TRACE: Temporal Retrieval with Anchored and Convergent Evidence for Long-Horizon Video Understanding
VES-Bench, a 600-question benchmark of Temporal Ordering and Event Counting items over 348 public long videos, is introduced and TRACE, a training-free agent that grounds answers in raw visual clips, builds an evidence bundle round by round, and stops only when the answer stabilises as the bundle grows and a final pass...