Latent Crystallographic Microscope: Probing the Emergent Crystallographic Knowledge in Large Language Models

Experiments

1. Format Understanding and Property Extraction

Evaluates LLM capabilities for:

• Format recognition (CIF vs POSCAR)
• Property extraction across multiple complexity tiers

2. Onset Analysis

Uses activation patching to identify where crystallographic reasoning emerges:

• Coordinate patching: Spatial coordinate correction
• Stability judge: Thermodynamic stability assessment
• Valence verifier: Charge neutrality reasoning

3. Onset Layer Intervention

Tests injecting stability vectors during generation to improve crystal structure stability.

Setup

1. Install dependencies:

pip install torch transformers vllm pandas numpy pymatgen

2. Store crystal stuctures to be tested in the ../data/ directory

3. Run experiments from each directories

Citation

@inproceedings{
    gan2025latent,
    title={Latent Crystallographic Microscope: Probing the Emergent Crystallographic Knowledge in Large Language Models},
    author={Jingru Gan and Yanqiao Zhu and Wei Wang},
    booktitle={Mechanistic Interpretability Workshop at NeurIPS 2025},
    year={2025},
    url={https://openreview.net/forum?id=28HkQ6mKxn}
}