DELTAVID gives video AI surgical precision in spotting local changes across clips

Current video multimodal large language models excel at open-ended understanding but fail at fine-grained spatiotemporal perception—especially when two videos share nearly identical global semantics and differ only in a brief moment or a small region. To address this, a team of researchers led by Yankai Yang introduces DELTAVID, a verifiable proxy-task framework that transforms cross-video spot-the-difference into a trainable perception signal. The model learns to pinpoint local changes, determine exact temporal boundaries, and organize spatial evidence by comparing similar videos. To make this scalable, they built DELTAVID-10K, a dataset of 10,000 real-video samples with controllable local differences and evidence labels, and DELTAVID-Bench for standardized evaluation.

Experimental results show that DELTAVID substantially outperforms existing models on cross-video difference understanding, and these gains transfer to general video understanding benchmarks, including MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench. The findings demonstrate that cross-video differences serve not only as a diagnostic tool for perception failures but also as effective proxy supervision, pushing video MLLMs from coarse semantic matching toward rigorous, evidence-based spatiotemporal reasoning—a critical capability for applications like surveillance, video editing, and automated quality control.