ALDEN: Agentic Long-document Document Intelligence

ALDEN is a multi-modal reinforcement learning framework designed for Agentic Visually-Rich Document Understanding (A-VRDU). Built upon Qwen2.5-VL, it introduces a novel fetch action, cross-level reward and a visual semantic anchoring mechanism to enable efficient navigation and reasoning over long, high-resolution documents.

This repository contains the official implementation of our paper: ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents..

🛠️ Installation

Installing the Training Environment

conda create -n alden python=3.10
conda activate alden
git clone https://github.com/gipplab/ALDEN.git
cd ./ALDEN
pip install -e .

Installing the Single-Vector Retriever Environment

conda create -n alden-sv python=3.10
cd ./ALDEN
pip install -r single-vec_retriever_requirements.txt
cd ./flashrag
pip install -e .

Installing the Multi-vector Retriever Environment

conda create -n alden-mv python=3.10
cd ./ALDEN
pip install -r multi-vec_retriever_requirements.txt
cd ./flashrag
pip install -e .

📂 Dataset Preprocessing

1. Corpus Building

We provide the processed training corpus on Hugging Face: SkyFishQ/ALDEN.

If you wish to build the corpus from scratch using your own data:

1. Modify the raw_data_path and target_path in rag_serving/build_corpus.py.
2. Run the build script:

python rag_serving/build_corpus.py

2. Image Index Building

We use flashrag to build the dense retrieval index for document images.

cd ./flashrag/flashrag/retriever

python index_builder.py \
    --retrieval_method vdr-2b-v1 \
    --model_path llamaindex/vdr-2b-v1 \
    --corpus_path /path/to/your/images_corpus/images.parquet \
    --save_dir /path/to/save/images_index \
    --max_length 512 \
    --batch_size 128 \
    --faiss_type Flat \
    --index_modal image \
    --sentence_transformer \
    --save_embedding

python index_builder.py \
	--retrieval_method gte-Qwen2-1.5B-instruct \
	--model_path Alibaba-NLP/gte-Qwen2-1.5B-instruct \
	--corpus_path /path/to/your/images_corpus/images.parquet \
	--save_dir /path/to/save/images_index \
	--max_length 4096 \
	--batch_size 128 \
	--faiss_type Flat \
	--index_modal text \
	--sentence_transformer \
	--save_embedding

python index_builder.py 
	--retrieval_method jina-colbert-v2 \
	--model_path jinaai/jina-colbert-v2 \
	--corpus_path /path/to/your/images_corpus/images.parquet \ 
	--save_dir /path/to/save/images_index \ 
	--max_length 4096 \
	--batch_size 128 \
	--faiss_type Flat \
	--index_modal text \
	--save_embedding

python index_builder.py 
	--retrieval_method colqwen2-v1.0 \
	--model_path vidore/colqwen2-v1.0 \
	--corpus_path /path/to/your/images_corpus/images.parquet \ 
	--save_dir /path/to/save/images_index \
	--max_length 4096 \
	--batch_size 128 \
	--faiss_type Flat \
	--index_modal image \
	--save_embedding

Note: Please replace /path/to/your/... with your actual file paths.

🚀 Launch RL Training

ALDEN uses a decoupled architecture where the environment (RAG tools) and the agent (RL training) run separately.

Step 1: Tool Environment Serving

First, launch the RAG environment server which handles the <search> and <fetch> actions.

1. Get the Server IP:

   hostname --ip-address

   Take note of this IP address, you will need to configure it in the training script.

2. Start the Service:

   python rag_serving/serving.py \
       --config rag_serving/serving_config_single-vec.yaml \
       --num_retriever 8 \
       --port 42354

   or

   python rag_serving/serving.py \
       --config rag_serving/serving_config_multi-vec.yaml \
       --num_retriever 8 \
       --port 42354

Step 2: RL Training

Once the tool server is running, start the training. Ensure the server URL in the training script points to the IP obtained in Step 1.

bash examples/baselines/qwen2_5_vl_7b_doc_agent_ppo.sh

⚡ Inference

To run inference on test sets:

bash examples/baselines/qwen2_5_vl_7b_doc_agent_generation.sh

💾 Model Utils

Merge Checkpoints in the Hugging Face Format

python3 scripts/model_merger.py \
    --local_dir checkpoints/easy_r1/exp_name/global_step_1/actor

📜 Citation

If you find this project useful, please cite our paper:

@article{yang2025alden,
  title={ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents},
  author={Yang, Tianyu and Ruas, Terry and Tian, Yijun and Wahle, Jan Philip and Kurzawe, Daniel and Gipp, Bela},
  journal={arXiv preprint arXiv:2510.25668},
  year={2025}
}

🙌 Acknowledgements

This work is built upon the following excellent open-source projects:

• EasyR1: For the RL infrastructure.
• VAGEN: For visual agent baselines.
• verl: For efficient RL training.
• ReCall: For RAG integration concepts.

We greatly appreciate their valuable contributions to the community.