@JChemseddine

PhD Student @ TU Berlin

 Berlin, Germany
jchemseddine.github.io
Joined August 2023
  • Tweets23
  • Following47
  • Followers51
6Media
This Monday, Oct 26, Gregor Kornhardt (@gregorkornhardt) will present his paper: SOL: Measuring Gaps between Text Distributions by Double Sliced Wasserstein Metrics
2
5
1
23
1,857
Jannis Chemseddine retweeted
Justin Deschenaux@jdeschenaOct 21
Nice to see that the spherical geometry xitter.cf/jdeschena/status/20546… can also be beneficial for fixed-point reasoners!
Arber Zela@ZelaArberOct 20
A looped transformer has to know when to stop, and its state has to stay bounded while it loops. We introduce Spherical Fixed-Point Reasoners (SFPRM) 🌐: loop on the unit sphere, where each step is an angle, and stop once the angle is small enough. 🧵 (1/9)
7
30
1,441
Jannis Chemseddine retweeted
Oscar Davis
@osclsdOct 14
Continuous Diffusion LMs went from niche to contender this year, but the recipe is shaky: why self-conditioning? How to embed tokens? 🤨 Our new work takes your one-hot flow from 4% to 50% on TinyGSM without self-conditioning, and explores the design space of CDLMs. 🔥🚀 (1/6)
13
54
6
211
17,733
Jannis Chemseddine retweeted
Mo Piening@moritz_pieningOct 9
Equivariance matters not only for networks, but also for interpolation in flow matching With Christian Wald, we show how graph alignment preserves geometry and improves few-step generation Preprint: arxiv.org/abs/2608.26961 Fig: Alignment preserves sparse graph structure (b, c)
2
3
197
Evaluating diffusion language models can be tricky… alongside reporting gen. PPL and entropy we suggest you try SOL 🌞
Gregor Kornhardt@gregorkornhardtOct 8
Recently there has been a lot of discussion about the difficulties of evaluating diffusion language models. We instead introduce: SOL, a sample based distance between generated and real text distributions. 🧵
3
1
11
707
Jannis Chemseddine retweeted
Max Zimmer@maxzimmerberlinOct 6
A bit late to the party, but two papers have been accepted to #NeurIPS2026! - A Free Lunch in LLM Compression: Revisiting Retraining after Pruning (arxiv.org/abs/2510.14444, camera ready update coming!) - A Trust Region Approach for Learning Schrödinger Bridges (soon on arXiv)

A Free Lunch in LLM Compression: Revisiting Retraining after Pruning

Post-training pruning can substantially reduce LLM inference costs, but it often degrades quality unless the remaining weights are adapted. Since global retraining is expensive at LLM scale,...

arxiv.org
1
1
7
250
Jannis Chemseddine retweeted
Justin Deschenaux@jdeschenaSep 29
🔺Happy to share a project done during my internship🔻 It was amazing to collaborate 🥹 @ValentinDeBort1 @ArnaudDoucet1 @agalashov @AndrewC_ML @JamesTThorn and Kevin Li (who I cannot find on Twitter)
Valentin De Bortoli@ValentinDeBort1Sep 29
🔥Welcome 🔺Simplex Diffusion Models🔻! Standard discrete diffusion models suffer from "information collapse" because they discard uncertainty at intermediate steps by sampling categorical tokens. (1/5)
4
27
86
6,254
Happy to share that our paper was accepted to #NeurIPS2026 as a spotlight! 🎉
Jannis Chemseddine@JChemseddineAug 12
A while ago we presented an early version of our work on spherical flows for categorical data at @diffusion_llms. The paper has grown a lot since, so it felt like time for a thread...
2
11
28
2,136
A while ago we presented an early version of our work on spherical flows for categorical data at @diffusion_llms. The paper has grown a lot since, so it felt like time for a thread...
2
6
1
23
3,843
more replies
We have also uploaded the code and checkpoints here: github.com/JChemseddine/SFCD

GitHub - JChemseddine/SFCD: Spherical Flows for Sampling Categorical Data

Spherical Flows for Sampling Categorical Data. Contribute to JChemseddine/SFCD development by creating an account on GitHub.

github.com
1
1
183
For more in this direction, also see the work on hyperspherical flows by @jdeschena and @caglarml (arxiv.org/abs/2605.11125)!

Language Modeling with Hyperspherical Flows

Discrete Diffusion Language Models progressed rapidly as an alternative to autoregressive (AR) models, motivated by their parallel generation abilities. However, for tractability, discrete...

arxiv.org
2
3
289
Load more