ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs
ProgRouter routes each workflow step by estimated progress gain, aiming to cut agent costs without giving up task quality.
The paper argues that one-shot cascade routing is a poor fit for multi-step LLM agent workflows because the best model choice changes as the task state changes. ProgRouter scores task progress from workflow outcomes, subtask completion, progress trends, and state quality, then predicts which candidate LLM is likely to move the task forward. Its routing decisions weigh that expected gain against time budgets and long-term operating cost. Experiments on HumanEval Plus, MBPP, MATH-500, and ASQA are reported as lowering operating cost versus key baselines while preserving strong performance. ArXiv · AI/CL/LG's note
The paper argues that one-shot cascade routing is a poor fit for multi-step LLM agent workflows because the best model choice changes as the task state changes. ProgRouter scores task progress from workflow outcomes, subtask completion, progress trends, and state quality, then predicts which candidate LLM is likely to move the task forward. Its routing decisions weigh that expected gain against time budgets and long-term operating cost. Experiments on HumanEval Plus, MBPP, MATH-500, and ASQA are reported as lowering operating cost versus key baselines while preserving strong performance. ArXiv · AI/CL/LG's note
score 5