VideoSEAL: Mitigating Evidence Misalignment in Agentic Long Video Understanding by Decoupling Answer Authority

VideoSEAL addresses evidence misalignment in long video understanding agents, where MLLMs produce correct answers unsupported by retrieved evidence; introduces temporal groundedness and semantic grounding diagnostics to characterize this failure mode in agentic long video QA systems.