AIM: Model Modulation with Logits Redistribution
Zihan Wang, Zhongkui Ma, Xinguo Feng, Zhiyang Mei, Zhiyong Ma, Derui Wang, Jason Xue, Guangdong Bai
Retraining-free model behaviour modulation through logits redistribution, including utility and focus modulation.
The problem
Serving different stakeholder needs by maintaining separately trained models can be costly. AIM studies changes to model behaviour without retraining or access to the original training data.
The method
AIM redistributes logits to support utility modulation and focus modulation. The former adjusts output utility; the latter shifts attention toward selected input features. These are behaviour-control mechanisms, not simply an authorisation switch.
Evidence and scope
The implementation covers image classification, semantic segmentation and text generation. The relevant evidence is how modulation changes these tasks under the evaluated settings, not a claim that any downstream policy can be enforced through logits alone.
See the original code and paper above. CoreLocker and AdaLoc address complementary model-access questions.
Citation
@inproceedings{wang2025ai,
title={{AI} Model Modulation with Logits Redistribution},
author={Zihan Wang and Zhongkui Ma and Xinguo Feng and Zhiyang Mei and Zhiyong Ma and Derui Wang and Jason Xue and Guangdong Bai},
booktitle={THE WEB CONFERENCE 2025},
year={2025},
url={https://openreview.net/forum?id=lOSomJvrc5}
}