CoG

[Paper][ACL 2026] CoG: Controllable Graph Reasoning via Relational Blueprints and Failure-Aware Refinement over Knowledge Graphs

Knowledge Graph and Datasets

Before running CoG, please set up Freebase on your local machine by following the installation guide.

We evaluate CoG on CWQ, WebQSP, and GrailQA. The corresponding data files should be placed in the data/ directory.

Code

Our codebase is built with reference to the open-source project PoG. We sincerely appreciate the authors for sharing their implementation.

Running

After completing all necessary configurations, you can run CoG using the following command:

python main_freebase.py \
  --dataset cwq \ # the dataset
  --max_length 4096 \ # the max length of LLMs output
  --temperature_exploration 0.3 \ # the temperature in exploration stage
  --temperature_reasoning 0.3 \ # the temperature in reasoning stage
  --depth 4 \
  --remove_unnecessary_rel True \ # whether removing unnecessary relations
  --LLM_type gpt-3.5-turbo \ # the LLM
  --opeani_api_keys sk-xxxx \ # your own api keys
  --num_workers 10 

Evaluation

We adopt Exact Match as the evaluation metric. After generating the final prediction file, you can evaluate the results with the following example command:

python eval.py \
  --dataset cwq \
  --output_file CoG_cwq_gpt-3.5-turbo.jsonl