Materials LLM Probe Eval

materials-llm-probe-eval

Probing Materials Knowledge in LLMs: From Latent Embeddings to Reliable Predictions — Venugopal et al. (2026) (arXiv:2603.01834, 2026)

What this evaluates

Evaluates large language models' ability to retrieve materials science knowledge and predict continuous physical properties. It probes the fundamental asymmetry in LLM behavior between symbolic tasks (classification, link prediction) and numerical regression tasks, assessing how fine-tuning affects accuracy and output consistency across modalities.

Datasets

• MatKG — total ?; splits: test (-1); repo https://github.com/olivettigroup/LLM-Probe
• Crystal System Classification — total ?; splits: test (-1); repo https://github.com/olivettigroup/LLM-Probe
• Bandgap Prediction — total ?; splits: test (-1); repo https://github.com/olivettigroup/LLM-Probe
• Dielectric Constant Prediction — total ?; splits: test (-1); repo https://github.com/olivettigroup/LLM-Probe

Metrics

• RMSE (primary) — range: other
  • Root Mean Square Error: sqrt(mean((y_pred - y_true)^2)). Lower values indicate better predictive accuracy for continuous property regression.
• Top-1 accuracy (primary) — range: percent
  • Fraction of instances where the model's predicted class or answer exactly matches the ground truth label. Higher values indicate better performance.
• Response entropy — range: other
  • Computed from the variance across 10 independent inference runs to quantify output consistency. Note: for numerical tasks, this reflects output format consistency rather than prediction confidence.

Input / output format

Input: Chemical composition or formula for numerical tasks; natural language questions derived from knowledge graph triples or chemical formulas for symbolic tasks.

Output: Continuous numerical values for bandgap and dielectric constant; class labels or natural language answers for crystal system classification and MatKG link prediction.

Scoring recipe

def compute_metrics(predictions, gold):
    import math
    rmse = math.sqrt(sum((p - g)**2 for p, g in zip(predictions, gold)) / len(gold))
    top1_acc = sum(1 for p, g in zip(predictions, gold) if p == g) / len(gold) * 100
    return {'rmse': rmse, 'top1_accuracy': top1_acc}

Common pitfalls

• Response entropy for numerical tasks reflects output format consistency rather than prediction confidence, so low entropy does not imply reliable predictions.
• Cross-task transfer is highly asymmetric; fine-tuning on symbolic tasks rarely improves numerical prediction and vice versa.
• Base models exhibit 'confident hallucination' on numerical tasks (low entropy, high RMSE), which fine-tuning improves in accuracy but not necessarily in reliability/entropy behavior.

Evidence (verbatim from paper)

Numerical property prediction—bandgap and dielectric constant—is evaluated using root mean square error (RMSE), with response entropy computed from the variance across 10 inference runs (Figure 2).

Citation

@misc{venugopal2026probing,
  title={Probing Materials Knowledge in LLMs: From Latent Embeddings to Reliable Predictions},
  author={Venugopal et al. (2026)},
  year={2026},
  note={arXiv:2603.01834}
}

• arXiv: 2603.01834