In this case, we are not arguing that min_p is better for "creative code" (you really don't want high temperature anywhere near your code generation, despite the "turning up the heat" framing of our paper) - at least in my post claiming min_p is strictly better than top_p above.
We are instead arguing that min_p handles truncating tokens that are more likely to lead to degeneration/looping because it is partially distribution aware. Fully distribution aware samplers like the ones I mentioned above (i.e. P-less decoding) are strictly superior due to using the whole distribution to decide the truncation at every time step.
Code hallucinations, like many LLM hallucinations, can be seen as accumulation of small amounts of "sampling errors".
