Research map / Reinforcement learning / RL from feedback
Dissecting Post-Training: Uncovering the Complementary Roles of SFT and RL for Document Parsing
Jun-Peng Jiang, An-Yang Ji, Shiyin Lu, Guodong Zheng, Weihong Zhang, Qing-Guo Chen, Weihua Luo, Kaifu Zhang, Long Chen, De-Chuan Zhan, Han-Jia Ye
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 |
|---|---|---|---|
| standard SFT on Markdown (100k synthetic pairs) | Ovis2.5-2B and Qwen2.5-VL-3B · OmniDocBench | Overall | No clear effect |
| standard RL (GRPO, edit-distance reward) | Ovis2.5-2B and Qwen2.5-VL-3B · OmniDocBench | Overall | Improved |
| add 50k general-purpose data to SFT (ablation) | SFT · OmniDocBench | Overall | No clear effect |
| add 10k targeted table/formula data to SFT (ablation) | Markdown SFT · OmniDocBench | TEDS and CDM | No clear effect |
| triple SFT data from 100k to 300k (ablation) | Markdown SFT · OmniDocBench | Overall | No clear effect |
| predict structured JSON instead of Markdown in SFT (ablation) | SFT · OmniDocBench | Overall | Improved |
| add table/formula sub-task data to JSON SFT (ablation) | Ovis2.5-2B · OmniDocBench | Overall: 71.58 -> 82.43 | Improved |
| add table/formula sub-task data to JSON SFT (ablation) | Qwen2.5-VL-3B · OmniDocBench | Overall: 79.41 -> 81.09 | Improved |
| use decomposed element-wise reward instead of global reward (ablation) | RL, Ovis2.5-2B · OmniDocBench | Overall: 66.59 -> 71.62 | Improved |
| use decomposed reward instead of global reward (ablation) | RL, Ovis2.5-2B · OmniDocBench | Table TEDS: 71.3 -> 60.97 | Hurt |
| use decomposed reward instead of global reward (ablation) | RL, Qwen2.5-VL-3B · OmniDocBench | Overall: 84.11 -> 84.29 | No clear effect |
| add table/formula data to RL with decomposed reward (ablation) | RL, Ovis2.5-2B · OmniDocBench | Overall: 71.62 -> 79.00 | Improved |
Closest papers by meaning
- Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (CVPR 2026 · similarity 0.88)
- Reading or Reasoning? Format Decoupled Reinforcement Learning for Document OCR (CVPR 2026 · similarity 0.87)
- MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing (NeurIPS 2026 · similarity 0.86)
- MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios (NeurIPS 2026 · similarity 0.85)
- Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (ICLR 2026 · similarity 0.85)
- Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead (ICLR 2026 · similarity 0.84)
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.
