This behavior is indeed strange and seemingly low hanging fruit. The model often is able to notice that it has given an incorrect answer and correct itself. But why not just give the correct answer to start with? It’s like it needs a second pass, but is there really that much extra information in its first wrong answer?