Momba: Network Modernization Improves Multi-Objective Reinforcement Learning

A new paper on arXiv (2608.07180) explores how modernizing neural network architectures can enhance multi-objective reinforcement learning (MORL). While recent advances in deep RL have shown that improving network architectures can boost sample efficiency and asymptotic performance without changing algorithms, MORL has largely focused on algorithmic innovations, leaving architectures underexplored. MORL aims to discover policies that balance trade-offs among conflicting objectives, and algorithms typically use simple feedforward networks conditioned on the trade-off. The paper questions whether more expressive function approximators could improve performance. The authors integrate recent advances in neural network design, including observation and feature normalization, into MORL. The paper is a cross-announcement, indicating it has been presented or published elsewhere. The work is relevant to the AI and machine learning community, particularly those interested in reinforcement learning and multi-objective optimization.