What they did

The authors developed a multi-view progress scorer that combines broad workflow outcome categories with finer signals including subtask completion, progress trends, and the quality of the current workflow state. A dual-path predictor and adaptive meta-gating mechanism then estimate how much progress each candidate LLM would add.

ProgRouter uses those estimates to route each workflow step while considering task-time budgets and longer-term operating costs. It was evaluated on HumanEval Plus, MBPP, MATH-500, and ASQA, covering agentic code generation, mathematical reasoning, and retrieval-augmented long-form answers.

Key findings

• ProgRouter reduced operating costs relative to the evaluated routing baselines while maintaining strong task-solving performance.
• Step-level routing adapted model choice to the workflow's evolving state, unlike one-shot cascade routers that select models only at the query level.
• The approach generalized across four benchmarks and three task families: programming, mathematics, and retrieval-augmented question answering.
• Its routing objective jointly considered predicted progress, available time, and long-term cost efficiency rather than model quality or price alone.

Why it matters

Multi-agent systems can repeatedly invoke expensive models and accumulate large contexts over long workflows. Progress-aware routing offers a way to allocate stronger models only when their expected contribution justifies the added cost, making orchestration responsive to what has already been accomplished.

Caveats

The supplied abstract does not provide numerical cost reductions, accuracy differences, statistical uncertainty, or details about the models and pricing assumptions used. The evidence is also limited to four benchmarks, so robustness under changing model pools, real-world workloads, and inaccurate progress estimates remains an open question.