How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models

Frontier language models are deployed as black-box services, where model weights cannot be modified and customization is limited to prompting. We introduce Advisor Models, a method to train small open-weight models to generate dynamic, per-instance natural language advice that improves the capabilities of black-box frontier models. Advisor Models improve GPT-5.2's performance on RuleArena (Taxes) by 27.4\%, reduce Gemini 3 Pro's steps taken in SWE agent tasks by 24.6\%, and outperform static prompt optimizers in personalizing GPT-5 to user preferences (85-100\% vs. 40-60\%). We also find that advisors are transferable: an advisor trained with a low-cost student model still transfers improvements to a frontier model. Moreover, Advisor Models are robust: we observe no degradation on other benchmarks than the pipeline is trained on. Our method shows how to perform parametric optimization for black-box frontier models in a practical and cost-effective way.

Today’s most powerful AI systems are increasingly available only through APIs. While people can use these models, they usually cannot directly modify or fine-tune them for specialized tasks or personal preferences. We introduce Advisor Models, a new method for improving black-box AI systems without changing their weights. Instead of relying on a single fixed prompt, we train a small “advisor” model to generate custom natural-language guidance for each individual task. A frontier model then uses this advice to produce better answers. Across tasks including tax reasoning, software engineering, and personalization, advisor models significantly improved the performance of frontier AI systems. For example, advisors improved GPT-5.2’s performance on a complex tax benchmark by 27.4% and helped software engineering agents solve tasks nearly 25% more efficiently. Importantly, advisor models can be trained using cheaper AI systems and later transferred to stronger frontier models, making the approach practical and cost-effective. Because the underlying frontier model itself is never modified, our method also avoids many of the risks associated with traditional fine-tuning. Our work suggests a new way to customize powerful AI systems: instead of retraining the model itself, we can teach smaller models to guide them more effectively.