Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
Supergroup Relative Policy Optimization (SGRPO) is a post-training reinforcement learning method that improves biomolecular generation by directly rewarding set-level diversity. It works by comparing sets of candidates generated under the same condition rather than individual candidates. This pushes the utility-diversity Pareto frontier outward, allowing models to achieve better utility or diversity trade-offs across small molecule and protein design tasks while maintaining broad generation coverage.