At a Glance
Track real-time task progress to pick the right AI helper at each step: ProgRouter keeps long-term costs under budget while preserving solution quality and meeting deadlines.
ON THIS PAGE
What They Found
Routing decisions that adapt to the evolving state of a multi-step task outperform one-shot or purely reactive strategies. A lightweight progress scorer plus a predictor that estimates marginal progress lets the coordinator choose cheaper or stronger models as needed. Across coding, math, and long-form question benchmarks, ProgRouter satisfied long-term cost limits while maintaining strong task performance and competitive run time. The approach learns from on-the-fly routing trajectories, so it improves as more tasks are served. progress-aware routing
Data Highlights
1Tested on 664 total tasks: 164 HumanEval Plus, 200 MBPP coding, 200 MATH-500 math, and 100 ASQA retrieval-augmented QA tasks.
2Model zoo covered up to 9 models for coding and 7 models for math and QA, with model sizes ranging roughly from 0.5 billion to 35 billion parameters.
3ProgRouter met the imposed long-term operating cost constraint across all four benchmarks while preserving strong task-solving performance and competitive execution time (consistent across the evaluated workloads).
What This Means
Engineers running multi-step AI workflows who want to reduce compute and energy costs without sacrificing results. Technical leaders evaluating agent orchestration strategies can use ProgRouter to trade off budget, deadlines, and solution quality more transparently. Researchers building coordinated agent systems will find the progress-aware routing and dual-path prediction ideas useful for online adaptation.
Need expert guidance?We can help implement this
Key Figures
Fig 1: Figure 1: Overview of LLM agent orchestration in collaborative multi-agent LLM workflows.
Fig 2: Figure 2: Overall procedure of ProgRouter . A coordinator LLM agent manages a collaborative multi-agent LLM workflow by dispatching worker agents and observing real-time workflow states. The online task progress predictor estimates multi-view task progress from structured and semantic state representations, and the progress-guided router selects suitable LLMs from the model zoo by jointly considering predicted progress gain, remaining task difficulty, and cost-budget constraints.
Fig 3: Figure 3: ProgRouter : Performance-cost tradeoff analysis.
Fig 4: Figure 4: LLM routing distribution by model families.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Generalization beyond the tested domains (code, math, retrieval QA) is not yet proven—open-ended web navigation or tool-heavy workflows may need extra validation. The progress scorer needs light domain adaptation: designers must specify observable milestones or coarse outcome regimes for new task types. ProgRouter only routes among existing models and does not change their inherent biases or reliability, so output verification and human oversight remain necessary. human oversight
Methodology & More
Multi-step tasks handled by multiple specialized AI agents create high operating costs because each step may call expensive models and workflows evolve in unpredictable ways. ProgRouter tackles that by scoring how much progress a workflow has already made and predicting how much additional progress each candidate model would provide at the next step. The key components are a multi-view progress scorer (combining coarse outcome regimes with fine-grained progress signals) and a dual-path progress predictor that fuses structured state features and semantic signals with an adaptive gating mechanism to estimate marginal progress gain per model. online routing policy
Using those progress estimates, ProgRouter runs an online routing policy that balances predicted progress gain against the remaining time and cost budget, learning from routing trajectories as it serves tasks. Experimental evaluation across 664 tasks from four benchmarks (coding, math, and retrieval-augmented long-form QA) with model zoos spanning small to large models (≈0.5B–35B parameters) shows ProgRouter can satisfy long-term operating cost constraints while keeping strong task performance and competitive run times compared to one-shot routing, escalation strategies, and naive history-based selection. The method reduces unnecessary calls to expensive models by invoking them only when progress estimates justify the cost, making multi-agent workflows more sustainable and practical—while requiring modest domain-specific progress signals and continued monitoring for model quality and bias. model zoos
Test your agentsValidate against real scenarios
Credibility Assessment:
ArXiv preprint, no institutional affiliations provided and authors have very low h-index values. Lacks strong author/venue credibility signals.