Researchers have discovered that GPT-3 models can execute algorithms involving loops by using iterative regimented self-attention (IRSA).This method employs three strategies: strong repetitive structures in execution paths, prompting with execution path fragments, or explicitly skipping self-attention to certain parts of generated text. IRSA leads to greater accuracy gains than upgrading to the more powerful GPT-4 model in dynamic program execution. This approach has promising applications in education, as the prompts and responses resemble data structures and algorithms assignments. The findings also highlight the crucial role of prompt design in large language model (LLM) performance, as prompts that do not cover one full task example can still trigger algorithmic behavior, allowing GPT-3 models to solve problems previously considered difficult for LLMs, such as logical puzzles.