Large language models usually reason by generating intermediate steps as tokens, a process known as chain-of-thought. Yet according to a new paper on arXiv, even the most advanced reasoning models encounter sudden performance drops when tasks require very long sequences of reasoning.
The paper explores implicit reasoning, where the model works through a problem within a single step rather than externalizing each sub-step. The authors propose using curriculum learning—training on progressively harder examples—to teach models this kind of long-horizon implicit reasoning.
The results show that curriculum learning does help, but only up to a point. The study notes a clear limit: beyond a certain difficulty or horizon length, the approach no longer prevents the performance cliffs seen in other methods. The paper does not claim a full solution, but rather highlights both the potential and the ceiling of this training strategy.