Tag
This paper investigates representation robustness in LLMs for mathematical problem solving by systematically varying surface representations of equivalent problems, finding substantial sensitivity and showing that code-augmented reasoning does not uniformly eliminate brittleness.