Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems
4.20T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.13870.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryARIA is a meta-controller for agentic systems that decides when a lower-cost executor should consult a stronger advisor. It frames consultation as sequential resource allocation via Monte Carlo control on cost-penalized outcomes. On SWE-bench Verified, Terminal-Bench 2.1, and τ²-bench, it recovers 44-81% of the success gap at less than half advisor cost, outperforming ten baselines.
Why it mattersConcrete cost-aware escalation policy for multi-model agent stacks, with a single tunable parameter and benchmarked gains over routing and cascading baselines.
Cited by
No citations on record.