README.md

---
license: apache-2.0
datasets:
- Frywind/GREAM_data
---

# 🧠 GREAM: Generative Reasoning Recommendation Model

**Paper:** *[Generative Reasoning Recommendation via LLMs,](https://arxiv.org/pdf/2510.20815) 2025.*  
**Authors:** Minjie Hong\*, Zetong Zhou\*, Zirun Guo, Ziang Zhang, Ruofan Hu, Weinan Gan, Jieming Zhu, Zhou Zhao†  
**Repository:** [https://github.com/Indolent-Kawhi/GRRM](https://github.com/Indolent-Kawhi/GRRM)  
**HF Papers Link:** [https://huggingface.co/papers/2510.20815](https://huggingface.co/papers/2510.20815)  

---

## 🧩 Model Summary

**GREAM** (Generative Reasoning Recommendation Model) is a **large language model (LLM)-based generative reasoning recommender** designed to unify *understanding, reasoning,* and *prediction* for recommendation tasks.  
It introduces a **reasoning-enhanced, verifiable reinforcement learning** framework that allows both high-throughput direct recommendations and interpretable reasoning-based outputs.

### Key Features
- **Collaborative–Semantic Alignment:** Fuses textual (titles, descriptions, reviews) and behavioral signals to align linguistic and collaborative semantics.  
- **Reasoning Curriculum Activation:** Builds synthetic *Chain-of-Thought (CoT)* data and trains via curriculum to develop causal reasoning for recommendations.  
- **Sparse-Regularized Group Policy Optimization (SRPO):** Enables stable RL fine-tuning using *Residual-Sensitive Verifiable Rewards* and *Bonus-Calibrated Group Advantage Estimation* for sparse feedback.

---

## 🧠 Model Architecture

| Component | Description |
|------------|--------------|
| **Backbone** | Qwen3-4B-Instruct|
| **Indexing** | Residual Quantization (RQ-KMeans, 5 levels, 256 values per level) |
| **Training Phases** | ① Collaborative–Semantic Alignment → ② Reasoning Curriculum Activation → ③ SRPO Reinforcement Learning |
| **Inference Modes** | - **Direct Sequence Recommendation:** low-latency item generation<br> - **Sequential Reasoning Recommendation:** interpretable CoT reasoning chains |
| **RL Framework** | Verl + SGLang backend|

---

## 📚 Training Data

| Data Type | Source | Description |
|------------|---------|-------------|
| **D<sub>align</sub>** | Amazon Review Datasets (Beauty, Sports, Instruments) | Sequential, semantic reconstruction, and preference understanding tasks |
| **D<sub>reason</sub>** | Synthetic CoT data generated via GPT-5 / Qwen3-30B / Llama-3.1 | Multi-step reasoning sequences with `<think>...</think>` and `<answer>...</answer>` supervision |
| **Text Sources** | Item titles, descriptions, and high-quality reviews | Combined and rewritten to form dense item semantics |


---

## 📊 Evaluation

### Datasets
- **Amazon-Beauty**
- **Amazon-Sports & Outdoors**
- **Amazon-Musical Instruments**

## Citation

```
@misc{hong2025generativereasoningrecommendationllms,
      title={Generative Reasoning Recommendation via LLMs}, 
      author={Minjie Hong and Zetong Zhou and Zirun Guo and Ziang Zhang and Ruofan Hu and Weinan Gan and Jieming Zhu and Zhou Zhao},
      year={2025},
      eprint={2510.20815},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2510.20815}, 
}
```