Princeton's ChefNMR2 Decodes Molecular Structures from NMR, Beating Frontier LLMs

A Princeton team (the labs of Ellen Zhong and Mo Seyedsayamdost), together with Edison Scientific, has released ChefNMR2 and its agent version: given a molecular formula (e.g., from mass spectrometry) plus a list of 1D ¹H/¹³C NMR peaks, it can infer the planar structure of a molecule from scratch, directly generate a 3D molecular structure, and provide calibrated uncertainty estimates. The work shows that specialized scientific AI systems now far outperform general-purpose LLMs on small-molecule structure elucidation.

Confirmed

• ChefNMR2 combines molecular generative models with a scientific reasoning agent to elucidate unknown small-molecule structures from a molecular formula plus 1D ¹H/¹³C NMR spectra, reportedly two to three orders of magnitude faster than frontier LLMs.
• Compared with the previous version, the diffusion-based generator has been upgraded to flow matching, with a new architecture built on a multimodal diffusion transformer.
• The team partnered with Edison Scientific for large-scale benchmarking: 5 test sets covering 2,000+ molecules (including experimental NMR data), compared against 6 frontier reasoning LLMs and specialized tools; ChefNMR2 achieved a Top-1 hit rate of 82.3% versus 52% for the best LLM, at roughly 1/360 of the cost.
• The authors note that since the release of Claude Opus 4.8 in June 2025, chatbots have shown a sudden jump in NMR interpretation performance, suggesting LLMs already possess general intelligence approaching chemist-level on this kind of scientific task — an observation that motivated this benchmark study.

Why it matters

• NMR structure elucidation is a frequent, time-consuming task in chemistry and drug discovery; ChefNMR2's accuracy and cost advantages suggest the specialized "generative model + reasoning agent" approach may substantially outperform general LLMs on scientific tasks.
• The sharp leap by frontier LLMs on this task also offers a quantifiable case study for when general intelligence can take over expert-level scientific work.

2026-11-03 ~ 2026-11-03 · 8 related posts

• [source] Princeton's ChefNMR2 elucidates small-molecule structures from NMR spectra, 100-1000x faster than frontier LLMs — ZhongingAlong · 2026-11-03
• LLMs suddenly got good at NMR since Claude Opus 4.8, says ChefNMR2 team — ZhongingAlong · 2026-11-03
• [source] ChefNMR2 released: generates 3D molecular structures from 1D NMR with calibrated uncertainty — ZhongingAlong · 2026-11-03
• [source] ChefNMR2 benchmark: 82.3% top-1 accuracy on 2,000+ molecules vs 52% for best LLM — ZhongingAlong · 2026-11-03
• ChefNMR2 Correctly Elucidates All Nine Test Natural Products, from Taxol to Vancomycin — ZhongingAlong · 2026-11-03

3 near-duplicate retellings: ZhongingAlong · ZhongingAlong · ZhongingAlong