Generalizable Reasoning through Compositional Energy Minimization

Alexandru Oarga*, and Yilun Du*, In Neural Information Processing Systems (NeurIPS), 2025 [Paper] [Project Page]

1. Requirements

hydra-core
torch
torch-geometric
torch-scatter
torch-sparse

2. Dataset

• Data and generation scripts are available in comp_reasoning/data/ folder.
• 3SAT data can be downloaded from here
• Crossword precomputed embeddings can be downloaded here

3. Training

• Train N-queens

  ./scripts/train_nqueens.sh

• Train 3-SAT

  ./scripts/train_3sat.sh

• Train Graph Coloring

  ./scripts/train_color.sh

• Train Crosswords

  ./scripts/train_color.sh

4. Evaluation

• Evaluate N-queens

  ./scripts/eval_nqueens.sh

• Evaluate 3-SAT

  ./scripts/eval_3sat_20.sh
  ./scripts/eval_3sat_50.sh

• Evaluate Graph Coloring

  ./scripts/eval_color.sh

  To evaluate on multiple graph files run:

  ./scripts/eval_colors_dir.sh

• Evaluate Crosswords

  ./scripts/eval_crosswords.sh

Bibtex

@InProceedings{oarga2025generalizable,
    author    = {Oarga, Alexandru and Du, Yilun},
    title     = {Generalizable Reasoning through Compositional Energy Minimization},
    booktitle = {Advances in Neural Information Processing Systems},
    year      = {2025}
}