MineralExp benchmark tests whether LLM agents can design geochemical experiments
The arXiv paper reformulates experimental design in mineral exploration as a selection problem, giving agents 277 expert-verified cases to score against.
In plain words
Before spending money on lab work, geologists must decide which experiments will test a hypothesis. The MineralExp benchmark checks whether AI agents can make that choice. Current agents tend to pick too many experiments, though they rarely miss the useful ones.
What was announced
The authors say automated geochemical experimental design — deciding which experiments to run before committing to costly laboratory work — has remained largely unexplored while LLM agents increasingly automate downstream geoscientific data analysis. Two obstacles made evaluation difficult: different geologists may choose different minerals and instruments to test the same hypothesis, so no single protocol serves as a reference answer, and open-ended plans resist representation as structured labels, making objective scoring hard.
MineralExp addresses this by recasting design as a hypothesis-based experiment selection problem, a form of multi-label classification. Instrument names and procedural details are abstracted into a taxonomy of 19 experimental types across five mineral-system processes: Source, Transport, Deposition, Modification, and Preservation. Built from published ore-deposit literature and verified by domain experts, the benchmark contains 277 cases and 1,572 reference plans. An agent receives a hypothesis and initial context, selects relevant tasks, and drafts concrete plans, scored by multi-label metrics plus an LLM-as-a-Judge metric. Results show current agents generally achieve high recall but low precision, proposing more experiments than necessary.
What it does not show
- The evidence is the paper's abstract only; full methodology, baselines, and detailed results are not covered.
For security and IT teams
- Scope
- The benchmark covers geochemical experimental design for mineral exploration, not downstream analysis of geoscientific data.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.