Speaker
Vasily Nesterov
(Moscow Institute of Physics and Technology)
Description
Large language models have achieved remarkable performance on a wide range of tasks that require natural language understanding. As recent studies show, they are able to solve tasks that require mathematical reasoning, such as solving problems and formalizing proofs. But how big are the language models needed for these tasks? We study whether it is possible to achieve comparable quality on open-source medium-sized models. We show that solving problems in natural language is possible on such models, while autoformalization requires larger ones.
Primary author
Vasily Nesterov
(Moscow Institute of Physics and Technology)
Co-author
Mr
Alexander Perepechko
(Moscow Institute of Physics and Technology)
