Improve Fine-R1 model card with paper link, authors, and structured content

Hi! I'm Niels from the Hugging Face community team.

I've updated the model card for Fine-R1 to make it more informative and better integrated with the Hugging Face ecosystem. Changes include:
- Updating the paper link to the official Hugging Face paper page.
- Adding the authors of the paper.
- Shortening and restructuring the abstract into a concise introduction and a dedicated methodology section, as per best practices.
- Adding the official BibTeX citation for easier referencing.
- Including descriptive metadata tags for better discoverability of the model.

Let me know if you have any questions!

README.md

@@ -1,20 +1,36 @@
 ---
+library_name: transformers
 license: mit
 pipeline_tag: image-text-to-text
-library_name: transformers
+tags:
+- fine-grained-recognition
+- chain-of-thought
+- vision-language
+- reasoning
+- qwen2-vl
+- arxiv:2602.07605
 ---
 
 # Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning
 
-This is the official model released for the paper **[Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning](https://openreview.net/pdf?id=kyzHM557gE)**.
+This is the official 3B model released for the paper **[Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning](https://huggingface.co/papers/2602.07605)**.
+
+**Authors**: Hulingxiao He, Zijun Geng, and Yuxin Peng.
+
+## Introduction
 
-## Abstract
+Fine-R1 is a Multi-modal Large Language Model (MLLM) specifically designed to excel in Fine-Grained Visual Recognition (FGVR). While traditional MLLMs often struggle with FGVR compared to contrastive models like CLIP, Fine-R1 bridges this performance gap by incorporating Chain-of-Thought (CoT) reasoning. It achieves state-of-the-art performance, even surpassing strong CLIP-like models, in identifying both seen and unseen fine-grained sub-categories with only 4-shot training.
 
-Any entity in the visual world can be hierarchically grouped based on shared characteristics and mapped to fine-grained sub-categories. While Multi-modal Large Language Models (MLLMs) achieve strong performance on coarse-grained visual tasks, they often struggle with Fine-Grained Visual Recognition (FGVR). Adapting general-purpose MLLMs to FGVR typically requires large amounts of annotated data, which is costly to obtain, leaving a substantial performance gap compared to contrastive CLIP models dedicated for discriminative tasks. Moreover, MLLMs tend to overfit to seen sub-categories and generalize poorly to unseen ones. To address these challenges, we propose Fine-R1, an MLLM tailored for FGVR through an R1-style training framework: (1) Chain-of-Thought Supervised Fine-tuning, where we construct a high-quality FGVR CoT dataset with rationales of "visual analysis, candidate sub-categories, comparison, and prediction", transition the model into a strong open-world classifier; and (2) Triplet Augmented Policy Optimization, where Intra-class Augmentation mixes trajectories from anchor and positive images within the same category to improve robustness to intra-class variance, while Inter-class Augmentation maximizes the response distinction conditioned on images across sub-categories to enhance discriminative ability. With only 4-shot training, Fine-R1 outperforms existing general MLLMs, reasoning MLLMs, and even contrastive CLIP models in identifying both seen and unseen sub-categories, showing promise in working in knowledge-intensive domains where gathering expert annotations for all sub-categories is arduous.
+## Methodology
 
+Fine-R1 employs an R1-style training framework consisting of two key stages:
+
+1.  **Chain-of-Thought Supervised Fine-tuning (SFT)**: This stage involves constructing a high-quality FGVR CoT dataset with rationales covering "visual analysis, candidate sub-categories, comparison, and prediction." This process trains the model to act as a strong open-world classifier.
+2.  **Triplet Augmented Policy Optimization (TAPO)**: This stage enhances the model's robustness and discriminative ability. It uses Intra-class Augmentation to improve robustness to intra-class variance and Inter-class Augmentation to maximize response distinction across sub-categories.
 
 ## GitHub Repository
 
+For code, data, and detailed training/evaluation instructions, please refer to the official repository:
 [https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026](https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026)
 
 ## Model Version
@@ -22,4 +38,17 @@ Fine-R1-3B
 
 ## Usage
 
-This model can be used with the Hugging Face `transformers` library. For detailed usage examples and how to integrate it into your projects, please refer to the official [GitHub Repository](https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026).
+This model can be used with the Hugging Face `transformers` library. For detailed usage examples and how to integrate it into your projects, please refer to the official [GitHub Repository](https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026).
+
+## Citation
+
+If you find this model or the research helpful, please consider citing:
+
+```bibtex
+@article{he2026finer1,
+  title={Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning},
+  author={He, Hulingxiao and Geng, Zijun and Peng, Yuxin},
+  journal={arXiv preprint arXiv:2602.07605},
+  year={2026}
+}
+```