CPT: Consistent Proxy Tuning for Black-Box Models
He, Yuanyang Huang, Zitong Khan, Salman Zuo, Wangmeng Shao, Ling Feng, Chun-Mei
He, Yuanyang
Huang, Zitong
Khan, Salman
Zuo, Wangmeng
Shao, Ling
Feng, Chun-Mei
Supervisor
Department
Computer Vision
Embargo End Date
Type
Conference proceeding
Date
License
Language
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Tuning proprietary black-box models, whose parameters are inaccessible, is a growing need. Proxy-tuning addresses this by applying the output logit difference from a tuned "proxy" model to the black-box model at test time. However, this creates a train-test inconsistency, which can limit performance. We propose Consistent Proxy Tuning (CPT), a simple yet effective black-box tuning method that resolves this issue. By leveraging the frozen large black-box model and an additional frozen small white-box model during training, CPT ensures the optimization objective is consistent with the test-time proxy mechanism. This simple alignment significantly boosts performance. Our method operates solely on logits, making it model-agnostic and applicable to any classification task. Extensive results confirm CPT’s superior performance for tuning LLMs and VLMs across diverse datasets.
Citation
Y. He, Z. Huang, S. Khan, W. Zuo, L. Shao, C.-M. Feng, "CPT: Consistent Proxy Tuning for Black-Box Models," 2026, pp. 18007-18011.
Source
ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Conference
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Keywords
46 Information and Computing Sciences, 4611 Machine Learning
Subjects
Source
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Publisher
IEEE