Subham Sahoo @ssahoo_

@iScienceLuvr Lol, same! I failed mine twice and got it on my third attempt only recently

2 hours into the week. fuck.

📢June 29 (Mon): Nemotron-Labs-Diffusion: Unifying AR, Diffusion, and Self-Speculation 🌟Nemotron-Labs-Diffusion is a tri-mode language model (LM) that unifies AR, diffusion, and self-speculation decoding within a single architecture. 💪Trained with a joint AR-diffusion objective, Nemotron-Labs-Diffusion can switch modes to sustain high throughput across deployment settings and concurrency levels. Their study shows that 1️⃣AR and diffusion objectives are complementary: diffusion improves lookahead planning, while AR provides left-to-right linguistic priors. 2️⃣In self-speculation mode, diffusion drafts while AR verifies, outperforming multi-token prediction (MTP) methods in both acceptance rate and real-device efficiency. 3️⃣A speed-of-light analysis further demonstrates diffusion’s long-term potential, with up to 76.5% more tokens per forward pass than self-speculation under an optimal sampler. 📈Scaling to 3B, 8B, and 14B parameters, the Nemotron-Labs-Diffusion family, including base, instruct, and vision-language models, consistently outperforms state-of-the-art open-source AR and diffusion LMs in both accuracy and speed. ⚡For example, Nemotron-Labs-Diffusion-8B decodes 6× more tokens per forward than Qwen3-8B with comparable accuracy, translating to 4× higher throughput on SPEED-Bench with SGLang on a GB200 GPU. This Monday, Yonggan Fu (@YongganFu) from NVIDIA Research will present Nemotron-Labs-Diffusion: Unifying AR, Diffusion, and Self-Speculation.

Join us for three days of demos, talks, and surprises at ICML 2026 in Seoul! The IFM team is bringing an interactive experience with live K2‑series demos and Pan‑series world model previews to the expo floor. The Institute of Foundation Models is a global research lab, bringing our work to Seoul from our labs in Silicon Valley, Paris, and Abu Dhabi. Want to work with us? Scan the QR code to explore career opportunities. See you at booth B402! #IFM #ICML2026 #WorldModels #FoundationModels

@jxmnop @srush_nlp Normies like me would say “graduated with a PhD” but you do you

It’s a great team!

Engram @EngramLab

6 days ago

x.com/i/article/2069…

170 220 2K 1.7M 1K

Now, imagine diffu..🤫🤐

Sarah Chieng @MilksandMatcha

a week ago

what if you could instantly clone any object? Gemma 4 is now on @cerebras Inference, running up to 10x faster than GPUs (1,500 tokens/sec). Multimodal generations you can iterate on in real time :)

22 16 271 32K 130

“Indian Software Engineer” 👹 intheweights.com

A technical dive inside our new "Midjourney Scanner"

@linqi_zhou I didn’t realize until now that you are at OpenAI, congrats to you too!

@juliarturc 😂😂

📢 June 15 (Mon): Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation 🤔 Discrete diffusion models are often trained through clean-data prediction, but the prediction can be used in different ways to define the reverse dynamics. In Masked Diffusion Models (MDM) these choices largely coincide, whereas in Uniform Diffusion Models (UDM) they do not. 💡 The authors show that the standard plug-in bridge parameterization for UDM is not optimized by the denoising posterior, but by a leave-one-out posterior that predicts each clean token without using its own noisy observation. This identifies a mismatch between the plug-in ELBO and the usual cross-entropy denoising objective. 🔧 The authors characterize the leave-one-out target and derive exact conversions between the denoiser, the leave-one-out posterior, and the score. These conversions allow them to disentangle parameterization and the training objective. 📈 Their results also lead to inference improvements without any additional training through an informed predictor-corrector sampler and improved temperature sampling based on the leave-one-out predictor. 🔧 The authors further introduce an absorbing-state reformulation of uniform diffusion that preserves the UDM joint law while decomposing it into masked-diffusion-like sampling operations, with simpler denoising posteriors, carry-over unmasking, and a natural remasking mechanism. 📈 On language modeling, leave-one-out parameterizations consistently improve UDM generation, while the absorbing construction matches or surpasses masked diffusion. These results suggest that the empirical gap between masked and uniform diffusion is driven less by the choice of marginals themselves than by parameterization and sampling design. This Monday, Samson Gourevitch (@samsongvch, samsongourevitch.github.io), Yazid Janati (@yjelid, yazidjanati.github.io), and Dario Shariatian (@dario_sha, darioshar.github.io) will present their paper "Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation".

Great work by @samsongvch @yjelid @dario_sha on Uniform-state Diffusion! Join us on Monday at 10 am PT

Discrete Diffusion Reading Group @diffusion_llms

2 weeks ago

📢 June 15 (Mon): Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation 🤔 Discrete diffusion models are often trained through clean-data prediction, but the prediction can be used in different ways to define the reverse dynamics. In Masked

1 6 27 11K 11

You have to be humble even when pursuing excellence. I think the arrogance with which Anthropic has pursued the latest release has universally landed poorly.

@beffjezos @snwy_me 😂😂

After DiffusionGemma dropped, views spiked on our @diffusion_llms Reading Group video: “The Diffusion Duality.” Video: youtu.be/FCO-nnqHOqQ?si… Join the Discord + mailing list: d-llms.com The diffusion duality: s-sahoo.com/duo/

Google Gemma @googlegemma

3 weeks ago

Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇

166 809 5K 959K 2K

@akashmjn Ofc 😊

Want to understand what powers DiffusionGemma? Start with these two tutorials on its core building blocks: 1. Uniform-state Diffusion: youtu.be/FCO-nnqHOqQ?si… 2. Block diffusion: iclr.cc/virtual/2025/o…

Google Gemma @googlegemma

3 weeks ago

Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇

166 809 5K 959K 2K

Curious what’s under the hood of DiffusionGemma? Begin with these two tutorials covering its foundational building blocks: 1. Uniform-state Diffusion: youtu.be/FCO-nnqHOqQ?si… 2. Block diffusion: iclr.cc/virtual/2025/o…