Still, we can look for telltale signs. Another symptom of memorization is that GPT is highly sensitive to the phrasing of the question. Melanie Mitchell gives an example of an MBA test question where changing some details in a way that wouldn’t fool a person is enough to fool ChatGPT (running GPT-3.5). A more elaborate experiment along these lines would be valuable.

In fact, we can definitively show that it has memorized problems in its training set: when prompted with the title of a Codeforces problem, GPT-4 includes a link to the exact contest where the problem appears (and the round number is almost correct: it is off by one). Note that GPT-4 cannot access the Internet, so memorization is the only explanation.

To benchmark GPT-4’s coding ability, OpenAI evaluated it on problems from Codeforces, a website that hosts coding competitions. Surprisingly, Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10 recent problems in the easy category. The training data cutoff for GPT-4 is September 2021. This strongly suggests that the model is able to memorize solutions from its training set — or at least partly memorize them, enough that it can fill in what it can’t recall.