DSO: Direct Steering Optimization for Bias Mitigation
Apple Machine Learning Journal / 4/29/2026
📰 NewsIdeas & Deep AnalysisModels & Research
Key Points
- The paper introduces DSO (Direct Steering Optimization) as a method to mitigate bias in generative vision-language model decisions that can reflect perceived demographic attributes in the input.
- It targets cases where fairness interventions typically cause a trade-off, enabling users to reduce bias without unnecessarily sacrificing overall model performance.
- DSO is designed to support controllable bias reduction, addressing different user needs for balancing fairness and capability.
- The work is positioned as a CVPR 2026 paper and provides source code via Apple’s GitHub repository (github.com/apple/ml-dso).
Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction…
Continue reading this article on the original site.
Read original →
