Generalized CycleGRPO for Fine-Grained Multimodal Understanding and Grounding
Abstract Overview
This paper presents Generalized CycleGRPO, a unified reinforcement learning framework for fine-grained multimodal understanding and grounding across spatial, temporal, and spatio-temporal region spaces. The method establishes a closed region-to-text-to-region loop where a model generates descriptions from an input region and subsequently localizes those descriptions back to the source region, using geometric overlap as a verifiable reward. This enables joint optimization of region captioning and localization using only region annotations, without requiring reference text, external reward models, or LLM judges during reinforcement learning. To assess occurrence-level distinctiveness in video captioning, the authors also introduce DOVE, a benchmark evaluating whether interval captions distinguish a target occurrence from semantically similar recurrences within the same video.
Novelty
The paper generalizes CycleGRPO from 2D image regions to temporal intervals and spatio-temporal trajectories under a unified, geometry-agnostic reinforcement learning formulation. It utilizes the model's own localization capability as a self-evaluating verifiable reward signal for generated descriptions, and introduces DOVE to specifically benchmark occurrence-level distinctiveness in temporal video captioning.
Results
Across spatial benchmarks, Generalized CycleGRPO improves SAMTok's overall gIoU from 57.5% to 68.2% on GroundingSuite and average gIoU from 71.2% to 78.7% on GRES. In temporal tasks, it achieves 51.9% tIoU on Charades and increases DOVE Temporal Evidence Coverage from 14.8% to 33.3% relative to its Qwen3-VL initialization, while also improving tIoU across Charades, ActivityNet, and QVHighlights when initialized from OMTG-4B. On spatio-temporal reasoning with Open-o3-Video-4B, it raises V-STaR mLGM from 47.2% to 49.8% and TVGBench mIoU from 21.1% to 28.5%.
Key Points
- Generalized CycleGRPO trains captioning and localization jointly through a closed region-text-region cycle with overlap-based rewards, bypassing the need for reference captions or external critics during reinforcement learning.
- The framework unifies spatial masks and boxes, temporal intervals, and spatio-temporal evidence under a shared reinforcement learning formulation using domain-specific serializers and overlap measures.
- The paper introduces DOVE, a human-verified benchmark measuring whether interval-conditioned video captions preserve occurrence-specific temporal evidence to differentiate target events from same-query recurrences.