Skip to content
DraconiticResearch
Search Sources API
203,234 studies31 sourcesUpdated 35 min ago
Peer-reviewedCC BYNature computational science

Less can be more for predicting properties with large language models.

  • Nawaf Alampara1
  • Santiago Miret2,3
  • Kevin Maik Jablonka4,5,6,7
7 affiliations
  1. 1Laboratory of Organic and Macromolecular Chemistry, Friedrich Schiller University Jena, Jena, Germany. [email protected]
  2. 2Intel Labs, Santa Clara, CA, USA
  3. 3Lila Sciences, Cambridge, MA, USA
  4. 4Center for Energy and Environmental Chemistry Jena, Friedrich Schiller University Jena, Jena, Germany. [email protected]
  5. 5Helmholtz Institute for Polymers in Energy Applications Jena, Jena, Germany. [email protected]
  6. 6Jena Center for Soft Matter, Friedrich Schiller University Jena, Jena, Germany. [email protected]
  7. 7Laboratory of Organic and Macromolecular Chemistry, Friedrich Schiller University Jena, Jena, Germany. [email protected]
Read full textPDFDOICite

Abstract

Predicting properties from coordinate-category data-sets of vectors paired with categorical information-is fundamental to computational science, manifesting in materials science as predicting formation energies or elastic moduli from crystal structures. While large language models (LLMs) have increasingly been applied to such tasks, optimal strategies for achieving reliable predictions remain elusive. Here we report fundamental limitations in LLM's ability to learn from coordinate information in coordinate-category data. Through systematic experiments using synthetic datasets with tunable coordinate and category contributions, and a comprehensive benchmarking framework (MatText), we find that LLMs consistently fail to capture coordinate information while excelling at category patterns. This geometric blindness persists regardless of model size (up to 70 billion parameters), dataset scale (up to 2 × 10⁶ structures) or text representation strategy. For material property prediction tasks dominated by structural effects, specialized geometric architectures consistently outperform LLMs, as evidenced by a clear 'Graph Neural Network Language Model (GNN-LM) wall' in performance benchmarks. Our analysis provides concrete guidelines for architecture selection in scientific machine learning.

Related

  1. Less is more in mixture modeling: every molecular vector is a map, yet the mixture is the territory
    Xiao Yang, Guanghao Song, Zihan Xu +6
    ChemRxivPreprintCC BY-NCFull textGeneral chemistryChemRxiv
  2. Predicting sensory texture properties of cosmetic emulsions by physical measurements
    Laura Gilbert, Géraldine Savary, Michel Grisel +1
    Chemometrics and Intelligent Laboratory SystemsPeer-reviewedPublisher termsFull text164 citedAnalytical chemistryMaterials & polymersHAL
  3. Large Language Models in Chemistry: Opportunities, Limitations, and Hallucination Risks
    Evalasting Michael Chukwuemeka
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  4. How Chemical Evidence Shapes Mechanistic Reasoning and Prediction by Large Language Models: Fine-Grained Inference in Nucleoside-Derived Supramolecular Gelation
    Ying Wang, Xi Yu
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  5. Dataset-level reliability and quality control for metallurgical data structured with large language models
    D. Wackerling, A. Oñate, D. Rojas +3
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv

Cite

BibTeXRIS
BibTeX
@article{alampara2026less,
  title = {{Less can be more for predicting properties with large language models.}},
  author = {Nawaf Alampara and Santiago Miret and Kevin Maik Jablonka},
  journal = {Nature computational science},
  year = {2026},
  month = oct,
  doi = {10.1038/s43588-026-01045-2},
  url = {https://europepmc.org/article/MED/42864435},
  note = {CC BY},
}
RIS
TY  - JOUR
TI  - Less can be more for predicting properties with large language models.
AU  - Nawaf Alampara
AU  - Santiago Miret
AU  - Kevin Maik Jablonka
PY  - 2026
DA  - 2026/10/12
JO  - Nature computational science
DO  - 10.1038/s43588-026-01045-2
UR  - https://europepmc.org/article/MED/42864435
AB  - Predicting properties from coordinate-category data-sets of vectors paired with categorical information-is fundamental to computational science, manifesting in materials science as predicting formation energies or elastic moduli from crystal structures. While large language models (LLMs) have increasingly been applied to such tasks, optimal strategies for achieving reliable predictions remain elusive. Here we report fundamental limitations in LLM's ability to learn from coordinate information in coordinate-category data. Through systematic experiments using synthetic datasets with tunable coordinate and category contributions, and a comprehensive benchmarking framework (MatText), we find that LLMs consistently fail to capture coordinate information while excelling at category patterns. This geometric blindness persists regardless of model size (up to 70 billion parameters), dataset scale (up to 2 × 10⁶ structures) or text representation strategy. For material property prediction tasks dominated by structural effects, specialized geometric architectures consistently outperform LLMs, as evidenced by a clear 'Graph Neural Network Language Model (GNN-LM) wall' in performance benchmarks. Our analysis provides concrete guidelines for architecture selection in scientific machine learning.
SN  - 2662-8457
ER  -