TimeProVe cuts long video reasoning costs by 93 percent

A new AI pipeline cuts long video reasoning costs by 93 percent. Long video question answering, which has to sift hours of untrimmed footage for sparse, relevant clues, has long battled between dense processing with big vision language models and sparse, caption driven reasoning that misses key moments. TimeProVe turns that compromise on its head with a hybrid design that keeps the expensive VLMs under control while preserving targeted accuracy.

The paper shows TimeProVe starts with lightweight modules to generate action grounded answer and evidence hypotheses. It then calls a powerful vision language model only for targeted verification, dramatically narrowing the amount of heavy computation required. The engine at the core is the Action-based Candidate Evidence, or ACE, module. ACE converts temporally localized actions into query conditioned candidate answers and supporting evidence windows through lightweight reasoning steps, effectively turning a long video into a handful of high quality, context rich prompts for the final verifier. This shift from dense, linearly scanned inference to a structured, event focused search is what unlocks the cost savings.

OpenTSUBench, or OTB, is introduced as an open ended benchmark to evaluate temporally grounded reasoning in real world Activities of Daily Living. The benchmark is designed to stress the ability to locate the right evidence within long, realistic videos rather than just scoring superficial caption matching. The authors frame OTB as a crucible for long video reasoning that mirrors practical adl tasks, where moments of interest are often sparse and motion driven. The combination of ACE and the OTB benchmark positions TimeProVe as a cost aware alternative for industry scale video QA pipelines.

Benchmarks indicate the strength of this design. The team reports TimeProVe outperforms the strongest baseline on OTB by 7.3 percentage points, while reducing VLM calls by 75 percent and overall inference cost by 93 percent. The approach also shows resilience in settings without explicit temporal grounding training, achieving competitive results on Charades-STA. When grounding VLMs are integrated, TimeProVe can reach state of the art on those tasks, underscoring the value of targeted verification as a lever to improve precision without a compute blow up.

For practitioners, the takeaways are concrete. First, lead with the engineering constraint: if you can confine the expensive model to a few high quality windows, you gain huge cost efficiency without sacrificing coverage. TimeProVe demonstrates that a hybrid pipeline can deliver strong results with far fewer VLM calls than dense processing. Second, the quality of the action grounding matters. The ACE module’s ability to translate temporally localized actions into viable evidence windows is the linchpin; mislocalization can prune away critical evidence and degrade answers. Third, evaluation matters: OpenTSUBench is designed to reflect real ADL scenarios, so success on that benchmark is a stronger signal for product viability than niche datasets. Fourth, there is room to push further with grounded VLMs. The paper shows gains when grounding-aware models are used for verification, suggesting a practical path to higher accuracy at large scale.

In short, TimeProVe reframes long video reasoning as a targeted verification problem rather than a brute force video sweep. For teams grappling with cost, latency, and scale, the approach offers a clear playbook: localize intent, narrow evidence windows, and verify only where it counts.