Congratulations! 🥳

That @void.comind.network sticker is awesome!

You‘re in luck, we offer a PyTorch implementation!

The paper, poster and code in @julialang.org and PyTorch can be found here:
neurips.cc/virtual/2025...

Joint work with Neal McKee, Johannes Maeß, Stefan Blücher and Klaus-Robert Müller, @bifold.berlin.

I'm excited to see whether our idea translates to general MC integration over Jacobians and gradients outside of XAI. Please don't hesitate to talk to us if you have ideas for applications!

Our proposed SmoothDiff method (see first post) offers a bias-variance tradeoff: By neglecting cross-covariances, both sample efficiency and computational speed are improved over naive Monte Carlo integration.

To reduce the influence of white noise, we want to apply a Gaussian convolution (in feature space) as a low-pass filter.
Unfortunately, this convolution is computationally infeasible in high dimensions. Naive Monte Carlo approximation results in the popular SmoothGrad method.

Input feature attributions are a popular tool to explain ML models, increasing their trustworthiness. In this work, we are interested in the gradient of a given output w.r.t. its input features.
Unfortunately, gradients of deep NN resemble white noise, rendering them uninformative:

An even scarier thought is that similar systems of checks and balances are likely silently failing across all of society. Academia is just uniquely transparent, making it look like patient zero.

Technical issues aside, it seems that the signal-to-noise ratio on both sides of the review process has reached a critical low point. And the academic community hasn’t yet figured out how to deal with this.

Judging by the code, I agree.
However, if I had to create these diagrams from scratch, I think I would prioritize Typst's error messages and almost instant compilation over TikZ' more concise syntax.

There is also a TikZ-like package called CeTZ, but I haven't found an excuse to try it yet:
github.com/cetz-package...

Yeah, that's fair. Some new formatting tweaks have recently been made available (e.g. bsky.app/profile/adri...), but vertical justification (especially with inline maths) still visibly lags behind TeX. I found this blog post by the Typst team to be both insightful and promising:

Which features do you feel are missing?

Wow, I had no idea that it had been around for so long! The documentation is excellent and a joy to read. Thank you for your efforts!

Made in Luxor.jl, of course!

The result of my first attempt:

Adding this to your document should get you started:

#set par(justify: true, justification-limits: (
spacing: (min: 66.67% + 0pt, max: 150% + 0pt),
tracking: (min: -0.01em, max: 0.02em),
))