←
Maria Esteban15 posts
Maria Esteban avatar

Maria Esteban

@Maria__Esteban

PhD student in Geometric deep learning at @AMLabUvA with @erikjbekkers 🌟

◷ Joined August 2020▧ 4 Photos and videos
Floor Eijkelboom avatar
⟳ Maria Esteban retweeted
Floor Eijkelboom@FEijkelboom·Nov 9

Our survey on continuous diffusion and flow language models is out! ❤️ A little animated tour of what’s inside 👇 From representations and training to few-step generation, scaling, and open problems. continuous-lm.github.io

New survey on Continuous Diffusion and Flow Language Models is out! 🌊💬 One of my first papers, back in early 2025, was on continuous simplex flows. After working on discrete diffusion, I wanted to try controllable sequence generation with flow matching, which naturally operates in continuous spaces. But at the time, I found them much harder to train well than discrete diffusion. 🥲 2026 has changed this picture completely! 🚀 It’s been super exciting to watch continuous LMs take off, but with the rapid advancements, it’s also becoming harder to keep up with all the new work across pre-training, few-step generation, guidance, and scaling 🥹 So together with @chandavidlee, @FEijkelboom, and @jw9730, we put together a survey that consolidates the continuous LM literature and its key insights into 37 pages. It covers everything from the early work before the so-called “winter” of continuous LMs to the wave of new papers in 2026 🌟 More in thread! 🧵

Post attachment
Maria Esteban avatar
Maria Esteban@Maria__Esteban·Nov 4

We introduce a new formulation of drifting models where the kernel's gradient directly determines the direction of motion. This provides theoretical guarantees and naturally extends the framework to data living on curved manifolds.

Alejandro García avatar
⟳ Maria Esteban retweeted
Alejandro García@algarciacast·Jul 21

🎉 Excited to share that RoVE was accepted as a spotlight at @TAGinDS! RoPE made attention scores position-aware but left the value pathway position-blind. RoVE fixes that with a parameter-free tweak, and it turns RoPE attention into an 𝐚𝐭𝐭𝐞𝐧𝐭𝐢𝐯𝐞 𝐜𝐨𝐧𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧🤘

Post attachment
Krishna Balasubramanian avatar
⟳ Maria Esteban retweeted
Krishna Balasubramanian@krizna_b·May 26

🚨 🧵 New paper: Finite-particle rates for drifting models. Drifting trains 1-step generators by moving model μ toward data ν with a conservative or non-conservative velocity field. Consider data on x-axis, model on the y-axis. These 2 fields yield 2 different motions. 1/7

Maria Esteban avatar
Maria Esteban@Maria__Esteban·May 20

2. Natural generalization to non-Euclidean geometries, and discrete data generation by working on the probability simplex 🌎

Post attachment
Maria Esteban avatar
Maria Esteban@Maria__Esteban·May 20

Why is this useful? This new reformulation gives us: 1. Strong theoretical guarantees, including a score-matching interpretation of the drifting dynamics for a broad class of kernels, as well as a Wasserstein gradient flow interpretation of the optimization 🧲