TimeProVe: Propose, then Verify for Efficient Long Video Temporal Reasoning in Activities of Daily Living

A new hybrid framework called TimeProVe cuts the cost of answering questions about long, untrimmed videos by 93 percent while outperforming existing models, according to a preprint submitted to arXiv on 18 June 2026 [1][2]. Long Video Question Answering (LVQA) demands that systems locate sparse, query-relevant evidence inside videos that can stretch for hours. Current methods either run large vision-language models (VLMs) across every frame — an approach the authors describe as incurring “prohibitive computational cost” — or rely on sparse captions that often miss motion-centric details [2]. TimeProVe, introduced in the preprint, takes a different path. It first uses lightweight modules to propose action-grounded answer-and-evidence hypotheses, then calls an expensive VLM only to verify those candidates [1][2]. The central component is the Action-based Candidate Evidence (ACE) module, which converts temporally localized actions into query-conditioned candidate answers and supporting evidence windows through lightweight large language model reasoning [2]. The researchers also built a new benchmark, OpenTSUBench (OTB), to test temporally grounded reasoning in real-world Activities of Daily Living (ADL) scenarios [1][2]. On OTB, TimeProVe outperformed the strongest baseline by 7.3 percent while cutting VLM calls by 75 percent and reducing inference cost by 93 percent [1][2]. Without any explicit temporal-grounding training, the framework posted competitive results on the Charades-STA benchmark and reached state-of-the-art performance when paired with grounding VLMs [1][2]. The paper appeared on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, and related fields [6]. arXiv, which began in August 1991, now receives roughly 24,000 submissions per month and does not subject papers to peer review before posting [6]. The preprint’s abstract page also surfaces experimental tools developed through arXivLabs, a framework launched in 2020 that lets community collaborators build features such as citation explorers and code finders directly on the site [4][5].