Iterative Graph Alignment

Fangyuan Yu¹, Hardeep Arora¹, Matt Johnson¹

¹Temus   

Iterative Graph Alignment (IGA) is an annotation-free alignment algorithm. A teacher model (VLM) iteratively generates logical graphs and reference answers using Iterative Graph Prompting (IGP). A student model (LLM) reviews its responses against these reference answers to identify hard cases where representation gaps exist. The student then collaborates with helper models to explore diverse ways to respond to these challenging queries by taking hints from the logical graphs and reference answers, before fine-tuning on the collected insights and proceed to the next iteration.

🆕 Updates

• [08/2024] [arXiv Preprint] https://arxiv.org/abs/2408.16667
• [08/2024] [Dataset Release] (https://github.com/fangyuan-ksgk/RuleEval) Rule-adherance ability evaluation suite.

Install Dependencies

bash set.sh

IGP prompting

python -m script.reason

SAIL training

python -m script.iter

🤗 Citation

If you find this work useful for your research, please kindly cite our paper:

@misc{yu2024iterativegraphalignment,
      title={Iterative Graph Alignment}, 
      author={Fangyuan Yu and Hardeep Singh Arora and Matt Johnson},
      year={2024},
      eprint={2408.16667},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2408.16667}, 
}