GGSS (Geodesic-Gated Spherical Steering) is an inference-time debiasing method for generative vision-language models. GGSS reduces demographic bias in a frozen VLM by installing a lightweight forward hook on the vision-to-language projection layer: it discovers a counterfactual bias subspace on the unit hypersphere and rotates visual token activations along geodesic arcs, with an adaptive per-token gate that concentrates the correction on tokens carrying demographic signal. Because the rotation is norm-preserving and token-selective, it avoids the capability damage that hard subspace projection causes on multimodal large language models (MLLMs).</p>\n","updatedAt":"2026-08-31T09:32:11.134Z","author":{"_id":"640eba51fdeaae139086dfda","avatarUrl":"/avatars/977fb33b34ef944735ae2f2b2c6adb34.svg","fullname":"Duke Sun","name":"dukesun99","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8122286796569824},"editors":["dukesun99"],"editorAvatarUrls":["/avatars/977fb33b34ef944735ae2f2b2c6adb34.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.25375","authors":[{"_id":"6a9549c394b7be1064d0f403","name":"Yiqun Sun","hidden":false},{"_id":"6a9549c394b7be1064d0f404","name":"Junyu Chen","hidden":false},{"_id":"6a9549c394b7be1064d0f405","name":"Pengfei Wei","hidden":false},{"_id":"6a9549c394b7be1064d0f406","name":"Lawrence B. Hsieh","hidden":false}],"publishedAt":"2026-08-26T04:58:40.000Z","submittedOnDailyAt":"2026-08-31T00:00:00.000Z","title":"GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models","submittedOnDailyBy":{"_id":"640eba51fdeaae139086dfda","avatarUrl":"/avatars/977fb33b34ef944735ae2f2b2c6adb34.svg","isPro":false,"fullname":"Duke Sun","user":"dukesun99","type":"user","name":"dukesun99"},"summary":"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.","upvotes":2,"discussionId":"6a9549c394b7be1064d0f407","githubRepo":"https://github.com/dukesun99/GGSS","githubRepoAddedBy":"user","ai_summary":"GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate, preserving visual-language accuracy.","ai_keywords":["Geodesic-Gated Spherical Steering","GGSS","counterfactual bias subspace","unit hypersphere","geodesic arcs","adaptive gate","visual tokens","generative VLMs","inference-time debiasing"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":1},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6945084e835f00df605d738b","avatarUrl":"/avatars/c5b734098c6589098cf287bae16ecd9d.svg","isPro":false,"fullname":"Yiqun Sun","user":"dukesunmtri","type":"user"},{"_id":"640eba51fdeaae139086dfda","avatarUrl":"/avatars/977fb33b34ef944735ae2f2b2c6adb34.svg","isPro":false,"fullname":"Duke Sun","user":"dukesun99","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.25375.md","query":{}}">
GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models
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 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.
Community
GGSS (Geodesic-Gated Spherical Steering) is an inference-time debiasing method for generative vision-language models. GGSS reduces demographic bias in a frozen VLM by installing a lightweight forward hook on the vision-to-language projection layer: it discovers a counterfactual bias subspace on the unit hypersphere and rotates visual token activations along geodesic arcs, with an adaptive per-token gate that concentrates the correction on tokens carrying demographic signal. Because the rotation is norm-preserving and token-selective, it avoids the capability damage that hard subspace projection causes on multimodal large language models (MLLMs).
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.25375 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.25375 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.25375 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.