RESEARCH · RESEARCH · #1342
MoFlow paper proposes multi-objective agentic workflow generation
The paper introduces MoFlow, which frames agentic workflow generation as a multi-objective Markov decision process and uses a Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups so each search node stores a set of reachable trade-offs. A single search approximates the Pareto front and lets MoFlow return workflows for different preference weights without retraining; it outperforms six baselines on six benchmarks (mathematics, code, question answering) by average hypervolume under an evaluation setup that favors the baselines.
KEY POINTS
- The paper introduces MoFlow, which frames agentic workflow generation as a multi-objective Markov decision process and uses a Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups so each search node stores a set of reachable trade-offs.
- A single search approximates the Pareto front and lets MoFlow return workflows for different preference weights without retraining; it outperforms six baselines on six benchmarks (mathematics, code, question answering) by average hypervolume under an evaluation setup that favors the baselines.
- MoFlow matters because it enables a single workflow generator to cover a range of trade-offs (accuracy, cost, latency, robustness) and return preference-specific workflows without retraining, which can simplify deployment of agentic systems.
WHY IT MATTERS
MoFlow matters because it enables a single workflow generator to cover a range of trade-offs (accuracy, cost, latency, robustness) and return preference-specific workflows without retraining, which can simplify deployment of agentic systems.