MoFlow proposes agent workflows for multiple objectives

An arXiv paper submitted on 29 Sep 2026 presents MoFlow, a proposed approach to generating agentic workflows with several objectives in view. The objectives named in the abstract are accuracy, cost, latency, robustness and consistency. The authors describe a limitation in existing workflow-generation methods: they often optimize accuracy alone or combine objectives into a weighted sum. In their account, each trained generator is then committed to one trade-off, and changing preferences can require retraining from scratch.

MoFlow frames the task as a multi-objective Markov decision process. Its search uses Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups. The key difference described in the abstract is how the method represents possible outcomes during search: each node keeps a collection of reachable trade-offs instead of reducing them to one weighted score. The authors say this lets a single search approximately cover the Pareto front, from which the system can select a workflow for a requested preference by lookup without retraining.

The paper evaluates MoFlow against six baselines across six benchmarks in mathematics, code and question answering. The authors note that the comparison is not straightforward because the baselines are designed to optimize a single scalar objective. To compare them across preferences, the evaluation reruns the baselines for each testing preference, which MoFlow does not see. The abstract characterizes this arrangement as favoring the baselines, yet reports that MoFlow achieves the highest average hypervolume under it.

For AI builders, the proposal is a way to generate workflows across multiple objective preferences rather than committing the generator to one weighted objective. In the paper’s account, the resulting choices can be retrieved by preference after a single search, without retraining the generator. That distinction is central to the method: it seeks to represent a range of trade-offs and make them available for selection, rather than return only one score-optimized workflow. The reported evaluation covers the stated six benchmarks and compares MoFlow with six baselines under the described preference-based setup.