Metraipolis › archive › dc_8730bd26f8df
EviAnchor: Mitigating Hallucinations in Large Vision-Language Models via Regional Visual Evidence Compensation — arXiv:2608.29092
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.29092
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-17T07:44:03.581Z
- SHA-256 of the archived text
9f5a74243bd89612b413f184c497cf12501eed20f66b9b7c40f5ea82bc7b3d4b- Stable URL of this page
https://metraipolis.com/documento/dc_8730bd26f8df
How to cite
EviAnchor: Mitigating Hallucinations in Large Vision-Language Models via Regional Visual Evidence Compensation — arXiv:2608.29092 — arXiv (en), https://arxiv.org/abs/2608.29092 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8730bd26f8df (Metraipolis archive dc_8730bd26f8df, sha256 9f5a74243bd89612b413f184c497cf12501eed20f66b9b7c40f5ea82bc7b3d4b)
BibTeX
@misc{dc8730bd26f8df,
title = {EviAnchor: Mitigating Hallucinations in Large Vision-Language Models via Regional Visual Evidence Compensation — arXiv:2608.29092},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.29092},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8730bd26f8df; sha256 9f5a74243bd89612b413f184c497cf12501eed20f66b9b7c40f5ea82bc7b3d4b},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.29092. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Large vision-language models (LVLMs) frequently generate content unsupported by visual inputs. Preliminary experiments show that visual evidence is primarily incorporated into answer-side representations in early-to-middle decoder layers, while its direct influence progressively weakens in later layers. This attenuation suggests that visual evidence acquired earlier may be insufficiently utilized during subsequent generation. Based on this observation, we propose EviAnchor, a training-free and single-branch inference framework that preserves and reactivates visual evidence throughout generation. EviAnchor introduces Regional Evidence Anchor (REA) slots to progressively aggregate dense visual tokens into spatially structured representations. It then strengthens the current decision state's access to these visual anchors through decision-conditioned evidence routing, mitigating excessive dependence on textual context. Finally, the model resumes its native Transformer computation to integrate the retrieved visual evidence with question semantics and generation history. Experiments across POPE, CHAIR, and MMHal-Bench demonstrate consistent improvements in visual grounding. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-01 in cs.AI). Categorie arXiv: Artificial Intelligence (cs.AI) ; Computer Vision and Pattern Recognition (cs.CV). Pagina del lavoro: https://arxiv.org/abs/2608.29092 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-17T07:28:26.511Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-01?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).