NexSight: verifiable long-video tasks show top models score only 54.5%

nex-agi · hf · 2026-11-19

NexSight (Zhejiang Univ., Tongji, Shanghai Qizhi) builds verifiable tasks for agentic long-video understanding. Its SightForge pipeline yields SightLine — 310 evidence-traceable tasks from 108 long videos — where the strongest model scores only 54.52% (79.35% with annotated evidence). Using solver trajectories as step-level supervision, Sight-35B nearly doubles interactive accuracy (21.29%→42.26%) with under 6K examples, beating its family's 397B model.

Original post →

More from Research

Research channel →