GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models

Research identifies that generative vision-language models (VLMs) can produce demographically biased outputs, even when visual inputs vary only in controlled attributes like perceived race or gender. Existing debiasing methods are often inadequate for these generative models. A new intervention, Geodesic-Gated Spherical Steering (GGSS), is proposed to address this by steering visual tokens along geodesic arcs within a counterfactual bias subspace, with an adaptive gate to focus corrections on strong demographic signals.

Why it matters

The persistent issue of demographic bias in generative vision-language models poses significant ethical and operational risks for organizations deploying such technologies. Addressing this bias is crucial for maintaining public trust, ensuring equitable application of AI, and complying with future regulatory frameworks concerning AI fairness and accountability. Effective debiasing mechanisms are therefore strategically important for responsible AI development and deployment.

Key insights

• Generative vision-language models (VLMs) frequently produce demographically biased outputs, particularly in human-centered applications.
• Bias can manifest even with minor, controlled variations in perceived demographic attributes (e.g., race, gender) within input images.
• Current inference-time debiasing techniques are largely designed for static embeddings or non-generative models, making them less effective for VLMs.
• The proposed GGSS method offers a norm-preserving intervention to discover and steer visual tokens within a counterfactual bias subspace.
• GGSS utilizes geodesic arcs and an adaptive gate to concentrate bias correction on tokens with significant demographic signals.
• The research evaluates GGSS against four generative VLMs and ten existing inference-time debiasing baselines, suggesting its potential efficacy for improving fairness.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.25375