Journal Watch · Nature Computational Science · 2026
Less can be more for predicting properties with large language models.
Our analysis provides concrete guidelines for architecture selection in scientific machine learning.
TL;DR
Our analysis provides concrete guidelines for architecture selection in scientific machine learning.
Key findings
- This geometric blindness persists regardless of model size (up to 70 billion parameters), dataset scale (up to 2 × 10<sup>6</sup> 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.
What they did
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.
Full abstract (original)
Background: 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.
What they did: 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.
Results & conclusions: This geometric blindness persists regardless of model size (up to 70 billion parameters), dataset scale (up to 2 × 10<sup>6</sup> 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.
Abstract © the authors / Nature Computational Science (2026). Reproduced here for discussion and indexing, with attribution. Read the original at https://doi.org/10.1038/s43588-026-01045-2 · PubMed 42864435.
Why it matters for pharma & biotech
- Directly relevant to teams in AI in drug discovery / diagnostics — the abstract reports new data in that space, so competitive and pipeline teams should read the full text.
- As earlier-stage work, its near-term value is target validation and assay design rather than clinical decisions.
- For searchers: the full abstract is reproduced below, so you can judge relevance before opening the publisher page.
Modalities & themes detected: AI in drug discovery / diagnostics
Limitations & what to check in the full text
- This brief is based on the published abstract only. Effect sizes, subgroup results and safety detail in the full text may qualify the headline findings.
- The abstract reports its conclusions qualitatively, without headline effect sizes — read the full text before citing a magnitude.
Frequently asked questions
What did this Nature Computational Science study find?
Our analysis provides concrete guidelines for architecture selection in scientific machine learning.
What type of study is this?
It is a research study, published in Nature Computational Science in 2026. Here we report fundamental limitations in LLM's ability to learn from coordinate information in coordinate-category data.
Where can I read the original paper?
The original is at the publisher via DOI 10.1038/s43588-026-01045-2. PubMed ID: 42864435. BioStudioAI reproduces the abstract for discussion and links to the source.
This brief: BioStudioAI Journal Watch, 29 October 2026. https://biostudioai.com/blog/papers/less-can-be-more-for-predicting-properties-with-large-language-models-010452/
Reading papers is step one. Analysing the data is the work.
BioStudioAI turns raw lab files into QC'd results, statistics, figures, reports and slides — hosted on your servers, every number computed, nothing typed.
Open BioStudioAI More guides