When an AI thinks for pages on end, why does it loop, wander, and pile on steps that don't help?

This explores why models that produce very long reasoning traces end up circling, drifting, or piling on steps that don't help. The corpus has no note that measures repetition loops directly, but several notes point to the same cause. Long traces are a habit the model was trained into and never taught to stop, and they are not a sustained line of thought that gets diluted over time.

The closest match to "loops" is Why do reasoning models abandon promising solution paths?. It describes two failures that feed each other. Models wander, meaning they explore invalid paths. They also underthink, switching away from a promising path too early. The paper's point is that the problem is disorganization, not too little compute. A penalty on thought-switching at decoding time improves accuracy without any retraining. That suggests the good path was often already in the trace and got abandoned. A model that drops a path and later restarts it looks like a loop from outside.

The second cause is that nothing in training teaches the model when to stop. Why do reasoning models overthink ill-posed questions? finds that reasoning models write long, redundant answers to questions with missing premises. Non-reasoning models correctly say the question can't be answered. Training rewards producing reasoning steps and never teaches disengaging. Does every correct chain-of-thought trace improve fine-tuning? shows the habit is also in the training data. Traces that keep exploring after the answer is already settled hurt fine-tuning even when the answer is right. Cutting only that tail helps more than cutting an equally long random chunk, so the harm comes from the pointless exploration and not from the length.

Third, a long trace is not a sign of hard thinking. Does longer reasoning actually mean harder problems? finds that trace length tracks difficulty only on problems close to the training data, and decouples from difficulty entirely outside it. Do language models fail at reasoning due to complexity or novelty? locates the breaking point in how unfamiliar the specific instance is, not in how complex the task is. Can reasoning models actually sustain long-chain reflection? shows the result. Frontier models manage only 20–23% on puzzles that require real backtracking, despite fluent "let me reconsider" language. When a problem is unfamiliar, the model keeps producing the form of deliberation without the ability to make progress. What makes chain-of-thought reasoning fail in language models? reaches a similar verdict, that format matters more than content.

Distraction is less harmful than you might expect, because much of the clutter is inert. Can models think longer by forgetting intermediate reasoning? finds that most intermediate tokens quickly stop mattering. Keeping only the instruction and a recent window gives a 3x speedup and lets traces run past 100k tokens. Do reasoning traces actually show how models think? adds that most erroneous steps in a trace never influence the final answer. The model largely ignores its own sprawl, and the cost is wasted tokens. The exception is input noise. Can longer reasoning chains eliminate model sensitivity to input noise? proves that longer chains reduce sensitivity to perturbed input but never remove it. What the corpus lacks is a direct account of repetition loops, such as degenerate cycles or attention dilution over huge contexts. The causes above come from training incentives and unfamiliarity, and none of these notes tests a mechanism inside long contexts.