← All notes

Summary

Minerva is a language model that starts from general natural-language pretraining and is then further trained on technical content. The authors report state-of-the-art performance on technical benchmarks without external tools. They also evaluate on over two hundred undergraduate-level problems in physics, biology, chemistry, economics, and other sciences, and they say the model can correctly answer nearly a third of them. [S1]

Claims

• id: C1 claim: 'Minerva is pretrained on general natural language data and further trained on technical content.' source: S1 evidence: "pretrained on general natural language data and further trained on technical content"
• id: C2 claim: 'They report state-of-the-art performance on technical benchmarks without external tools.' source: S1 evidence: "state-of-the-art performance on technical benchmarks without the use of external tools"
• id: C3 claim: 'They say the model can correctly answer nearly a third of over two hundred undergraduate-level quantitative problems.' source: S1 evidence: "evaluate our model on over two hundred undergraduate-level problems"

Verification

• Code available: No official code URL was named on the abs page or HTML. The authors say they plan to release problems as part of an open-source dataset in an upcoming manuscript; that later release was not verified here.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live notes 2201-11903-chain-of-thought-prompting and 2203-11171-self-consistency. This is the paper-verification note for Minerva itself.
• Limits: Headline results are from mid-2022 models further trained on technical text, without tools. 'Nearly a third' is the authors' summary of their undergraduate-level set, not a general guarantee on later exams.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 29 Jun 2022 (v1); last revised 1 Jul 2022 (v2). Sources: [S1] https://arxiv.org/abs/2206.14858, v2 (1 Jul 2022). checked_on: 2026-11-11.

When to apply

Use when looking at continued training on scientific and mathematical text for quantitative reasoning, as distinct from tool-using program-aided methods.

Facts

• id: F1 claim: 'Minerva continues training a general language model on technical content.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors report state-of-the-art performance on technical benchmarks without external tools.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report that the model correctly answers nearly a third of over two hundred undergraduate-level quantitative problems they evaluated.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2206.14858, https://arxiv.org/abs/2206.14858. License: CC BY 4.0. v2 (1 Jul 2022). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_8xv80fe1,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Solving Quantitative Reasoning Problems with Language Models}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2206-14858-minerva/},
  note = {Permanent ID iil:note:8xv80fe1. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2206-14858-minerva.md.