Google's new paper Minerva: solving quantitative reasoning problems with language models

Quantitative reasoning is a field where language models are still far below human level . Solving mathematical and scientific problems requires comprehensive skills , Including using natural language and mathematical symbols to correctly analyze problems 、 Recall relevant formulas and constants , And generate step-by-step solutions involving numerical calculations and symbolic operations . Because of these challenges , it is generally agreed that , Using machine learning to solve quantitative reasoning problems will require significant advances in model architecture and training techniques , Allow the model to access external tools , Such as Python Interpreter , Or a deeper paradigm shift may be needed .

stay “ Use language models to solve quantitative reasoning problems ”（ Soon to be in arXiv Publish on ） in , We introduced Minerva, A language model that can use progressive reasoning to solve mathematical and scientific problems . We show that , By focusing on collecting training data related to quantitative reasoning problems 、 Large scale training model and first-class reasoning technology , We have achieved significant performance improvements in various difficult quantitative reasoning tasks . Minerva Solve such problems by generating solutions that include numerical calculations and symbolic operations , Without relying on external tools such as calculators . The model combines natural language and mathematical symbols to analyze and answer mathematical problems . Minerva Combined with a variety of technologies , Including small sample tips 、 Thought chain or register prompt and majority voting , In the STEM Achieve the most advanced performance on reasoning tasks .