CtrlHGen

This is the official code repository for [Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs] (ICLR 2026).

Environment

conda create -n ctrlhgen python=3.9
conda activate ctrlhgen
pip install -r requirements.txt 

Training

As described in the paper, you can run the code in the following steps:

1. Sampling
2. Supervised training
3. Reinforcement learning

Step 1: Sampling

bash scripts/sample/sample_full.sh

Step 2: Supervised training

1. Without condition

bash scripts/train/wn-g2.sh

2. With condition

bash scripts/cond-train/wn-g2-pattern.sh

Step 3: Reinforcement learning

Example scripts:

bash scripts/optim/wn-g2.sh

For training with multi-gpu:

bash scripts/optim/wn-g2-multi.sh

Evaluation

Example scripts:

bash scripts/test/wn-g2.sh

bash scripts/optim-test/wn-g2.sh

Citation

Welcome to cite our work!

@article{gao2025controllable, title={Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs}, author={Gao, Yisen and Bai, Jiaxin and Zheng, Tianshi and Sun, Qingyun and Zhang, Ziwei and Li, Jianxin and Song, Yangqiu and Fu, Xingcheng}, journal={arXiv preprint arXiv:2505.20948}, year={2025} }