Research map / Large language models / Reasoning and chain of thought

Reading or Reasoning? Format Decoupled Reinforcement Learning for Document OCR

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FD-RL, an SFT-then-RL pipeline for document OCR that filters training samples by output entropy and uses format-decoupled rewards (edit distance for text, BLEU/expression correctness for formulas, TEDS for tables), reaching 90.41 on OmniDocBench.

Problem

End-to-end document OCR models show high output entropy on formatted content (formulas, tables) and token-level SFT struggles with format-sensitive documents where multiple valid formats exist.

Approach

Use per-sample average token entropy from the SFT model to select high-entropy format-intensive samples for RL, and replace a single token-level reward with format-decoupled rewards specialized per content type (text, formula, table).

Key results

• FD-RL achieves 90.41 overall on OmniDocBench, surpassing dots.ocr (88.41) among end-to-end models.
• RL on top-50% entropy-filtered data raises overall from 87.13 (SFT only) to 90.41; ablation shows 0% filtration gives 88.47.
• Formula CDM 88.67 and TEDS 87.35 are best among end-to-end VLMs in the table.

Surprising findings

• RL applied directly to the base Qwen3-VL-4B without SFT yields only 49.37 overall, far below SFT alone.

Limitations

• Formula reward uses BLEU after LaTeX conversion, a proxy that may not capture semantic equivalence fully.
• Entropy filtration threshold and rate tuned on the benchmark ablation (50% optimal), with performance sensitive to rate.

Experimental setup

Datasets: PDFA DocStruct4M DocGenome IAM ORAND-CAR HME OmniDocBench

Benchmarks: OmniDocBench

Metrics: Overall Text Edit Distance Formula CDM Table TEDS Table TEDS-S Reading Order Edit Distance

Baselines: dots.ocr Deepseek-OCR PaddleOCR-VL MinerU2.5 POINTS-Reader olmOCR-7B Nanonets-OCR-s PP-StructureV3 Gemini-2.5 Pro Qwen3-VL-4B Logics-Parsing olmOCR 2

Builds on: GRPO Qwen3-VL EasyR1 LLaMA Factory

Closest papers by meaning

• DocR1: Evidence Page-Guided GRPO for Multi-Page Document Understanding (AAAI 2026 · similarity 0.85)
• VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
• Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.84)
• Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.84)
• Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning (AAAI 2026 · similarity 0.83)
• Table2LaTeX-RL: High-Fidelity LaTeX Code Generation from Table Images via Reinforced Multimodal Language Models (NeurIPS 2025 · similarity 0.83)