Update README.md

README.md

@@ -15,32 +15,13 @@
 
 We are delighted to release Infinity-Parser2-2B, our latest state-of-the-art document understanding model. Compared to our prior model, Infinity-Parser-7B, we have deeply optimized our data engine and multi-task reinforcement learning. We have successfully condensed robust multi-modal parsing capabilities into a highly efficient 2B-parameter model, offering massive speedups and brand-new zero-shot capabilities for real-world business scenarios.
 
-## 🌟 Key Features
-
-### 📚 Upgraded Data Engine
-- **Massive & Diverse Data:** Added **1M+ full-text samples** across 9 document types (academic, financial, books, etc.).
-- **Targeted Enrichment:** Injected 170K synthetic financial tables, 900K formulas, and 5K negative samples to mitigate hallucinations.
-- **Adaptive Sampling:** Dynamically adjusts data distribution based on task importance and dataset size for balanced learning.
-
-### 🧠 Multi-Task Reinforcement Learning
-- **Verifiable Reward System:** Designed a novel reward mechanism to support Joint Reinforcement Learning (RL).
-- **Unified Optimization:** Simultaneously co-optimizes multiple tasks, ranging from full-text and table parsing to layout analysis and Document VQA.
-
-### 📈 Breakthrough Parsing Performance
-Despite its compact 2B size, it significantly outperforms our previous 7B model:
-- **Domain SOTA:** Achieves SOTA on financial benchmarks (`FinDocBench`, `FinTabBench`), surpassing frontier models like DeepSeek-OCR2, GLM-OCR, and PaddleOCR-VL-v1.5.
-- **Public Benchmarks:** Achieves SOTA on `olmOCR-Bench` and `PubTabNet`, with highly competitive results on `OmniDocBench-v1.5` and `UniMERNet`.
-- **General Multimodal:** Scores **66.06** on average across 7 benchmarks (e.g., MathVista, MMMU), beating the Qwen3-VL-2B base (+3.2pt).
-
-### 🚀 Massive Inference Acceleration (3.68x Faster)
-- **Optimized Architecture:** Transitioned to the highly efficient **Qwen3-VL-2B** base model.
-- **Blazing Fast:** Inference throughput surged by **3.68x** (from 441 to **1,624 tokens/sec**), slashing latency and deployment costs without accuracy drop.
-
-### ✨ Expanded Capabilities (Zero-to-One Additions)
-Unlocked entirely new skills in this release:
-- **Chart Parsing:** Scores 79.91 on `Chart2Table`.
-- **Chemical Structures:** Scores 68.05 on our new `ChemDraw-198` and 52.78 on `CoSyn-Chemical`.
-- **Layout Analysis:** Achieves 64.92 on `DocLayNet` and 73.16 on `OmniDocBench-v1.5-layout`, matching dedicated layout models like DocLayout-YOLO.
+## Key Features
+
+- **Upgraded Data Engine**: We comprehensively upgraded our data engine by adding over 1 million diverse full-text samples, 170K synthetic financial tables, 900K formulas, and targeted negative samples to mitigate hallucinations. Combined with a dynamic adaptive sampling strategy, this ensures highly balanced and robust multi-task learning across various document types.
+- **Multi-Task Reinforcement Learning**: We designed a novel verifiable reward system to support Joint Reinforcement Learning (RL), enabling the model to seamlessly and simultaneously co-optimize multiple complex tasks, including full-text parsing, table and formula extraction, layout analysis, and document VQA.
+- **Breakthrough Parsing Performance**: Despite its compact 2B size, it significantly outperforms our previous 7B model. It achieves State-of-the-Art (SOTA) results on both in-house financial benchmarks (`FinDocBench`, `FinTabBench`)—surpassing frontier models like DeepSeek-OCR2 and GLM-OCR—and public sets like `olmOCR-Bench` and `PubTabNet`, while maintaining highly competitive general multimodal capabilities.
+- **Massive Inference Acceleration (3.68x Faster)**: By transitioning to the highly efficient Qwen3-VL-2B architecture, our inference throughput has surged by **3.68x** (jumping from 441 to 1,624 tokens/sec), dramatically slashing deployment latency and costs without compromising core parsing accuracy.
+- **Expanded Capabilities**: We have unlocked entirely new zero-shot skills in this release, achieving strong benchmark results in chart parsing (`Chart2Table`), chemical structure recognition (including our new `ChemDraw-198`), and layout analysis, where it successfully matches the performance of dedicated specialized models like DocLayout-YOLO.
 
 # Architecture