{"paper":{"title":"When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Vision-language models hallucinate because they over-align visual embeddings to text, and removing a linguistic bias subspace fixes it.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Dianbo Liu, Harshvardhan Saini, Samyak Jha, Yiming Tang","submitted_at":"2026-05-07T10:09:18Z","abstract_excerpt":"Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grai"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace, so that explicitly projecting this subspace from visual representations removes the bias without discarding task-critical visual information.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Decoder-based VLMs hallucinate because visual embeddings are over-aligned to a text manifold; projecting out the top principal components of a universal linguistic subspace reduces this bias and improves benchmark performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Vision-language models hallucinate because they over-align visual embeddings to text, and removing a linguistic bias subspace fixes it.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f9c88b5b806f759216d3786e2a091569f7924c8d22eb035ff0b8f41d9379ce2b"},"source":{"id":"2605.08245","kind":"arxiv","version":3},"verdict":{"id":"e93a481b-41ca-4ba1-882c-196df3dda3b8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:01:42.168645Z","strongest_claim":"We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence.","one_line_summary":"Decoder-based VLMs hallucinate because visual embeddings are over-aligned to a text manifold; projecting out the top principal components of a universal linguistic subspace reduces this bias and improves benchmark performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace, so that explicitly projecting this subspace from visual representations removes the bias without discarding task-critical visual information.","pith_extraction_headline":"Vision-language models hallucinate because they over-align visual embeddings to text, and removing a linguistic bias subspace fixes it."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.08245/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T19:31:19.305479Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T13:06:17.537298Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"9c0a38d34e46c4e2b6b3736b7aa9f3551be0c448b3582020c5ac9c3288f88ef8"},"references":{"count":23,"sample":[{"doi":"","year":null,"title":"Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C","work_id":"fe2a10cc-b8d6-4236-b413-9ae3cdc9470f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Neeraj Anand, Samyak Jha, Udbhav Bamba, and Rahul Rahaman","work_id":"131b814a-7192-4c6c-893a-b6c325135269","ref_index":2,"cited_arxiv_id":"1505.00468","is_internal_anchor":true},{"doi":"","year":null,"title":"Qwen3-VL Technical Report","work_id":"1fe243aa-e3c0-4da6-b391-4cbcfc88d5c0","ref_index":3,"cited_arxiv_id":"2511.21631","is_internal_anchor":true},{"doi":"","year":2023,"title":"David M Chan, Suzanne Petryk, Joseph E Gonzalez, Trevor Darrell, and John Canny","work_id":"7733671b-ca00-46d6-ba19-73ef7f876327","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","work_id":"51960d72-c69f-4db8-8efd-e90e8b4d9524","ref_index":5,"cited_arxiv_id":"2309.08600","is_internal_anchor":true}],"resolved_work":23,"snapshot_sha256":"1129fe4c1831a55e4958aeeb2e9e24c60b5f1b4d8556dff32f89f963f76f14fd","internal_anchors":9},"formal_canon":{"evidence_count":2,"snapshot_sha256":"85181579ed5b55d53a85445141e278cf394e6cae5cb6f9c17f8612369bf0549f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}