MTA: A Merge-then-Adapt Framework for Personalized Large Language Model

MTA is a framework for Personalized Large Language Models (PLLMs) that addresses two key challenges in user-centric LLM adaptation: (1) storage costs that scale linearly with the number of users, and (2) suboptimal performance for users with sparse data. MTA comprises three stages:

1. Meta-LoRA Bank Construction — Select anchor users and pre-train meta-personalization traits in meta-LoRA modules.
2. Adaptive LoRA Fusion — Retrieve and dynamically merge the most relevant anchor meta-LoRAs to synthesize a user-specific adapter, eliminating per-user storage.
3. LoRA Stacking — Apply an additional ultra-low-rank LoRA on top of the merged LoRA for effective few-shot personalization.

Paper: arXiv:2511.20072 | Accepted by ACL 2026 Main Conference 🎉

🛠️ Environment Setup

Create and activate a new conda environment named MTA:

conda create -n MTA python=3.10
conda activate MTA

Install required dependencies:

pip install -r requirements.txt

Note: LlamaFactory will be installed separately in the next step.

📥 Model Download

Download the required models to the ./model directory:

# Create model directory
mkdir -p ./model

# Download BGE embedding model for dense retrieval
git clone https://huggingface.co/BAAI/bge-small-en-v1.5 ./model/bge-small-en-v1.5

# Download DeBERTa v3 Large model
git clone https://huggingface.co/microsoft/deberta-v3-large ./model/deberta-v3-large

# Download Llama 3 8B Instruct model
git clone https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct ./model/Meta-Llama-3-8B-Instruct

Alternative download with huggingface-hub:

# Install huggingface-hub if not already installed
pip install huggingface-hub

# Download models using huggingface CLI
huggingface-cli download BAAI/bge-small-en-v1.5 --local-dir ./model/bge-small-en-v1.5
huggingface-cli download microsoft/deberta-v3-large --local-dir ./model/deberta-v3-large
huggingface-cli download meta-llama/Meta-Llama-3-8B-Instruct --local-dir ./model/Meta-Llama-3-8B-Instruct

📊 Dataset

We use public available data from the LaMP benchmark for anchor selection. The dataset processing approach follows the same methodology as Per-Pcs.

The processed datasets for anchor selection and training should be placed under the ./data directory with the following structure:

./data/
├── citation/
├── movie_tagging/
├── news_headline/
├── product_rating/
└── scholarly_title/

Each task directory contains:

• user_anchor_candidate.json: Candidate users for anchor selection
• Test data for evaluation

🚀 Usage

🎯 Select Anchor Users

python select_anchor.py --candidate_path <data_path> --task_name <task> --k <num_anchors>

Example:

python select_anchor.py --candidate_path ./data/movie_tagging/user_anchor_candidate.json --task_name movie_tagging --k 50

Parameters:

• --candidate_path: Path to candidate user JSON file
• --task_name: Task name (movie_tagging, news_headline, citation, etc.)
• --k: Number of anchor users to select

🧠 Train Meta-LoRA Bank

git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory && git checkout v0.9.2 && pip install -e . && cd ..
python prepare_meta_lora.py --task_name <task> --top_k 1
bash run_<task>_k1_training.sh

Example:

git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory && git checkout v0.9.2 && pip install -e . && cd ..
python prepare_meta_lora.py --task_name movie_tagging --top_k 1
bash run_movie_tagging_k1_training.sh

Parameters:

• --task_name: Task name (same as step 1)
• --top_k: Number of history items for BM25 retrieval (default: 1)

🔍 Dense Retrieval Matching

python dense_retrieval_matcher.py --task_name <task> --top_k <num_matches>

Example:

python dense_retrieval_matcher.py --task_name movie_tagging --top_k 2

Parameters:

• --task_name: Task name (movie_tagging, citation, news_headline, product_rating, scholarly_title)
• --top_k: Number of top similar anchor users to return (default: 2)

Output: Results saved to ./dense_retrieval_results/{task_name}/matches_results.json

🧪 Complete MTA Experiment

After completing the above three steps, you can run the complete MTA experiment using the unified script:

bash run_experiment.sh <task_name> <num_users>

Example:

bash run_experiment.sh citation 10

Parameters:

• <task_name>: Task name (movie_tagging, citation, news_headline, product_rating, scholarly_title)
• <num_users>: Number of few-shot users to process

Output Structure:

./merge_lora/hybrid_{task_name}/           # Fused LoRA models
├── merge_lora_{user_id}/
└── ...

./adapt_lora/hybrid_{task_name}/           # Adapted stacked LoRA models  
├── adapt_lora_{user_id}/
└── ...

./final_eval_results/hybrid_{task_name}/   # Evaluation results
├── results_{user_id}_alpha_{weight}.json
├── final_aggregated_results_{task_name}.json
└── ...

📝 Citation

@misc{li2025mtamergethenadaptframeworkpersonalized,
      title={MTA: A Merge-then-Adapt Framework for Personalized Large Language Model},
      author={Xiaopeng Li and Yuanjin Zheng and Wanyu Wang and wenlin zhang and Pengyue Jia and Yiqi Wang and Maolin Wang and Xuetao Wei and Xiangyu Zhao},
      year={2025},
      eprint={2511.20072},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.20072},
}