Like, it's clearly a thing in the culture to "implement algo X, from scratch in numpy / jax etc !" .. because a popular algorithm is so obscurely presented in papers (with a lot of mathiness and information theoretic principles yada yada) that this needs to be a thing

I'm afraid there is often also an element of writing to create a tribe and looking smart, instead of writing for clarity / being understood etc.

It is a very unfortunate trait of much of ML lit, that algorithmic details and implementations are hidden in badly written supplementaries, or much worse in large codebases. It can be very different to read an algo proposed in siam journal or sigproc. vs in ML.

Harsh. I choose to work with person X or Y because I really like their topic !

I didn't even mention the concept of "compiti delle vacanze", holiday homework ...

If the reviewer / meta-reviewer does a bad job / is not fit for the paper, there is no structure that will fix it of course.
And, the people reviewing will have a significant overlap (I guess) from NeurIPS/ICML/etc..
Still, worth trying to have (healthier?) incentives/structure as in TMLR

and forces the reviewer to focus on requested changes.
Of course, I also like TMLR because of its spirit, which for me is about correctness and details rather than significance or importance, which are typically in the eyes of the (powerful) beholder.

24. arxiv.org/abs/2510.00389
'Zero variance self-normalized importance sampling via estimating equations'
- Art B. Owen

Even with optimal proposals, achieving zero variance with SNIS-type estimators requires some innovative thinking. This work explains how an optimisation formulation can apply.

were you reading:
epubs.siam.org/doi/abs/10.1...

"Conditional Causal Discovery"

(don't be fooled by the title :D )

openreview.net/forum?id=6IY...

"Estimating the Probabilities of Rare Outputs in Language Models"

arxiv.org/abs/2410.13211

"Stochastic Optimization with Optimal Importance Sampling"

arxiv.org/abs/2504.03560