Gilad (@giladturok) / X

Gilad

1,357 posts

Gilad

@giladturok

CS PhD student @cornell | diffusion language models | Prev: @Uber, @FlatironInst, @Columbia

NYC 🗽🇺🇸

Joined November 2023

Pinned
Gilad
@giladturok
Mar 12
1/ 🚨 New paper! DUEL: Exact Likelihood for Masked Diffusion via Deterministic Unmasking We give masked diffusion models (MDMs) proper likelihood — and therefore proper perplexity — for the first time. Turns out MDMs are closer to autoregressive models than previously thought.
23K
Gilad
@giladturok
Oct 27, 2024
Amazingly *clear* monograph on Monte Carlo & sampling: arxiv.org/pdf/2405.16359. Highly recommend skimming if you want to learn these topics!
74K
Gilad
@giladturok
Jun 29, 2025
sometimes ppl forget: flow matching and diffusion are mathematically equivalent!
44K
Gilad
@giladturok
Aug 14, 2024
Cool blog post on convergence of the Unadjusted Langevin Algorithm (ULA): fa.bianp.net/blog/2023/ulaq/ If you're intimidated by complex convergence proofs, this one is excellently written! ULA underpins diffusion models and is related to the Markov chain Monte Carlo literature.
13K
Gilad
@giladturok
Nov 21, 2024
What are SOTA algorithms for sampling from *discrete* distributions? I already know about Gibbs and Gibbs with Gradients
16K
Gilad
@giladturok
Nov 8, 2024
"we don't know what to do with those chains yet, but I have the gut feeling that once we do, probabilistic programming will have another breakthrough" - @remilouf on parallelizing 100s of MCMC chains with JAX
12K
Gilad
@giladturok
Dec 17, 2024
Thinkin bout normalizing flow to generate samples from probability distributions. What methods meet this criteria: 1. learn entire trajectory, not just final distribution p(x) 2. unbiased 3. doesn't back-propogate through ODE solver LFIS (link below) is the only one I found
8.8K
Gilad
@giladturok
May 24, 2024
Replying to @AnthropicAI
Golden Gate Claude has a big secret. It might struggle on MMLU, but it excels on our new challenging benchmarks: - What is the best bridge (WITBB) - What precious metal is coolest (WPMIC) - Why do fences suck (EDFS)
7.2K
Gilad
@giladturok
Jul 14, 2024
Had a great time presenting a poster at @ISBA_events on dynamic step sizes for Markov chain Monte Carlo methods. Joint work with Bob Carpenter and Chirag Modi.
2.9K
Gilad
@giladturok
Jun 20, 2025
Replying to @giffmana
I love Jax so freakin much it’s a shame that no one uses it in production besides Google Guess I’ll have to get a job at Google one day to make my Jax dreams come true 🤷‍♂️😎
4.1K
Gilad
@giladturok
Sep 11, 2024
Saying something *actually helpful* about Scientific Communication is pretty hard. Most articles out there don't actually tell you anything useful. The one exception is the amazing presentation from Trevor Campbell. Here's a summary, a short🧵:
docs.google.com
How to Explain Things
How to Explain Things Guidelines for Effective Scientific Communication Trevor Campbell Associate Professor UBC Statistics 1
2.6K
Gilad
@giladturok
Jul 28, 2024
I'm currently in my JAX era: - Write my own JAX projects for ML + stats - Do serious deep learning with {model, tensor, data} parallelism - Write JAX-style code (functional) with grad, vmap, pmap, jit - Understand why JAX is fast (jit, SDMP, lax) - PyTorch vs JAX trade-offs
10K
Gilad
@giladturok
May 1, 2025
🚨Exited to present my 1st paper at @aistats_conf in Thailand! Joint work with Chirag Modi + Bob Carpenter We develop an MCMC method that *efficiently* adapts step size for multiscale (high curvature) probability distributions: arxiv.org/abs/2406.02741 DM me to chat at AIStats!
arxiv.org
Sampling From Multiscale Densities With Delayed Rejection...
Hamiltonian Monte Carlo (HMC) is the mainstay of applied Bayesian inference for differentiable models. However, HMC still struggles to sample from hierarchical models that induce densities with...
2.2K
Gilad
@giladturok
Jul 29, 2024
Replying to @jeremyphoward @Railway and 2 others
Seems amazing but also
4K