AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

SupGRPO: Enhancing GRPO with Matching-based Online SFT for Text Spotting

Xudong Xie, Yuzhe Li, Jing Shi, Zhifei Zhang, Curtis Wigington, Zhaowen Wang

ECCV 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add ATS data to training (ablation)SupGRPO 3B · ATS detectiondetection score: 65.3 -> 71.6Improved
add ATS data to training (ablation)SupGRPO 3B · ATS recognitionrecognition score: 75.7 -> 84.4Improved
add GRPO joint training to SFT (SupGRPO vs SFT) (ablation)Qwen2.5-VL-3B · ATS (Qwen2.5-VL-3B) detectiondetection score: 63.3 -> 71.6Improved
add SFT joint training to GRPO (SupGRPO vs GRPO) (ablation)Qwen2.5-VL-3B · ATS (Qwen2.5-VL-3B) recognitionrecognition score: 80.0 -> 84.4Improved
use GRPO instead of SFT (ablation)Qwen2.5-VL-3B · ATS (Qwen2.5-VL-3B) detectiondetection score: 63.3 -> 62.0Hurt
add data augmentation to SFT (ablation)SFT · ATS (Qwen2.5-VL-3B)detection score: 63.3 -> 65.1Improved
apply online SFT to all tokens on top of GRPO (ablation)GRPO · ATS (Qwen2.5-VL-3B) detectiondetection score: 62.0 -> 64.0Improved
apply online SFT to text tokens only on top of GRPO (ablation)GRPO · ATS (Qwen2.5-VL-3B) detectiondetection score: 62.0 -> 60.8Hurt
apply online SFT to location tokens with matching on top of GRPO (ablation)GRPO · ATS (Qwen2.5-VL-3B) detectiondetection score: 62.0 -> 71.6Improved
use IoU and text matching instead of IoU-only matching (ablation)SupGRPO location SFT · ATS (Qwen2.5-VL-3B) detectiondetection score: 66.5 -> 71.6Improved
use IoU and text matching instead of text-only matching (ablation)SupGRPO location SFT · ATS (Qwen2.5-VL-3B) detectiondetection score: 68.6 -> 71.6Improved
add text reward to baseline model with GRPO (ablation)Qwen2.5-VL-3B · ATS (Qwen2.5-VL-3B) detectiondetection score: 26.3 -> 61.1Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy