New ExecTune method optimizes how guide models can effectively steer black-box LLMs while reducing costly inference expenses.

Researchers propose Guide-Core Policies (GCoP), a system where a guide model generates structured strategies executed by expensive black-box core models to amortize inference costs

The framework unifies base, supervised, and advisor-style approaches, differing mainly in how the guide model is trained

Analysis reveals that guide-averaged executability—the probability a generated strategy can be faithfully executed by the core model—is the key factor governing end-to-end performance

Current GCoP implementations often fail to optimize for executability under real deployment constraints, leading to brittle strategies and computational inefficiency