A new benchmark checks whether AI video agents actually gather evidence before answering a question, not just whether they happen to guess right.
The benchmark, called AgentVidBench, is a multi-hop video question answering test that requires spatial, temporal, and causal reasoning across multiple steps, unlike older benchmarks built around single-step scene queries or simple summaries. It also ships step-by-step solution traces, so evaluators can score the reasoning path an agent takes and not just its final answer. Researchers tested 12 proprietary and open-source multimodal large language models (MLLMs) and found single-turn performance was weak across the board. Wiring the same models into state-of-the-art agentic workflows - where a model can take multiple steps, pull evidence, and revise - generally improved both accuracy and the quality of the reasoning trail, and the team also published a simple baseline agentic strategy that already holds its own on the benchmark.
This matters because most video-AI benchmarks measure the wrong thing: whether a model can summarize a clip or name an object, not whether it can chain evidence together to solve a multi-step problem, the kind of reasoning needed for tasks like reviewing surveillance footage, breaking down sports plays, or auditing a robot's actions from video logs. Trajectory scoring closes a real gap, since a model can land on the correct answer for the wrong reasons, and that distinction has been mostly invisible on prior leaderboards.
The accuracy gains from bolting on agentic scaffolding are real, but trajectory scores still trail accuracy scores - a reminder that getting the right answer and being able to show your work remain two different achievements.