plumspringrollsEven research level stuff (let's say getting published in NeurIPS/ICLR/ICML etc.) doesn't need to have a ton of math depending on what you do.
Is this even true lol. For the conferences you mentioned almost all papers are primarily math.
Just looking at the ICLR 2024 outstanding papers I would say only one is math-heavy (generalization of diffusion models). The vision transformers need registers in particular is nice and has basically nothing. Somewhat ironically the more math heavy papers were relegated to the honorable mentions lol.
Obv every paper is gonna require you to latex something I'm not saying that, but in my experience the way you approach a theory/math paper vs the other types of papers is just so different with the math being extremely secondary (means to an end type deal). Compare the protein discovery paper vs the robust agents paper.
Also (and if you're familiar I'm sure you've seen this as well) these conferences are usually flooded with garbage like "we changed some model marginally in pytorch and got some small non-generalizable improvements" and sometimes that stuff gets published so.