Research map / Vision-language and multimodal / Multimodal LLMs
Syn-GRPO: Self-Evolving Data Synthesis for MLLM Perception Reasoning
Qihan Huang, Haofei Zhang, Rong Wei, Yi Wang, Rui Tang, Mingli Song, Jie Song
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 online image synthesis with diversity reward | GRPO · LISA-Grounding (3B) | accuracy: 62.24 -> 68.28 | Improved |
| add online image synthesis with diversity reward | GRPO · LISA-Grounding (7B) | accuracy: 65.26 -> 70.14 | Improved |
| add diversity reward to Syn-GRPO (ablation) | Syn-GRPO · LISA-Grounding (3B) | accuracy: 64.60 -> 68.28 | Improved |
| add diversity reward to Syn-GRPO (ablation) | Syn-GRPO · RefCOCOg (3B) | accuracy: 85.85 -> 87.45 | Improved |
| add entropy loss to GRPO | GRPO · RefCOCO+ (3B) | accuracy: 84.10 -> 82.55 | Hurt |
| add entropy loss / entropy advantage to GRPO | GRPO · LISA-Grounding (3B) | accuracy: 62.24 -> 64.23 | Improved |
| replace offline GPT-4o descriptions with online self-evolving synthesis | GRPO + Offline Generation · LISA-Grounding (3B) | accuracy: 64.23 -> 68.28 | Improved |
| train with diversity smoothing (gamma 0.7 vs 0.9) (ablation) | Syn-GRPO · LISA-Grounding (7B) | accuracy: 69.24 -> 70.14 | Improved |
| use asynchronous instead of synchronous data server (ablation) | Syn-GRPO · REC training time (3B) | hours: 13.16 -> 6.45 | Improved |
| use Syn-GRPO instead of GRPO | GRPO · OVD COCO_filter | mAP: 18.66 -> 23.74 | Improved |
| use Syn-GRPO instead of GRPO | GRPO · ISR perception | mAP: 33.46 -> 35.77 | Improved |
| increase training data size 400 to 2000 (ablation) | Syn-GRPO · REC LISA-Grounding | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VLM-R1 — VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation (NeurIPS 2025 · similarity 0.89)
- Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.89)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.89)
- Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · similarity 0.88)
- Perception-R1: Pioneering Perception Policy with Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.88)
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.
