Researchers at Mila and the Université de Montréal posted a paper on Sept. 11, 2026, arguing that reinforcement learning, learning by trial and error with a reward for getting it right, does not improve large language models evenly. Those are the kind of AI behind chatbots such as Claude and Gemini, trained on huge amounts of text. The authors, Michael Noukhovitch, Hamish Ivison, Nathan Lambert and Aaron Courville, say this training lifts easy problems far more than hard ones, which bears on how AI developers spend their training budgets.
The authors studied publicly released AI models already trained this way for math, coding, and software work the model carries out on its own. In each case, the gains tracked how good the model was at the start: problems it could already solve improved most, while hard ones barely moved. They call this the Matthew Effect, after the rich-get-richer pattern. They argue that giving every problem the same fixed number of attempts wastes computing power on problems the model already solves and leaves too little to learn from on the ones it cannot yet crack.
Standard training methods give each problem a set number of attempts. On easy problems, most of those attempts teach the model little. On hard problems, the model may get none right and so learns nothing from them. Never Give Up, or NGU, keeps asking the model for new attempts at a problem until one is correct, which the authors say spends fewer attempts on easy problems and more computing power on hard ones. NGU is a proposal in a preprint, not yet an established way to train models.
On Deepscaler, a set of math problems, the authors say NGU got more out of the same computing budget than the standard method, with the biggest gains on the hardest competition problems. On Manufactoria, a coding task where each passed test earns a reward, they say the standard method stalled short of full solutions, while NGU kept clearing harder tests until it solved problems completely. In a Sept. 15 blog post, first author Michael Noukhovitch reports a similar edge on the hardest questions of a separate math test.
The authors warn that NGU is unlikely to help when most training problems are very hard. Retrying in rounds can take longer than asking for more attempts up front, so by the time a full set of attempts is ready, the model in training has often moved further on from the version that made them. No outside group has reported repeating the gains, and the code repository linked from the paper showed a 404 page-not-found error as of Sept. 16, 2026.