Frontier Language Models are not Robust to Adversarial Arithmetic, or "What do I need to say so you agree 2+2=5?

C. Daniel Freeman, Laura Culp, Aaron Parisi +27