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

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 diagnostics, temporal groundedness and semantic groundedness, and use them to reveal two pressures that amplify misalignment: prompt pressure from shared-context saturation at inference time and reward pressure from outcome-only optimization during training. These findings point to a structural root cause: the coupled agent paradigm conflates long-horizon planning with answer authority. We therefore propose the decoupled planner–inspector framework, which separates planning from answer authority and gates final answering on pixel-level verification. Across four long-video benchmarks, our framework improves both answer accuracy and evidence alignment, achieving 55.1% on LVBench and 62.0% on LongVideoBench while producing interpretable search trajectories. Moreover, the decoupled architecture scales consistently with increased search budgets and supports plug-and-play upgrades of the MLLM backbone without retraining the planner. Code and models are available at https://github.com/Echochef/VideoSEAL.

AI video agents are increasingly used to answer questions about long videos. Ideally, such agents should find the relevant visual evidence first and then answer based on what they actually saw. However, we find that they can sometimes give correct answers without having found or checked the right evidence. In other words, the system appears to succeed according to the final answer, but it does not follow the intended process of evidence-based reasoning. This work frames this problem as a form of reward hacking in long-video question answering. When an agent is mainly rewarded for producing the correct final answer, it may learn to exploit shortcuts, guesses, or prior knowledge instead of carefully verifying the video. This creates a mismatch between what the system is optimized to do and what users actually want: answers that are not only correct, but also supported by trustworthy evidence. To address this, we propose VideoSEAL. The key idea is to separate the agent into two roles. One part searches through the long video to find possible evidence, while another part checks the actual visual content and has the authority to give the final answer. This prevents the search component from making unsupported guesses. Our experiments show that this design improves both answer accuracy and the reliability of the evidence behind the answer, making long-video AI agents more transparent and trustworthy.