Can a reasoning model fix its own mistakes just because you tell it 'wait, keep thinking'?

This explores whether a tiny test-time nudge, such as appending "Wait" so a model keeps reasoning instead of stopping, can make a model genuinely fix its own errors. The collection has no paper that tests that exact trick head-on. It does have a lot of evidence on the question underneath it: when a model is prompted to reconsider, does the reconsidering actually correct anything? That evidence leans clearly toward no.

Start with what happens when reasoning models reflect on their own. Across eight reasoning models, reflections rarely changed the answer. They mostly confirmed what the model had already decided, so the "second look" works more as a ritual than as a repair Is reflection in reasoning models actually fixing mistakes?. Forcing extra revision can make things worse. In QwQ, R1 and LIMO, most revisions kept wrong answers wrong, smaller models often switched correct answers to incorrect ones, and longer chains with more revisions went with lower accuracy Does self-revision actually improve reasoning in language models?. If a single word buys more thinking time, these results suggest it mostly buys more confirmation, and sometimes it gives the model more chances to talk itself out of a right answer.

The deeper reason is that a model judging its own work isn't a neutral checker. Models over-trust answers they generated themselves, because a high-probability answer also *feels* correct when the same model evaluates it Why do models trust their own generated answers?. That fits the broader finding that without an outside signal telling the model it got something wrong, self-correction lowers reasoning accuracy across GPT-3.5, GPT-4 and Llama-2. Even multi-agent debate turned out to be majority voting under another name Can language models fix their own reasoning mistakes?. A prompt word can't supply the missing ingredient, which is information the model doesn't already have.

So where does real self-correction come from? The corpus points to training, not prompting. Models learn to correct mistakes when they practice on their *own* errors through online RL. Fine-tuning on pre-made correction examples fails because those mistakes don't look like the ones the model actually makes at test time Why does self-correction training on offline data fail?. Scale matters too: a 1-trillion-parameter model discovered self-verification on its own under plain RL, while a 104B model needed hand-designed rewards to get there Does scale alone teach models to reason without hand-crafted rewards?. That suggests a nudge can only draw out a checking habit the model already has.

That last point is what makes the "one word" idea interesting rather than simply wrong. A cheap trigger can work as a key, but only if training has already put the habit behind the door. Related work shows something similar during training: about 1,000 examples of "how to deepen shallow reasoning" can activate latent ability that then supports repeated self-improvement Can models improve themselves on tasks without verifiable answers?. The better question is less "does the word work?" and more "what did the model learn before the word arrived?"