MoFlow studies the generation of agentic workflows that simultaneously optimize multiple objectives such as accuracy, cost, latency, robustness, and consistency. Existing workflow generators usually target a single objective or a weighted sum, which forces a complete retraining whenever the preference changes. MoFlow formulates workflow generation as a multi‑objective Markov decision process (MOMDP) and solves it with Convex‑Hull Monte Carlo Tree Search (MCTS) equipped with optimistic set‑valued backups. Each node in the search tree stores a set of reachable trade‑offs instead of a single scalar score, allowing a single search to approximate the Pareto front. Given any preference, a workflow can be retrieved by lookup in this set without any further training.
We evaluate MoFlow on six benchmarks covering mathematics, code, and question answering, against six strong baselines. Because the baselines are single‑scalar optimizers, we adopt an evaluation protocol that favors them: they are rerun for every test preference, while MoFlow never sees those preferences. Even under this stringent setup, MoFlow achieves the highest average hypervolume, demonstrating its superiority in multi‑objective optimization.
Review