In a test involving 27 organic molecules and 810 spatial conformations, several frontier LLMs ranked molecular stability more accurately than the classical UFF force field and came close to modern GFN-FF. This is worth paying attention to because the boundaries of language models' chemical reasoning may be closer to professional tools than we previously believed.
What This Is
Conformation ranking determines which spatial arrangement of the same molecule is more stable. It directly affects drug molecule design, reaction pathways, and materials calculations, and it must account for three-dimensional relationships beyond atomic connectivity.
A team from the Russian Institute of Organic Chemistry and Florida State University used DFT results—calculating molecular properties through quantum-mechanical methods—as the benchmark to test LLMs from different generations. GPT-5.6 Sol, Kimi K3, and Gemini 3.6 Flash surpassed the classical force field, while GPT-6 Astra and Claude Opus 5 approached modern force fields.
The failure patterns are even more revealing: models performed well on well-defined, thoroughly quantified chemical concepts such as hydrogen bonding, but generally made errors on concepts with fuzzy boundaries, such as ring strain. In other words, the models did not first acquire a complete chemical intuition; they developed stronger capabilities in the parts of chemistry that can be described and computed in language.
Industry View
The findings support a more restrained conclusion: text-only training does more than teach models to repeat knowledge; it may also enable them to form transferable representations of three-dimensional structures from papers and textbooks. The correlation between this capability and performance on scientific, coding, and abstract-reasoning tests provides further support for that explanation.
But the counterarguments are equally clear. The test covered only 27 small molecules and cannot represent the vast chemical space. The models still read textual descriptions of geometry, which is not the same as genuinely perceiving three dimensions. Nor does correlation with reasoning-test performance prove that reasoning is the cause of conformation-ranking ability. So-called “emergence” remains a hypothesis, not a confirmed mechanism.
Accordingly, we would not interpret this as “models have already replaced professional chemistry software.” The more defensible conclusion is that AI has entered verifiable professional workflows. Its next value will depend on whether its results hold up across more molecules, experiments, and real-world projects.
Impact on Regular People
For enterprise IT: AI evaluations at pharmaceutical, materials, and R&D companies should shift from question-answering ability toward verifiable tasks. For companies that already have structured data and professional software, LLMs are more likely to become search, explanation, and option-screening layers first, rather than autonomous decision-making systems.
For individual careers: Roles in chemistry, pharmaceuticals, materials, and research support will encounter model-assisted ranking and validation earlier than many other fields. In the short term, models are strongest on problems with clear criteria, while professionals must still review boundary conditions and failure cases.
For consumer markets: Models participating in molecular screening could shorten some early-stage R&D cycles, but this will not translate directly into cheaper medicines or materials. Drug development remains constrained by much longer chains involving synthesis, clinical trials, safety, and regulation.