ReMoT: Reinforcement Learning with Motion Contrast Triplets

TL;DR

ReMoT introduces a novel contrastive learning framework with a large-scale motion-contrast dataset and a new benchmark, significantly improving spatio-temporal reasoning in vision-language models.

Contribution

The paper presents ReMoT, a unified training paradigm combining a large motion-contrast dataset and a group relative policy optimization method for enhanced spatio-temporal reasoning.

Findings

Achieved state-of-the-art results on new and existing benchmarks.

Constructed the first fine-grained motion contrast triplet benchmark.

Realized a 25.1% performance improvement on spatio-temporal tasks.

Abstract

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.