{
  "id": 4475,
  "url": "https://arxiv.org/abs/2605.12571v1",
  "title": "VideoSEAL: Mitigating Evidence Misalignment in Agentic Long Video Understanding by Decoupling Answer Authority",
  "summary": "Long video question answering requires locating sparse, time-scattered visual evidence within highly redundant content. Although current MLLMs perform well on short videos, long videos introduce long-horizon search and verification, which often necessitates multi-turn, agentic interaction. We show that existing LVU agents can exhibit \"evidence misalignment\": they produce correct answers that are not supported by the retrieved or inspected evidence. To characterize this failure, we introduce two ",
  "authors": "Chenhao Qiu, Yechao Zhang, Xin Luo, Shien Song, Xusheng Liu",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-12T10:37:49.000Z",
  "fetched_at": "2026-07-14T16:31:03.577Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4475",
  "original_url": "https://arxiv.org/abs/2605.12571v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}