The paper introduces GGSS (Geodesic-Gated Spherical Steering), a norm-preserving intervention method for reducing demographic bias in generative vision-language models (VLMs) during inference. Unlike existing debiasing approaches designed for static embeddings or CLIP-like models, GGSS discovers a counterfactual bias subspace and applies geodesic-gated spherical steering to mitigate biased outputs while preserving model performance. The method addresses the gap in inference-time debiasing techniques specifically tailored for generative VLMs.