arxivcs.CVcs.AI2026-06-26
Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding
Shuimu Chen, Yuteng Chen, Yuanshen Guan, Zebang Cheng, Zeyu Zhang, Shengqian Qin, et al.
Current multimodal reflection mechanisms for long video understanding predominantly rely on closed-loop self-reflection within internal parameters. Lacking objective external evidence, models are frequently trapped in blind confidence and often fail to correct errors. Furthermore…