ReMoT: Reinforcement Learning with Motion Contrast Triplets
This addresses a critical failure point in navigation, robotics, and autonomous driving by providing a systematic solution with significant gains.
The paper tackled the problem of spatio-temporal consistency in vision-language models (VLMs) by introducing ReMoT, a training paradigm that includes a rule-based dataset generation method and a novel optimization technique, resulting in a 25.1% performance improvement on spatio-temporal reasoning tasks.
We present ReMoT, a unified training paradigm to systematically address the fundamental shortcomings of VLMs in spatio-temporal consistency -- a critical failure point in navigation, robotics, and autonomous driving. ReMoT integrates two core components: (1) A rule-based automatic framework that generates ReMoT-16K, a large-scale (16.5K triplets) motion-contrast dataset derived from video meta-annotations, surpassing costly manual or model-based generation. (2) Group Relative Policy Optimization, which we empirically validate yields optimal performance and data efficiency for learning this contrastive reasoning, far exceeding standard Supervised Fine-Tuning. We also construct the first benchmark for fine-grained motion contrast triplets to measure a VLM's discrimination of subtle motion attributes (e.g., opposing directions). The resulting model achieves state-of-the-art performance on our new benchmark and multiple standard VLM benchmarks, culminating in a remarkable 25.1% performance leap on spatio-temporal reasoning tasks.