ARIA: learning when to consult stronger advisor models in agentic systems

A meta-controller trained via Monte Carlo control over compact workflow states recovers 44 to 81% of the executor-advisor gap at under half the advisor cost.

Primary source: arXiv:2610.13870: https://arxiv.org/abs/2610.13870

What’s new: Multi-model agent architectures often pair an inexpensive executor model with an expensive frontier advisor model, but leaving the decision to seek guidance to the executor fails because models struggle to detect their own reasoning errors. Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems (arXiv:2610.13870, 9 Oct 2026) introduces ARIA, an external meta-controller that formulates advisor consultation as a sequential resource-allocation problem. At recurring workflow contact points, ARIA observes a compact state abstraction (combining execution progress, tool feedback status, and historical consultation depth) and selects whether to remain silent or trigger advisor guidance.

ARIA learns action values using tabular Monte Carlo control with sample-average updates driven by terminal task success penalized by a consultation budget parameter. Crucially, neither the executor nor the advisor model undergoes retraining. Evaluated across SWE-bench Verified, Terminal-Bench 2.1, and $\tau^2$-bench, ARIA recovers 44 to 81% of the performance gap between pure-executor and pure-advisor systems while incurring less than half the advisor model’s computational cost. It outperforms ten baseline strategies, including prompt-based self-escalation, input-time query routers, cascade verifiers, and matched-volume random consultations.

Why it matters: Sustaining long-horizon agent workflows with frontier models is economically prohibitive for production tasks, while cheaper models frequently derail on subtle edge cases. Regulating advice through an inspectable, policy-driven meta-controller allows teams to tune an explicit price parameter balancing task completion against API expenditure without altering underlying weights.

Caveats: The policy operates on discrete, predefined contact points within structured agent frameworks (such as Magentic-One), rather than continuous action streams. The state representation abstracts away raw text into coarse execution categories, which may discard subtle domain signals. Performance gains depend on the advisor providing actionable corrections that the executor is capable of parsing and following.