REFLEX Uses Jev as a Typed Decision Layer and Cuts Strong-Model Calls 72.7% at 95% Task Success

arXiv·medium signalXBlueskyLinkedInCopy link

Wu and Lim route an agent's bounded decisions through Jev and call a strong LLM only when confidence is low or free-form generation is needed. On a frozen 100-task benchmark REFLEX reached 95% success with 72.7% fewer strong-model calls than a strong-only agent, across three fallback model families. Reliability fell as action sets grew and when near-valid alternatives sat close to authorization boundaries. On BFCL and τ-style evals it had little edge over a cheap generative cascade. That gives a concrete test for where Jev-style classification helps and where it doesn't.