Google Research continued pretraining a large language model on scientific papers and mathematical web content, then evaluated chain-of-thought-style solutions across mathematics, physics and other quantitative domains. The system operated in natural language rather than a proof assistant.
Minerva achieved then-state-of-the-art performance on several quantitative reasoning benchmarks and correctly solved nearly one third of a separate set of more than two hundred undergraduate-level science problems evaluated by the researchers.
Minerva helped establish natural-language mathematical reasoning as a serious scaling target for large language models, providing an important precursor to later Olympiad-level end-to-end reasoning systems.
Significance rationale: Minerva showed that a general language-model architecture further trained on technical content could make a large jump on mathematical and scientific quantitative reasoning benchmarks without external symbolic tools.
—
The quantitative reasoning results and evaluation methodology were released in the research paper.
No corrections recorded.
