Researchers propose MoFlow to generate agentic workflows optimized across multiple objectives

Researchers from IBM Research and the University of Notre Dame propose MoFlow, a method to generate agentic workflows that jointly optimize multiple objectives such as accuracy, cost, latency, robustness, and consistency. Existing workflow generators typically focus on optimizing accuracy alone or a weighted sum of objectives, which commits each trained generator to a fixed trade-off and necessitates retraining when preferences change.

MoFlow formulates workflow generation as a multi-objective Markov decision process and solves it using Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups. In this approach, every node stores a set of reachable trade-offs rather than a single weighted score, allowing a single search to approximately cover the Pareto front.

The authors evaluate MoFlow against six strong baselines across six benchmarks spanning mathematics, code, and question answering. Because the baselines are single-scalar optimizers by design, the authors adopt an evaluation setup that favors the baselines: they rerun the baselines for each testing preference, which MoFlow does not require. Even under this stringent setup, MoFlow achieves the highest average hypervolume, a metric used to measure the quality of a set of non-dominated solutions in multi-objective optimization.

The method enables retrieval of a workflow for any preference by lookup from the approximated Pareto front without retraining, addressing a key limitation of existing approaches that require retraining when preferences shift.