masked-diffusion

Embedding Inversion via Conditional Masked Diffusion: recover original text from embedding vectors using parallel denoising. Live demo + training pipeline + technical report.

Pytorch Implementation of MD4: Simplified and Generalized Masked Diffusion for Discrete Data

The simplest masked-diffusion language model you can actually train, debug, and learn from — ~1100 lines of plain PyTorch, char-level, with an honest head-to-head against a matched autoregressive baseline. Watch text crystallize from noise.

[EMNLP 2026] Length-Adaptive Decoding for Masked Diffusion Machine Translation — a training-free target-length selector for masked diffusion LMs.

A compact, from-scratch trainer for masked diffusion and GPT-style autoregressive language models in one shared PyTorch Transformer.

🛠 Reconstruct original text from text embeddings using conditional masked diffusion to reveal reversible embedding representations efficiently and accurately

Official implementation of Attention-Discounted Adaptive Sampler (ADAS) for masked diffusion language models.

Minimal masked diffusion LM in PyTorch. Pretraining + SFT + a scaling family from 50M to 350M, LLaDA recipe.

DiffLM Lab — Interactive Diffusion Language Model Laboratory. 6 modules: denoising playground, forward absorbing process, sampling strategies, AR-vs-diffusion race, block diffusion, masked-diffusion ELBO. Real in-browser denoiser.

What parallel commitment costs a masked diffusion LM: a 4-page note, measured on a task with a decidable verifier. Produced by the Rust stack (candle-mi -> canvas).

Masked diffusion language model (201M params, RTX 3090) — parallel token generation validated by an AR oracle, 1.8-3.9x faster forward pass than autoregressive. Weights on HuggingFace.

Steer masked diffusion LLMs toward a topic without mentioning it in the prompt — energy fields injected at each denoising step

PyTorch implementation of "Train for the Worst, Plan for the Best." Investigating adaptive token ordering in Masked Diffusion Models (MDMs) to sidestep hard subproblems and elicit reasoning in discrete domains.

[ECCV2026] GryphOne: Symbol-Aware Masked Diffusion for Structural Refinement in Offline Handwritten Mathematical Expression Recognition

Empirical study: logit-level energy guidance on LLaDA-8B masked diffusion (13 experiments, +39% semantic steering). Extends Diffusion-LM classifier guidance to 8B scale, gradient-free.