Researchers have introduced GGSS (Geodesic-Gated Spherical Steering), a novel inference-time intervention designed to mitigate demographic bias in generative vision-language models (VLMs). Unlike previous debiasing methods tailored for static CLIP-style embeddings, GGSS operates on the unit hypersphere to preserve the norm of visual tokens. The method functions by identifying a counterfactual bias subspace and steering visual tokens along geodesic arcs to counteract demographic signals. To maintain model utility, the framework employs an adaptive gate that selectively applies corrections only to tokens exhibiting high demographic signal strength. The authors evaluated GGSS across four generative VLMs against ten inference-time baselines and prompt-based techniques using categorical, pairwise, and occupation-gender bias metrics. Experimental results demonstrate that GGSS achieves the lowest average bias across all tested models and shows significant statistical significance in paired permutation tests for three out of four backbones. Crucially, the intervention maintains high general visual-language performance, with MMStar accuracy remaining within a tight margin of +/- 0.6 percentage points compared to unsteered baselines. This approach provides a way to reduce social bias in generative outputs without requiring retraining or fine-tuning of the underlying vision-language architectures.