MostafaAboalkasim/cpp-overflow-blindspot

As a competitive programmer I have noticed that language models often make mistakes with integer overflow in C++. For example, they may use long long correctly but still calculate (a * b) % m without considering that the multiplication can overflow before the modulo is applied, so when a, b, and m can be as large as 1e18, the product can exceed 64bit limits. Using int128 or a safe modular multiplication algorithm can prevent this.

I chose Qwen/Qwen2.5-Coder-3B-Instruct as it was trained on 5.5 trillion code and math tokens, and tested it with 10 simple problems that have modular arithmetic with large inputs up to 1e18, also used the official chat template with greedy decoding. Compiled the generated C++17 code and then used Python to check the results.

The model has generated valid C++ code and used long long but all 10 solutions failed because of intermediate integer overflow, this results show a potential blind spot.

I think to fix this problem we need Include tests with large input values that expose integer overflow and reward models for passing tests that check edge cases not just producing code that compiles.