Back to AI Research

AI Research

EvoRouter learns how to allocate models and agents under a compute budget

Key Takeaways

  • A routing framework combines execution-derived insights, on-policy distillation and cost-aware reinforcement learning.
  • [EvoRouter](https://arxiv.org/abs/2610.35690) addresses the configuration problem in multi-agent systems: choosing models, collaboration styles, agent counts, interaction rounds and skills together.
  • Those choices multiply quickly, while a final success or failure signal gives little detail about which planning decision helped.
  • ## Turn execution feedback into routing guidance
  • The framework uses an expert component to synthesise performance and cost feedback across configuration trajectories.

EvoRouter addresses the configuration problem in multi-agent systems: choosing models, collaboration styles, agent counts, interaction rounds and skills together. Those choices multiply quickly, while a final success or failure signal gives little detail about which planning decision helped.

Turn execution feedback into routing guidance

The framework uses an expert component to synthesise performance and cost feedback across configuration trajectories. It turns that experience into conditional insights about capability needs, scaling resources and interactions among choices.
A privileged teacher combines these insights with current execution feedback. Through on-policy distillation, it supplies token-level supervision for the router’s own planning trajectories. Cost-aware reinforcement learning then optimises task performance together with computational cost.
The updated router produces new execution trajectories, providing material for later insight extraction. The proposed improvement loop therefore depends on actual execution feedback rather than only a fixed set of hand-written routing rules.

Results retain their comparison boundary

The authors evaluate EvoRouter on 19 in-domain datasets covering five domains and two out-of-domain benchmarks. They report average in-domain accuracy of 68.69%, a 7.23-percentage-point improvement over the strongest routing baseline, alongside a 50.95% reduction in inference compute cost.
Those figures describe the reported study and its comparison. Inference compute cost should not automatically be interpreted as a bill reduction for a different provider or as the total cost of training and operating the system.

Measure the full configuration

For teams evaluating routing, the useful question is which combination achieves an acceptable result with the resources it consumes. That requires measuring the agent count, interaction pattern and model choices together.
The paper provides a method and reported results for that comparison. Applying it to a new workload would still require representative tasks, explicit cost accounting and an evaluation of how the learned routing behaviour transfers beyond its training distribution.

Comments