ChefNMR2 elucidates molecular structures from NMR at 82.3% top-1, 360x cheaper than LLMs
ZhongingAlong · x · 2026-11-03
A team with Princeton's Seyedsayamdost group unveiled ChefNMR2 and ChefNMR2-Agent, which infer planar molecular structures from 1D NMR spectra plus a molecular formula.
- Benchmarks: across 2,000+ molecules in 5 test sets, ChefNMR2 hits 82.3% top-1 accuracy vs 35.4% for the best specialized de novo model, 60.1% for best retrieval, and 52.0% for the best LLM — with up to 2,170× lower median runtime and 360× lower cost per molecule. It solved all nine famous natural products tested (taxol, rapamycin, vancomycin aglycone, etc.).
- Calibration: sampling frequencies of multiple candidates serve as confidence estimates and closely track actual correctness; Claude Opus 5 is notably underconfident, GPT-5.6-Sol extremely overconfident.
- Agent: a general LLM evaluates ChefNMR2 candidates via chemical reasoning, chemical-shift prediction, and literature search, adapting sampling budget by confidence. It elucidated four previously unpublished natural products (nocobactin AI-1/AI-2, hybridamides A/B).
- Limitations: requires a molecular formula, no stereochemistry prediction, and accuracy drops for larger molecules dissimilar to training data.
More from Models
- Olmo 3 ships with Nextcloud's llm2 backend for centrally managed local AI features — allen_ai · 2026-11-03
- Nextcloud picked Olmo 3 partly because it runs on consumer-grade hardware — allen_ai · 2026-11-03
- Olmo 3 earns Nextcloud's top green Ethical AI rating: open code, weights, and data — allen_ai · 2026-11-03
- Nextcloud picks Ai2's fully open Olmo 3, earning its highest Ethical AI green rating — allen_ai · 2026-11-03
- Qwen 3.8 27B spontaneously uses Blender MCP to code its way through a physics problem — Balloon_Fan · 2026-11-03
- AA-PersonalityBench goes live: build your own AI personality rankings across 41 models — ArtificialAnlys · 2026-11-03