ARK: Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning

Figure 1: Overview of the ARK framework.

Overview

Query Construction & Contrastive Finetuning: ARK constructs a knowledge graph from documents, extracts query-based subgraphs, and generates augmented queries to mine hard negative chunks for contrastive learning.

Figure 2: Query augmentation via KG-driven subgraph extraction.

Answer-Centric Alignment & Curriculum Learning: We rank chunks by forward/backward alignment scores (whether a chunk helps generate the correct answer), then progressively train the retriever from easy to hard negatives across three curriculum stages.

Figure 3: Answer-centric alignment and curriculum-based contrastive learning.

Quick Start

pip install -r requirements.txt
cp .env.example .env  # Setup API keys

Pipeline

# 1. Knowledge Graph Construction
bash experiments/1-kg/generate_kg_{vllm,openai}.sh
bash experiments/1-kg/augment_kg.sh
bash experiments/1-kg/generate_comm.sh

# 2. Training Data Generation
bash experiments/2-training/generate_pos.sh
bash experiments/2-training/generate_query.sh
bash experiments/2-training/generate_neg.sh
bash experiments/2-training/generate_training.sh

# 3. Train
bash experiments/2-training/train_qwen_{single,multi}.sh

# 4. Evaluate
bash experiments/3-eval/base/*.sh
bash experiments/3-eval/finetuned/run_qwen.sh

# 5. Stats
bash experiments/4-stats/kg.sh
bash experiments/4-stats/coverage.sh

Config

Configuration files in src/config/:

• alignment.yaml - Alignment score computation, subgraph extraction, query_generation, training samples
• kg.yaml - Knowledge graph construction
• llm.yaml - LLM API and inference settings
• retrieval_model.yaml - Retrieval model settings
• training.yaml - Training params

Results

All experiments use Mistral-7B v0.2 as the reader and Qwen3-Embedding-0.6B as the default ARK retriever. For fair comparison, all methods retrieve Top-5 chunks (chunk size = 512, overlap = 12).

LongBench (F1 Score):

Model	NarrativeQA	Qasper	MuSiQue	2WikiMQA	HotpotQA
Qwen3-embedding	19.58	23.90	14.19	21.24	35.27
BGE-M3	18.37	23.33	21.13	22.86	38.64
Stella-v5	20.90	23.39	17.08	22.13	35.45
HippoRAG	11.51	21.90	13.09	30.96	28.71
ARK (Ours)	21.57	24.04	20.60	23.41	42.35

UltraDomain (F1 Score):

Model	Biology	Fiction	Music	Technology	Philosophy
Qwen3-embedding	32.99	29.41	34.90	38.03	34.04
BGE-M3	32.52	31.72	35.34	39.13	35.97
Stella-v5	33.85	32.41	35.02	35.16	34.09
HippoRAG	36.13	29.23	32.94	27.15	29.06
ARK (Ours)	36.19	32.59	38.03	40.16	37.86

Acknowledgments

Models (Base retriever and training framework):

• Qwen3-Embedding-0.6B
• MS-SWIFT
• LLM Providers: Gemini, DeepSeek, GPT

Datasets (Evaluation benchmarks):

• LongBench
• UltraDomain

Baselines (Comparison methods):

• HippoRAG
• LightRAG
• MemoRAG
• GraphRAG
• Nano-GraphRAG

Citation

@article{zhou2025ark,
  title={ARK: Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning},
  author={Zhou, Jiawei and Ding, Hang and Jiang, Haiyun},
  journal={arXiv preprint arXiv:2511.16326},
  year={2025}
}

@misc{zhou2025ark_repo,
  author={Zhou, Jiawei and Ding, Hang},
  title={Code Repo for ARK: Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning},
  year={2025},
  url={https://github.com/valleysprings/ARK},
}

License

MIT License