The latter. A reasoning model has been finetuned to use the scratchpad for intermediate results (which works better than just prompting a model to do the same).

I'd expect the same (fine tuning to be better than mere prompting) for most anything.

So a model is or is not "a reasoning model" according to the extent of a fine tune.

Are there specific benchmarks that compare models vs themselves with and without scratchpads? High with:without ratios being reasonier models?

Curious also how much a generalist model's one-shot responses degrade with reasoning post-training.

> Are there specific benchmarks that compare models vs themselves with and without scratchpads?

Yep, it's pretty common for many models to release an instruction-tuned and thinking-tuned model and then bench them against each other. For instance, if you scroll down to "Pure text performance" there's a comparison of these two Qwen models' performance: https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking

Thanks for the Qwen tip. Interesting how much of a difference reasoning makes for coding.

> Are there specific benchmarks that compare models vs themselves with and without scratchpads? High with:without ratios being reasonier models?

Yes, simplest example: https://www.anthropic.com/engineering/claude-think-tool

The question is: fine-tuning for what? Reasoning is not a particular task, it is a general-purpose technique for directing more compute at any task.

Pivot tokens like 'wait', 'actually' and 'alternatively' are boosted in order to force the model to explore alternate solutions.

Any model that does thinking inside <think></think> style tokens before it answers.

This can be done with finetuning/RL using an existing pre-formatted dataset, or format based RL where the model is rewarded for both answering correct and using the right format.