MoFlow paper proposes one search for many workflow trade-offs

A research paper posted to arXiv on 29 September 2026 describes MoFlow, a method for generating agentic workflows that balance several goals at once. The authors, Yining Lu, Aurelie Lozano, Xi Yang, Naoki Abe, Yu Deng and Meng Jiang, name accuracy, cost, latency, robustness and consistency as the objectives.

The authors say existing workflow-generation methods typically optimize accuracy alone or a weighted sum of objectives. In their account, each trained generator is therefore tied to one fixed trade-off and must be retrained from scratch when preferences change.

The paper proposes a different framing. MoFlow treats workflow generation as a multi-objective Markov decision process and solves it with Convex-Hull Monte Carlo Tree Search using optimistic set-valued backups. Each node in the search stores a set of reachable trade-offs instead of a single weighted score.

According to the authors, that lets one search approximately cover the Pareto front, the set of options where improving one objective would cost another. A workflow for a given preference can then be returned by lookup, without retraining.

On evaluation, the authors say they tested MoFlow against six baselines on six benchmarks covering mathematics, code and question answering. Because the baselines are built to optimize a single scalar objective, they acknowledge that a like-for-like comparison is hard. They instead reran the baselines for each testing preference, which MoFlow never sees, a setup they describe as favoring the baselines. They report that MoFlow still achieves the highest average hypervolume, a measure of how well a set of trade-offs covers the objectives.

What is confirmed and what is not: everything above comes from the paper's abstract page, so these are the authors' own descriptions and results. The abstract gives no numerical results, and we have not reviewed the full paper, its code or its benchmark details. We found no independent coverage or replication, so the performance claim should be treated as unverified. This article rests on one source.

Why it matters: teams that build agent systems often have to trade answer quality against cost and speed, and those priorities can change. If the approach holds up, a single search could produce a menu of workflows to choose from, rather than a new training run for every preference. That remains a claim for other researchers to test.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.