AI News

MoFlow addresses the problem of generating agentic workflows that must balance accuracy, cost, latency, robustness, and consistency. Existing workflow-generation methods generally optimize accuracy alone or use a weighted sum, meaning each generator is tied to one trade-off and must be retrained when preferences change. The paper formulates workflow generation as a multi-objective Markov decision process. It solves this problem with Convex-Hull Monte Carlo Tree Search and optimistic set-valued backups, allowing each search node to retain reachable trade-offs instead of a single weighted score. One search approximately covers the Pareto front, so a workflow for a new preference can be selected by lookup without retraining. MoFlow is evaluated against six strong baselines on six benchmarks spanning mathematics, code, and question answering. Because the baselines are single-scalar optimizers, the evaluation reruns each baseline for every testing preference, even though MoFlow does not see those preferences. Under this setup, which the authors describe as favoring the baselines, MoFlow achieves the highest average hypervolume.