Published in the Proceedings of the 41st ACM/SIGAPP Symposium on Applied Computing (SAC '26), Thessaloniki, Greece, pages 908-916.
The paper proposes an in-context learning method that iteratively picks few-shot examples to help LLMs solve algebraic expressions under non-standard rules in which the priority of addition and multiplication is swapped. LLMs handled these tasks poorly, but iterative example selection combined with explicit reasoning instructions helped, and some models generalized better when given simpler examples.
Authors: Stefano Fioravanti, Matteo Zavatteri, Roberto Confalonieri, Kamyar Zeinalipour, Paolo Frazzetto, Alessandro Sperduti, Nicolò Navarin.