SPARK: Adaptive Parametric Knowledge Participation for Faithful Retrieval-Augmented Generation

Ruochang Li^1*, Pengcheng Huang^1*, Zhenghao Liu^1†, Yukun Yan²,
Huiyuan Xie², Yu Gu¹, Ge Yu¹, Maosong Sun²

¹Northeastern University, ²Tsinghua University

📖 Overview

SPARK is a lightweight framework that learns adaptive knowledge participation policies for faithful retrieval-augmented generation. With the backbone LLM frozen, its participation policy dynamically regulates layer-wise FFN contributions, determining when and to what extent parametric knowledge participates during generation. Optimized through faithfulness-guided reinforcement learning using response-level rewards, SPARK mitigates catastrophic forgetting and better preserves general generation capabilities compared to full-parameter supervised fine-tuning.

⚙️ Setup

1. Create Conda Environment

conda create --name spark python==3.10.0
conda activate shift
git clone https://github.com/OpenBMB/SPARK.git
cd SPARK

2. Install PyTorch

pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124

3. Install Flash Attention

pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

4. Install the rest of the dependencies

pip install -r requirements.txt

5. Patch vLLM model files
   This project requires modifications to the vLLM implementations of Qwen3 and LLaMA. After installing the requirements, run the patch script:

bash vllm/patch_vllm.sh

Note: Please run this script after 'pip install -r requirements.', because installing or reinstalling vLLM may overwrite the patched files.

⚡ Data

Our training data can be downloaded from SPARK_Training_Data. After downloading, place the files into the dataset folder.

To construct the data from scratch, download the files from MRQA-Shared-Task-2019.
Use the downloaded data to synthesize it with FlashRAG.

🔥 Training

GRPO with a single GPU:

python single_gpu.py

GRPO with multiple GPUs:

python multi_gpu.py

📊 Evaluation

We provide the SFT checkpoints and the trained checkpoints(https://huggingface.co/ITcoder/SPARK) on Hugging Face.

To evaluate on the MRQA and ConfiQA benchmarks:

python eval.py

To evaluate on the MMLU benchmark, use lm-evaluation-harness.

🎯 Analysis

We also provide the t-SNE visualization pipeline for participation vectors in SPARK, with corresponding figures available under the figs folder:

python run_batch_tsne.py

For Qwen-3-1.7B:

For Qwen-3-8B:

🎉 Acknowledgement

Our work is built on the following codebases, and we are deeply grateful for their contributions.

• vllm
• nano-aha-moment
• lm-evaluation-harness
• gated_attention
• FlashRAG

🥰 Citation

If you find this work useful, please cite our paper and give us a shining star 🌟

@misc{li2026sparkadaptiveparametricknowledge,
      title={SPARK: Adaptive Parametric Knowledge Participation for Faithful Retrieval-Augmented Generation}, 
      author={Ruochang Li and Pengcheng Huang and Zhenghao Liu and Yukun Yan and Huiyuan Xie and Yu Gu and Ge Yu and Maosong Sun},
      year={2026},
      eprint={2606.27786},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.27786}, 
}

📧 Contact

If you have questions, collaboration opportunities, or potential PhD opportunities in the United States, please feel free to email:

ruochangli@gmail.com