Conditional Multi-Event Temporal Grounding in Long-Form Video

A new benchmark called CoMET-Bench aims to close a gap in video understanding by testing how well AI models locate every event in long-form video that matches compositional temporal and spatial conditions, according to a paper posted to arXiv on June 13 [1]. The benchmark comprises 2,789 queries over 600 videos that average 33.8 minutes in length, spanning five real-world domains [1]. Each query is built from four temporal conditions and three spatial conditions, and the dataset includes a dedicated negative-query subset to test whether models can correctly return no results [1]. The authors argue that existing benchmarks localize only a single moment per query, count without temporal conditions, or treat grounding and counting as separate tasks, leaving a gap for applications that require finding every event that satisfies a compound description [1]. The paper introduces a unified evaluation protocol that jointly measures counting, grounding, and negative-query recognition [1]. A new metric called Rejection-F1 is designed to prevent trivial gaming by models that always return an empty set [1]. When the researchers benchmarked a broad suite of multimodal large language models, agent-based methods, and grounding-specialized approaches, they found that existing systems remain far from solving the task [1]. To address the shortcomings, the team built CoMET-Agent, a training-free agentic framework that reformulates the task as structured search-and-aggregate [1]. The framework improved the [email protected] score by 6.1% over GPT-5 using structural reasoning alone, without additional training [1]. A failure analysis surfaced three open directions for future work: fine-grained entity tracking, position-uniform retrieval, and causal event pairing [1]. The paper appeared on arXiv, the open-access repository of electronic preprints that has hosted scientific papers since 1991 and now receives roughly 24,000 submissions per month [7]. The work was shared through arXivLabs, a framework that allows community collaborators to develop and share experimental tools directly on the arXiv website under guidelines that emphasize openness, community, excellence, and user data privacy [5].