Paper of the Week — MoFlow: Multi-Objective Agentic Workflow Generation
Practitioners building multi-objective agentic workflows can now route at inference time across accuracy, cost, latency, and robustness without retraining separate generators.
MoFlow: Multi-Objective Agentic Workflow Generation
Yining Lu, Aurelie Lozano, Xi Yang, Naoki Abe, Yu Deng, Meng Jiang. Published 2026-09-29. arXiv:2609.38294
One sentence summary
MoFlow generates agentic workflows that are steerable across multiple competing objectives at inference time, without retraining a new generator for each tradeoff.
Why this paper
Most production agentic systems make a silent design choice: optimize for accuracy and ignore cost and latency until they become operational fires. MoFlow gives teams a principled way to expose and navigate that tradeoff from day one.
What they did
Today’s workflow generators commit to a single objective (or a hand-tuned weighted sum) at training time, which means changing the cost/accuracy tradeoff requires a whole new model. MoFlow instead trains one generator conditioned on a preference vector — a direction in objective space — so you can steer at inference time by adjusting weights across accuracy, cost, latency, robustness, and consistency. The approach borrows ideas from multi-objective optimization and uses a flow-based generation backbone to produce diverse Pareto-optimal workflow candidates.
Key findings
- MoFlow consistently produces workflows on or near the Pareto frontier across all evaluated objective pairs, outperforming single-objective baselines that collapse to one corner of the tradeoff space
- Adjusting the preference vector at inference time shifts the generated workflow’s operating point predictably — accuracy vs. cost tradeoffs are steerable without retraining
- Robustness and consistency objectives, often ignored in prior work, are tractably jointly optimized alongside accuracy and latency
- A single trained MoFlow generator matches or exceeds task-specific generators on their own target objective while also delivering multiple other Pareto points
- The framework generalizes across workflow types, including sequential chains, branching DAGs, and tool-augmented pipelines
Why it matters for practitioners
If you’re building agentic systems with GPT-6 Astra or Claude Fable 5.1 on the hot path and cheaper models (GPT-4.1 Nano, Gemini 3.8 Flash, DeepSeek V4-Flash) for cheaper steps, you’re already implicitly solving a multi-objective workflow problem — just without tooling. MoFlow gives you a framework to formalize that routing logic and optimize it end-to-end rather than hand-tuning model assignments per task type. This is especially relevant as Sonnet 5’s ~40% real-cost premium (due to its new tokenizer emitting more tokens) makes cost-accuracy tradeoffs harder to ignore.
What you can use today
- The preference-vector conditioning pattern is implementable on top of existing workflow frameworks today: encode your current cost/accuracy tradeoff as a vector and pass it as a conditioning signal when generating or selecting workflow templates
- When evaluating agentic pipelines, track all five objectives (accuracy, cost, latency, robustness, consistency) from the start — MoFlow’s framing makes clear that optimizing one while blind to others produces dominated solutions
- Watch the authors’ repo for released code; the paper describes a training recipe compatible with standard instruction-tuning toolchains, meaning adaptation to your task domain should not require significant infrastructure changes