Yes - impressive how good the small models are getting, and this "reasoning distillation" seems to have given them a significant boost.
Even though humor is largely about the unanticipated punchline, I'd have guessed (maybe wrongly) that there'd be enough analytical discussion of humor in the training set for a reasoning model to come up with a much more plausible attempt at a formulaic type of joke.
From the example given it seems there's too much "thought" put into "what do I have to work with here", and not enough into conceiving/selecting a template for the joke. Maybe part of the problem is that the LLM doesn't realize that, being an LLM, it's best chance at being funny to a human is to closely stick to a formula that humans find funny, and not try to be too smart in trying to deconstruct it.
Even though humor is largely about the unanticipated punchline, I'd have guessed (maybe wrongly) that there'd be enough analytical discussion of humor in the training set for a reasoning model to come up with a much more plausible attempt at a formulaic type of joke.
From the example given it seems there's too much "thought" put into "what do I have to work with here", and not enough into conceiving/selecting a template for the joke. Maybe part of the problem is that the LLM doesn't realize that, being an LLM, it's best chance at being funny to a human is to closely stick to a formula that humans find funny, and not try to be too smart in trying to deconstruct it.