Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties

This is the official reposaitory for paper Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties.

Setup

docker build -t topology-of-reasoning .
pip install -r requirements.txt

Evaluation

bash scripts/eval.sh

you can get csv file in eval_result/dataset_name/your_model/your_time

Cluster and visualize reasoning graph

bash scripts/cluster_generated.sh

you can get csv file in extract_steps/your_model/dataset_name/target_layer_ratio=0.9
and visualize the reasoning graph with extract_steps/your_model/dataset_name/target_layer_ratio=0.9/visualize

Graph characteristics are calculated in extract_steps/your_model/dataset_name/target_layer_ratio=0.9/results.json

Citation

• To cite our paper:

  @misc{minegishi2025topologyreasoningunderstandinglarge,
    title={Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties}, 
    author={Gouki Minegishi and Hiroki Furuta and Takeshi Kojima and Yusuke Iwasawa and Yutaka Matsuo},
    year={2025},
    eprint={2506.05744},
    archivePrefix={arXiv},
    primaryClass={cs.AI},
    url={https://arxiv.org/abs/2506.05744}, 
  }