Reproduce “NexSight: Verifiable Tasks for Fine-Grained Agentic Long-Video Understanding” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tianci Xie, Li Chen, Pinxue Guo, Rui Zheng

NexSight combines SightForge, which verifies tasks by reacquiring video evidence, with SightLoop, which records agents’ observations. In a matched-budget test on SightLine, agentic window selection raised temporal evidence-hit recall from 19.84% to 40.63%. Accuracy rose from 36.13% to 40.97%, but the difference was not statistically significant, showing improved localization without establishing an answer-accuracy gain.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.