llada

[ECCV 2026] A diffusion-based framework for document OCR that replaces autoregressive decoding with block-level parallel diffusion decoding.

LLaDA implementation

OpenAI-compatible HTTP serving for diffusion language models. Continuous batching + LocalLeap acceleration.

GUI for LLaDA Diffusion LLM with Quantization for low end GPU and CPU options. Now with Prototype Training and Vector DB.

Minimal quickstart fork of the LLaDA repository. Watch text diffusion with a single line of code (& flag).

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

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

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

This project is an implementation of a character-level diffusion model (LLaDA) for text generation, based on the principles outlined in the take-home exercise. It also includes a standard autoregressive Transformer model as a baseline for comparison.

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.

LLaDA based ASR

Experiments with diffusion LLMs on Apple Silicon

Official Repository for The Paper, Optimizing Denoising Trajectories in dLLMs: A Lightweight Evolutionary Heuristic Approach

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

Formal study proving Discrete Latent Diffusion Models (LLaDA 1.5 + Path-Guided Unmasking) outperform autoregressive models on complex cross-file code refactoring tasks requiring global structural integrity.

Research project benchmarking diffusion language models vs autoregressive transformers on IMDB, focusing on minimum denoising steps for quality-latency parity.

PyTorch Lightning implementation of LLaDA, a masked-diffusion language model that generates text without autoregression. Includes training, greedy/multinomial sampling, and a generation demo.

Empirical study of logit-level guidance on LLaDA-8B masked diffusion model. Extends Diffusion-LM (Li & Liang 2022) classifier guidance to 8B scale with gradient-free energy vectors. 13 experiments, +39% semantic steering.