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

Parth Asawa, Alan Zhu, Abigail O’Neill, Matei Zaharia, Alex Dimakis, Joseph E. Gonzalez
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:4068-4086, 2026.

Abstract

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-asawa26a, title = {How to Train Your Advisor: Steering Black-Box {LLM}s with Advisor Models}, author = {Asawa, Parth and Zhu, Alan and O'Neill, Abigail and Zaharia, Matei and Dimakis, Alex and Gonzalez, Joseph E.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {4068--4086}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/asawa26a/asawa26a.pdf}, url = {https://proceedings.mlr.press/v306/asawa26a.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models %A Parth Asawa %A Alan Zhu %A Abigail O’Neill %A Matei Zaharia %A Alex Dimakis %A Joseph E. Gonzalez %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-asawa26a %I PMLR %P 4068--4086 %U https://proceedings.mlr.press/v306/asawa26a.html %V 306 %X 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.
APA
Asawa, P., Zhu, A., O’Neill, A., Zaharia, M., Dimakis, A. & Gonzalez, J.E.. (2026). How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:4068-4086 Available from https://proceedings.mlr.press/v306/asawa26a.html.

Related Material