{
  "id": 425,
  "url": "https://arxiv.org/abs/2606.31054v1",
  "title": "ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs",
  "summary": "Multimodal Large Language Models (MLLMs) are critically hampered by hallucination, generating content inconsistent with the provided image. In this paper, we identify an internal signature of hallucination: progressive degradation of text-to-image cross-attention during generation, leading to specific failure patterns like unfocused or biased attention. Existing mitigation strategies are largely outcome-driven and do not explicitly target this failure mode. To address this problem, we propose AD",
  "authors": "Zhiyuan Yao, Zheren Fu, Zhixiao Zheng, Jiajun Li, Yi Tu, Zhendong Mao",
  "category": "research",
  "topics": "bias-fairness,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-06-30T02:46:10.000Z",
  "fetched_at": "2026-07-14T14:14:32.646Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/425",
  "original_url": "https://arxiv.org/abs/2606.31054v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}