ChronoReplay lets one VLM answer spatial queries in real time

arXiv preprint 2610.34283v1 introduces ChronoReplay, a replay-based training loop that lets a single vision-language model answer categorical, metric and navigation queries on an unbounded video stream. On an A100 it processes a frame in 18 ms, reaching 94.7% side-of-object accuracy, 0.22 m distance error and 0.77 mIoU.

• Frame processing at 512 px takes 18 ms on an A100 at 55 fps
• Side-of-object accuracy 94.7%, distance error 0.22 m, mIoU 0.77
• A 5% replay buffer prevents forgetting across task heads
• Model plus buffer uses 9 GB of GPU memory vs 10 GB baseline