README.md

---
base_model:
- Qwen/Qwen2.5-32B-Instruct
datasets:
- liuwenhan/reasonrank_data_sft
- liuwenhan/reasonrank_data_rl
- liuwenhan/reasonrank_data_13k
language:
- en
license: mit
pipeline_tag: text-ranking
library_name: transformers
tags:
- reranking
- reasoning
- qwen
---

# ReasonRank: Empowering Passage Ranking with Strong Reasoning Ability

## Introduction
This is the model trained in our paper: **ReasonRank: Empowering Passage Ranking with Strong Reasoning Ability** ([📝arXiv](https://arxiv.org/abs/2508.07050)).

Large Language Model (LLM) based listwise ranking has shown superior performance in many passage ranking tasks. With the development of Large Reasoning Models, many studies have demonstrated that step-by-step reasoning during test-time helps improve listwise ranking performance. ReasonRank addresses the scarcity of reasoning-intensive training data by proposing an automated reasoning-intensive training data synthesis framework. To empower the listwise reranker with strong reasoning ability, we further propose a two-stage post-training approach, which includes a cold-start supervised fine-tuning (SFT) stage for reasoning pattern learning and a reinforcement learning (RL) stage for further ranking ability enhancement.

Please refer to our [🧩GitHub repository](https://github.com/8421BCD/ReasonRank) for detailed usage instructions and code.

Project page: [https://brightbenchmark.github.io/](https://brightbenchmark.github.io/)

## Model Performance
<p align="center">
<img width="90%" alt="image" src="https://8421bcd.oss-cn-beijing.aliyuncs.com/img/image-20250810163757771.png" />
</p>

## Sample Usage

You can use this model with the `transformers` library. Here is a basic example to perform inference. Note that the exact prompt construction for ReasonRank is critical for performance and should ideally follow the `create_prompt` function in the original [GitHub repository's `rerank/rank_listwise_os_llm.py` file](https://github.com/8421BCD/ReasonRank/blob/main/rerank/rank_listwise_os_llm.py). The example below provides a simplified structure for demonstration.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import torch

# Load the model and tokenizer
model_id = "liuwenhan/reasonrank-32B" # Assuming this is the model being documented
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # or torch.float16 depending on your GPU and needs
    device_map="auto",
    trust_remote_code=True # Required for custom modeling files (Qwen components)
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

# Example query and passages
query = "What is the capital of France?"
passages = [
    "Paris is the capital and most populous city of France.",
    "London is the capital of England and the United Kingdom.",
    "The Eiffel Tower is a famous landmark in Paris.",
    "France is a country in Western Europe."
]

# Construct the input messages for Qwen's chat template.
# For ReasonRank's specific prompt structure, refer to the original GitHub repository's
# `rerank/rank_listwise_os_llm.py` file and `add_prefix_prompt`/`add_post_prompt` functions.
# This example uses a general Qwen-like structure for demonstration.
system_prompt = "You are a helpful and intelligent assistant."
user_prefix = f"For the query: '{query}', please rank the following passages from most relevant to least relevant.\
"
passage_list_str = "\
".join([f"[{i+1}] {p}" for i, p in enumerate(passages)])
user_suffix = "\
Now, please generate the reasoning process and the ranked list of passages."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": f"{user_prefix}{passage_list_str}{user_suffix}"}
]

# Apply the chat template to get the final prompt string
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Tokenize the input
inputs = tokenizer(prompt, return_tensors="pt", padding=True).to(model.device)

# Generate response
# Use generation_config from the model if available, otherwise define
generation_config = model.generation_config if model.generation_config else GenerationConfig()
generation_config.max_new_tokens = 512
generation_config.do_sample = False # For greedy decoding
generation_config.temperature = 0.1 # Keep temperature low for ranking tasks
generation_config.top_p = 0.95


with torch.no_grad():
    outputs = model.generate(
        **inputs,
        generation_config=generation_config
    )

# Decode the output
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(f"Query: {query}\
Response:\
{response}")

# Expected (simplified) output might look like:
# Response:
# Reasoning: The query asks for the capital of France. Passage [1] directly states "Paris is the capital and most populous city of France."
# This makes it the most relevant. Other passages are less direct or irrelevant.
# Ranked List:
# 1. [1] Paris is the capital and most populous city of France.
# 2. [3] The Eiffel Tower is a famous landmark in Paris.
# 3. [4] France is a country in Western Europe.
# 4. [2] London is the capital of England and the United Kingdom.
```

## Citation

If you find this work helpful, please cite our papers:

```bibtex
@misc{liu2025reasonrankempoweringpassageranking,
      title={ReasonRank: Empowering Passage Ranking with Strong Reasoning Ability}, 
      author={Wenhan Liu and Xinyu Ma and Weiwei Sun and Yutao Zhu and Yuchen Li and Dawei Yin and Zhicheng Dou},
      year={2025},
      eprint={2508.07050},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2508.07050}, 
}
```

## License

This project is released under the [MIT License](https://opensource.org/licenses/MIT).

## Acknowledgement

The inference codes and training implementation build upon [RankLLM](https://github.com/castorini/rank_llm), [Llama Factory](https://github.com/hiyouga/LLaMA-Factory) and [verl](https://github.com/volcengine/verl). Our work is based on the [Qwen2.5](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) model series, and we sincerely thank the Qwen team for their outstanding contributions to the open-source community.