[Submitted on 28 Feb 2026 (v1), last revised 10 Jun 2026 (this version, v3)]

ReMoT: A Reinforcement Learning Framework Using Motion Contrast Triplets to Improve VLM Spatio-Temporal Reasoning