MoFlow
An arXiv paper proposing a multi-objective agentic workflow generation method that formulates workflow generation as a multi-objective Markov decision process, employing convex-hull Monte Carlo tree search to approximate the entire Pareto front within a single search.
Timeline (last 90 days)
MoFlow is a research paper published on arXiv that proposes a multi-objective agent workflow generation method
MoFlow models workflow generation as a multi-objective Markov decision process
MoFlow employs convex hull Monte Carlo tree search with optimistic set-valued backups to solve the problem
Each node in MoFlow's search tree stores a set of reachable trade-off solutions rather than a single weighted score
MoFlow's set-valued representation allows a single search to approximately cover the entire Pareto front
When faced with any new preference, MoFlow requires only a single lookup to return the corresponding optimal workflow, without any retraining
The research team compared MoFlow against six strong baseline methods on six benchmarks covering three types of tasks: mathematics, code, and question answering
The experiment adopted a baseline-favorable evaluation setup: baseline methods were re-run for each test preference, while MoFlow had never seen these test preferences
Even under the stringent conditions that favor baselines, MoFlow still achieved the highest average hypervolume
The computational overhead of convex hull MCTS as the number of objectives increases and the scalability of set-valued backups still need further verification in larger-scale scenarios
All Facts (10)
MoFlow is a research paper published on arXiv that proposes a multi-objective agent workflow generation method
50%UnverifiedMoFlow models workflow generation as a multi-objective Markov decision process
50%UnverifiedMoFlow employs convex hull Monte Carlo tree search with optimistic set-valued backups to solve the problem
50%UnverifiedEach node in MoFlow's search tree stores a set of reachable trade-off solutions rather than a single weighted score
50%UnverifiedMoFlow's set-valued representation allows a single search to approximately cover the entire Pareto front
50%UnverifiedWhen faced with any new preference, MoFlow requires only a single lookup to return the corresponding optimal workflow, without any retraining
50%UnverifiedThe research team compared MoFlow against six strong baseline methods on six benchmarks covering three types of tasks: mathematics, code, and question answering
50%UnverifiedThe experiment adopted a baseline-favorable evaluation setup: baseline methods were re-run for each test preference, while MoFlow had never seen these test preferences
50%UnverifiedEven under the stringent conditions that favor baselines, MoFlow still achieved the highest average hypervolume
50%UnverifiedThe computational overhead of convex hull MCTS as the number of objectives increases and the scalability of set-valued backups still need further verification in larger-scale scenarios
50%