of-distribution exploitation. To maintain pressure on unresolved weaknesses as the model improves, we propose a Prioritized Adversarial Trajectory (PAT) buffer that re-ranks trajectories based on prediction error, action fidelity, and learning progress, focusing training on unresolved failure modes rather than repeatedly revisiting solved cases. We implement our approach in the MineRL framework and evaluate it on held-out out-of-distribution trajectories; PROWL improves robustness over models trained on passive data alone, reveals reward-hacking behaviors under weak behavioral constraints, and demonstrates that effective adversarial world-model training critically depends on balancing exploratory failure discovery with explicit behavioral regularization. Our results suggest that scalable world models benefit not only from larger datasets, but also from selectively generating informative training data.

한국어 요약

📋 한 줄 요약

**[월드 모델 / 적대적 커리큘럼]** KL 제약 적대 정책으로 희귀 실패를 능동적으로 유도해 비디오 월드 모델의 견고성을 끌어올리는 PROWL 제안.

🎯 핵심 기여도

💡 핵심 아이디어

월드 모델의 견고성은 데이터를 더 모으는 문제가 아니라 "어떤 실패를 능동적으로 찾아낼 것인가"의 문제이며, 적대 정책은 행동 분포 근처에 묶여야 OOD 익스플로잇 없이 의미 있는 학습 신호가 된다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: 확장 가능한 월드 모델 학습이 데이터 규모뿐 아니라 정보가 풍부한 학습 데이터의 선택적 생성에서도 이득을 본다는 메시지를 제시. **한계**: MineRL 도메인 위주 평가, 더 복잡한 실세계 환경(자율주행·로보틱스)으로의 확장은 후속 과제.

🚀 실용적 활용