Research map / Vision-language and multimodal / Multimodal LLMs
Actor as Its Own Critic: Unifying Region Understanding and Localization via CycleGRPO
Xin Zhang, Haochen Wang, Yikang Zhou, Zhuochen Wang, Robby T. Tan, Xiangtai Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cycle-consistency reward (CycleGRPO) on top of SAMTok instead of no RL | SAMTok-4B · DLC-Bench | Avg. score: 61.9 -> 67.7 | Improved |
| use cycle reward instead of standard GRPO with IoU reward on generated captions (ablation) | SAMTok-4B · DLC-Bench | Avg. score: 63.4 -> 67.7 | Improved |
| add CycleGRPO instead of standard GRPO (ablation) | SAMTok-4B · GroundingSuite | gIoU All: 62.7 -> 67.6 | Improved |
| add CycleGRPO | SAMTok-4B · GRES | N-acc (avg): 58.7 -> 92.1 | Improved |
| add standard GRPO (no cycle) to SAMTok (ablation) | SAMTok-4B · GRES | gIoU (avg): 71.2 -> 70.8 | Hurt |
| add CycleGRPO | SAMTok-4B · GAR-Bench-VQA | Overall accuracy: 64.2 -> 65.1 | Improved |
| add CycleGRPO | SAMTok-4B · GAR-Bench-VQA (Position and Relation) | accuracy: Position -3.2, Relation -1.0 | Hurt |
| add standard GRPO to SAMTok (ablation) | SAMTok-4B · GAR-Bench-VQA | Overall accuracy: 64.2 -> 63.9 | No clear effect |
| add CycleGRPO | SAMTok-4B · GCG (Val) | CIDEr: 48.2 -> 54.7 | Improved |
| use cycle reward instead of captioning-only GRPO with GT captions (ablation) | SAMTok-4B · GroundingSuite (DAM-triples training) | All: 56.9 -> 68.5 | Improved |
| use cycle reward instead of localization-only GRPO with GT captions (ablation) | SAMTok-4B · DLC-Bench (DAM-triples training) | Avg. score: 61.8 -> 68.0 | Improved |
| increase rollout counts G=K from 2 to 6 (ablation) | CycleGRPO · GroundingSuite | All: 62.5 -> 67.6 | Improved |
Closest papers by meaning
- RegionGPT: Towards Region Understanding Vision Language Model (CVPR 2024 · similarity 0.87)
- Grounding Everything in Tokens for Multimodal Large Language Models (CVPR 2026 · similarity 0.86)
- Breaking the Regional Perception Bottleneck of Multimodal Large Language Models via External Reasoning Framework (CVPR 2026 · similarity 0.86)
- Hugging Visual Prompt and Segmentation Tokens: Consistency Learning for Fine-Grained Visual Understanding in MLLMs (CVPR 2026 · similarity 0.86)
- RegionReasoner: Region-Grounded Multi-Round Visual Reasoning (ICLR 2026 · similarity 0.86)
- VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
