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

Hugging Face Daily Papers Papers

Summary

GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate during inference, preserving visual-language accuracy.

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS---Geodesic-Gated Spherical Steering---a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.
Original Article
View Cached Full Text

Cached at: 09/01/26, 12:03 PM

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

Source: https://huggingface.co/papers/2608.25375

Abstract

GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate, preserving visual-language accuracy.

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather thangenerative VLMs. We proposeGGSS---Geodesic-Gated Spherical Steering---a norm-preserving intervention that discovers acounterfactual bias subspaceon theunit hypersphere, steersvisual tokensalonggeodesic arcs, and uses anadaptive gateto focus correction on tokens that carry stronger demographic signal. We evaluate fourgenerative VLMsagainst ten adaptedinference-time debiasingbaselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability.GGSSachieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.

View arXiv pageView PDFGitHub1Add to collection

Get this paper in your agent:

hf papers read 2608\.25375

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.25375 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.25375 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.25375 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

GAPS: Dimension-Level Gates for Conditional Activation Steering

arXiv cs.CL

GAPS introduces dimension-level gating for conditional activation steering in language models, combining static and dynamic gates to selectively intervene and improve behavior-capability trade-off, with significant gains in toxicity mitigation and concept removal tasks.