1 paper
Woojin Lee, Jin-Xia Huang
State-of-the-art code generation frameworks rely on mental simulation, where LLMs internally trace execution to verify correctness. We expose a fundamental limitation: the Mental-R…