BYOD Turns Any Instruction-Tuned LLM Into a Diffusion Model in 4 Hours

BYOD, short for Build Your Own DLM, converts existing instruction-tuned autoregressive models into masked diffusion language models by training only LoRA adapters. The method freezes the parent model, uses no more than 102.4 million training tokens, and completes each reported conversion in 3.1 to 4.0 hours on one GPU.

The approach sharply reduces the compute required to study diffusion-based text generation. Purpose-built systems such as LLaDA train billions of parameters from scratch; LLaDA used 2.3 trillion pretraining tokens and 4.5 million instruction pairs. BYOD preserves the knowledge in an existing model and trains adapters to support bidirectional attention and iterative denoising.

From left-to-right generation to denoising

Autoregressive language models predict one token at a time from left to right. Masked diffusion models begin with a fully or partially masked sequence, predict missing tokens using context from both directions, and refine the sequence over multiple model passes. The process supports infilling and parallel token updates because several positions can change during each pass.

Bidirectional context can also help with tasks that expose the reversal curse, where a model trained on “A is B” fails to recover “B is A,” and with constrained generation tasks such as reconstructing a poem in reverse order. These capabilities come with a compute trade-off: each denoising step evaluates the sequence again, so parallel updates do not automatically produce lower latency.

A frozen backbone and large LoRA adapters

BYOD replaces the causal attention mask with bidirectional attention and trains on a masked-diffusion objective. Random spans in the target sequence are hidden, and the model learns to recover them from the visible context.