Excited to share CatSBench (#ICLR2026) - a ready-to-use benchmark package for EOT/SB on discrete spaces.
It covers the full benchmarking cycle: data, metrics, ground-truth solutions, and all current solvers + utils.
Paper: https://t.co/w7Usb2d7Kd
GitHub: https://t.co/5MPaAZQLge
Built Tough AI Paper Reviewer from my own research experience: the prompt we use to check our papers before submission.
8 sub-agents check proofs, experiments & novelty against relevant literature. Try it!
For Claude Code/Codex and similar tools 🦉
https://t.co/daJbsyZGTt
Masked diffusion LMs unmask several tokens per step, but sample them independently - so few-step text falls apart.
We introduce E-MoE: the routing decisions of a Mixture-of-Experts backbone become a discrete shared latent. 2× lower generative perplexity at 1–2 steps. 🧵
🔬Can we learn how a population evolves without tracking a single individual? And what if the data is intrinsically discrete?
Introducing Discrete Action Matching (DAM): from population snapshots to stochastic dynamics on a graph.
📄 https://t.co/ifZCGYwY6U
🖼️ Details below 👇
⚡ Want more speed and better quality from your continuous or discrete diffusion LMs? Our new method works across both, scaling up to 16B with DMax! Check out our joint work below👇
Introducing a simple method to post-train both discrete and continuous diffusion LMs.
It's based on Maximum Mean Discrepancy (MMD) that matches token-level feature distributions. Our method improves generation and speed on OWT/GSM8K and scales to 16B models. 🧵
We introduce Inverse Distillation Unlearning (IDU): distilling a multi-step flow/diffusion teacher into a one-step generator that suppresses unwanted outputs while preserving retained-data quality: https://t.co/E5gW02kClQ
A short teaser below. Details in the next post.
Is factorization itself the problem for few-step discrete diffusion on downstream tasks?
AlphaDLM keeps factorization but changes the loss: 34.6% on GSM8K across 4 model evaluations, and 7.66% in one!
Our new work on how factorization and the training objective interact 🧵
Continuous diffusion for text is taking off.
But most methods still diffuse in spaces that were never learned for denoising and this limits their performance.
In LDLM, we train the encoder, diffusion model, and decoder jointly, so the denoising objective shapes the representation directly.
❌ Not: learn a space → run diffusion on it.
✅ Instead: learn a diffusable space.
On OpenWebText & LM1B, LDLM achieves a better quality–speed trade-off than MDLM, Duo, CANDI, and FLM.
Up to 13× faster sampling.
🧵 Recipe and insights below ↓
Latent diffusion for text.
Better quality? Faster sampling?
Seen one? Didn't think so.
Meet LDLM — beats MDLM, Duo, CANDI, FLM on OpenWebText & LM1B.
2–13× faster.
The trick is:
Don't run diffusion in a latent space.
LEARN a latent space FOR diffusion.
Paper: https://t.co/LYgkxTWJuE
Furthermore, we hope the GitHub repo can serve as a starting point for future work on EOT/SB in discrete spaces. It provides reusable code and configs for existing generative solvers (DLightSB, DLightSB-M, CSBM, alpha-CSBM) and can be adapted as a template for new solvers.
Excited to share CatSBench (#ICLR2026) - a ready-to-use benchmark package for EOT/SB on discrete spaces.
It covers the full benchmarking cycle: data, metrics, ground-truth solutions, and all current solvers + utils.
Paper: https://t.co/w7Usb2d7Kd
GitHub: https://t.co/5MPaAZQLge
On the evaluation side, CatSBench provides TorchMetrics-based metrics, including Trend and Shape scores. To keep the benchmark efficient in practice, we also implemented a Triton-based logsumexp operation for key internal computations.