GPT-5.5 study suggests there is no ideal prompt for competitive programming
A preprint tested nine instruction styles on problems from the International Olympiad in Informatics and on optimization problems. The best prompt depended on the task and on how the results were measured.

Photo: Unknown author · CC0 1.0 · via Rawpixel
Anyone who uses a language model for programming tends to settle on a favorite way of asking it for things. A new study, published as a preprint on arXiv, suggests that such a universal formula may not exist, at least in competitive programming. According to its results, the best instruction style changes with the type of problem and even with how the results are summarized.
The paper, by Saad B. Alarifi and Mohammed A. Alam, was posted to arXiv on October 21, 2026. The authors describe it as an exploratory study. They tested one prompt (the instruction given to the model) for each of nine styles on GPT-5.5, across five tasks: three accuracy tasks drawn from the International Olympiad in Informatics (IOI) and two execution-speed tasks based on NP-hard problems, a class of problems that are notoriously difficult to solve optimally.
According to the abstract, what is new is that previous work had compared prompting strategies only on accuracy. This study adds the speed at which the generated code runs and sets the results against human performance.
The central finding is that no prompt was best across the board. The winner depended on the task and on whether one looked at the mean or the median. In addition, many differences in the mean came from a few outlier runs. The authors therefore warn that evaluating a prompt with a single metric can hide how much its effectiveness varies.
The comparison with human IOI participants is deliberately cautious. The competitors worked under different conditions, so the comparison is descriptive. Even so, the AI was neither clearly better nor clearly worse: it outperformed the average participant on one task and fell short on the other two.
There are limitations worth keeping in mind. The available material does not specify what the nine styles are, how many times each test was repeated, or exactly which tasks were used. Nor does it indicate that the paper has undergone peer review. The sample is small (one prompt per style, a single model), and the authors themselves limit their conclusions to "these tasks."
For people in Latin America who use language models to train for programming contests or to write performance-sensitive code, the practical lesson is to try several phrasings and measure more than one thing. The 10-page paper states that the code and data are available, so other teams could try to replicate it with other models and more repetitions.
Source: arXiv