Nesta Midavaine

AI Engineer @ Prosus. MSc AI graduate from UVA. Interested in ML research, Diffusion LMs, and Post-training.

[Christian A. Naesseth @canaesseth]

During a MSc project back in early 2025 @N_Midavaine did some great work with @GrigoryBartosh in this space where we learn noising paths and embeddings jointly. Later published at a #EurIPS2025 workshop:

Towards Latent Diffusion Suitable For Text
https://t.co/QuqgpMU68D

[Alex Choi @AlexChhk]

i only just discovered there is stack overflow for agents tf

[Nesta Midavaine @N_Midavaine]

This is the first paper from the Large Commerce Model work at Prosus: pretrain on commerce behaviour once, specialise into small task models that are cheap to serve. Presenting Thursday in Naples. Come say hi.

[Nesta Midavaine @N_Midavaine]

We taught a 4B LLM to hand out food-delivery discounts. One month in production at iFood: +1% conversion, −6% promotion cost vs the existing system. Paper accepted at ECML PKDD 2026 (Industry Track). How it works 🧵

[Nesta Midavaine @N_Midavaine]

Why it beat the baseline: the old policy forecast orders per segment × discount cell, and every new cell needed exploration traffic. 50 segments × 10 discounts was the ceiling. With context in the prompt, the same weights price all promotion combinations. Twice a day, in 30 mins.

[Birchlabs @Birchlabs]

how was your day and why are you still using pytorch

[Sander Dieleman @sedielem]

My first blog post in over a year is a deep dive on flow maps🗺️, or how to learn the integral of a diffusion model to enable faster sampling and several other cool tricks.

It's the longest one yet👀 Let me know what you think!

https://t.co/O8bBGZ9qjC

[Sander Dieleman @sedielem]

Sorted! @FEijkelboom and @GrigoryBartosh will be your #ICLR2026 diffusion circle organisers 🙏 stay tuned!

[Nesta Midavaine @N_Midavaine]

Result: 1.1163 bpb, -0.0031 over last Friday's top score. All under ~16MB with 10 minutes of eval time. Code + writeup: https://t.co/hQpLY1M6lr

More information coming later in a @Prosusgroup
tech blog.

[Nesta Midavaine @N_Midavaine]

I brought depth recurrence back in @OpenAI's #ParameterGolf challenge.

Last week an analysis showed why recurrence doesn't work under the 16MB/10min constraints — compounding quantization errors + fewer training steps. +0.025 bpb worse.

Two ideas to fix it: 🧵

[Nesta Midavaine @N_Midavaine]

Second fix: a cheap Jacobian proxy loss that keeps hidden states from exploding across passes. I used a finite-difference approximation, and measured the impact by tracking the growth ratio across several depths.

I managed to keep it stable even up to 6x and 8x recurrence.

[Discrete Diffusion Reading Group @diffusion_llms]

📅 Weekly meetings on Mondays starting November 17, 10–11 AM ET (4–5 PM CET).

Details about our first session are coming soon! 🚀

[Nesta Midavaine @N_Midavaine]

@tenderizzation Improves efficiency on rotated mnist

[Diego Canez @dgcnz]

llm assistant for legal research? maybe… 🔜

[The ML Reproducibility Challenge @repro_challenge]

#MLRC2025 happening this week (August 21st)! The second keynote talk [2/4] will be from @soumithchintala at 10:30AM EST

[Diego Canez @dgcnz]

Adding nvtx.push/pop calls just to profile GPU code is such unnecessary friction

Built a VSCode extension that lets you mark profiling ranges visually, no more modifying your source code

🧵
[1/4]

[Nesta Midavaine @N_Midavaine]

This thread mainly contains vague posting, I will do a better write-up later on. Also, ignore the watermarks please...

[Nesta Midavaine @N_Midavaine]

Diffusion Language Models have the potential to break the AR LLM paradigm. Models like Gemini Diffusion have already shown incredible token/s.

I just finished the last couple months of work on Latent Diffusion Models for language. We found competitive likelihood results and reasonable sample quality for latent diffusion. This example shows 32-step generation of 64 tokens.

Also, if you mess with the way you destroy information, you can do fun stuff like autoregressive latent diffusion (see 🧵),

[Nesta Midavaine @N_Midavaine]

Autoregressive Latent Diffusion. Generation derails around 30 tokens. I do not recommend doing this in practice.
However, it is interesting that you can train a diffusion model to generate autoregressively.