What gets a reasoning model stuck on one result?

Hi @lmccall Yeah, this looks like one of those cases where the model gets stuck in a self-correction loop instead of continuing the task. What might be happening is that it latches onto a token, then realizes it’s repeating or conflicting with the instructions. Instead of just moving on, it tries to fix itself mid-generation and those corrections show up in the output. Models on the smaller side (like 5.4-mini) do seem a bit more prone to this. When the prompt is fairly open-ended, they can end up juggling a few competing goals at once, avoiding repetition, staying accurate, and continuing the list and sometimes they don’t fully commit to moving forward, which is where the looping comes from. One thing that often helps is tightening the prompt so there’s less room to backtrack. Asking for a fixed number of items, explicitly saying no duplicates, and telling it not to include corrections or meta commentary can all make a difference. Even something like “don’t revise earlier entries, just continue forward” can help it stay on track. ~Smith