BYOD Converts Four Open Models to Diffusion With LoRA Alone

Researchers have presented a LoRA-only method for turning existing autoregressive language models into diffusion language models, without updating the base models. The work matters because diffusion generation offers a different way to produce text, and a cheaper conversion route could make that approach available to far more model families.

According to the paper posted to arXiv under Computation and Language by Ruurd J.A. Kuiper and Robert A. Bagheri, the method is called Build Your Own DLM, or BYOD. The authors describe it as a generalizable recipe for parameter-efficient conversion from autoregressive, or AR, models to diffusion language models, or DLMs.

What is BYOD and what did the researchers do?

Most large language models in use today are autoregressive. They generate text token by token, with each new token conditioned on the tokens that came before it. That sequential process is familiar and effective, but it constrains how generation can be controlled and how context can be used.

Diffusion language models take a different approach. According to the paper, DLMs permit parallel token updates, flexible generation, and bidirectional attention. In practical terms, a diffusion model can update many tokens at once rather than committing to them strictly left to right, and it can use context from both directions when it generates.

The trade-off, as the authors note, is that leading diffusion systems have relied on costly pretraining or full-model fine-tuning. That makes diffusion generation expensive to adopt if a team already has a strong autoregressive model it wants to reuse.

BYOD is designed to avoid that cost. Instead of retraining or fine-tuning every parameter in the base model, the recipe trains only LoRA adapters. LoRA, or low-rank adaptation, adds small trainable modules to a frozen model. In BYOD, those adapters are what induce diffusion generation behaviour in a model that was originally trained to generate autoregressively.

The authors stress that the recipe works without architecture-specific tuning. That claim is central to the paper: the same general approach was applied across models from different families, rather than being hand-tailored to one architecture.

Which models were converted, and how much training did it take?

According to the paper, BYOD was applied to four models: Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B. The authors say that, to their knowledge, this is the first reported AR-to-DLM conversion of a Mistral model, referring to Ministral 8B.

The scale of the conversion is deliberately modest. The adapters comprise 4.1 to 5.8 percent of the final parameters, while every base-model parameter remains frozen. That means the original model weights are not changed during conversion; only the added adapter parameters are trained.

Each conversion ran for 25,000 iterations. According to the paper, each one needed at most 102.4 million tokens and took between 3.1 and 4.0 hours on one Google Colab GPU. For researchers and developers, that is the most practical number in the paper. It suggests that converting an existing model to diffusion generation does not necessarily require a large training cluster, at least for the models and settings tested here.

The authors also say they provide a standalone training and evaluation implementation, along with interactive demos for all four conversions, to enable further experimentation and inspection. That matters because a conversion recipe is only useful if other people can reproduce it, test it on their own models, and see where it fails.

How well do the converted models perform?

The paper reports that all four converted models generate from fully masked sequences. That is a basic check that the diffusion behaviour has been induced: the model can start from a sequence in which the tokens to be generated are masked, and produce text from that starting point.

The authors also report that all four models use bidirectional context in a controlled attention test. That test is meant to check whether the converted models really do attend in both directions, as diffusion models are expected to, rather than simply imitating autoregressive behaviour.

In addition, the converted models retain substantial parent-model knowledge, according to the paper. That is an important result for a frozen-base approach. If conversion destroyed what the original model knew, the adapters would be of limited use. The authors' evaluation suggests that much of the parent model's knowledge survives the change in generation style.

The strongest conversion was BYOD-Gemma. On the authors' nine-task evaluation, it reached 50.6 percent average accuracy, compared with 51.9 percent for LLaDA 8B Instruct in the same harness. The gap to a dedicated diffusion model is therefore small on that benchmark suite, at least for the tasks the authors chose.

That comparison should be read carefully. It is a comparison within the authors' own evaluation harness, on nine tasks, and against one reference diffusion model. It does not show that BYOD conversions match dedicated diffusion models in general, and the paper itself points to a clear weakness in open-ended generation, discussed below.

What is the main limitation?

The principal limitation, according to the paper, is open-generation quality at low network function evaluation budgets. Network function evaluation, or NFE, refers to how many times the model network is evaluated during generation. Lower NFE budgets mean faster generation in principle, but they can also mean lower quality if the model needs more steps to refine its output.

The authors measure this with perplexity when generating 128 tokens in 64 steps. BYOD perplexity ranges from 8.3 to 11.1, compared with 4.1 for LLaDA and 2.2 to 3.0 for the AR parent models. Lower perplexity generally indicates a better fit to the data being modelled, so on this measure the BYOD conversions lag behind both the dedicated diffusion model and their own autoregressive parents.

That gap is the key caveat for anyone considering BYOD. The conversions can generate from masked sequences, use bidirectional context, and retain much of their parent knowledge, but their open-ended generation quality is not yet at the level of the models they are compared with. The authors describe this as the gap that efficient conversion must next close.

For builders, the distinction matters. If a task depends mainly on knowledge retention and structured evaluation, BYOD may already be interesting. If a task depends on fluent, high-quality open-ended text at low generation budgets, the paper's own numbers suggest caution.

What does this mean for builders, and how can you explore it in Grafy?

BYOD offers a practical route for extending diffusion generation to other existing model families, in the authors' words. That is the broader significance. Diffusion language models have so far been limited by the cost of pretraining or full fine-tuning. A LoRA-only conversion that works across several families, with a frozen base, lowers the barrier to experimenting with diffusion generation on models that already exist.

For teams that already rely on Gemma, Llama, Qwen, or Mistral models, that could mean testing diffusion-style generation without starting from scratch. It also gives researchers a way to study how generation behaviour changes when the same underlying knowledge is accessed through a diffusion process rather than an autoregressive one.

Readers who want to dig into the paper itself can use Grafy Reader to read, search and question PDFs and papers with AI, including checking how the authors define their evaluation tasks and where they report limitations. For a hands-on comparison of how different models respond to the same prompt, Grafy Chat lets you chat with many AI models side by side. Developers who want to inspect or adapt the released implementation can use Grafy Code, an AI coding studio, to explore training and evaluation code in a workspace.

The Grafy News Desk view is straightforward: BYOD is best understood as an efficiency result, not a claim that converted models have overtaken dedicated diffusion models. Its value lies in showing that diffusion generation can be induced in existing models with a small adapter and a modest training run. Whether that is enough for production use will depend on closing the open-generation quality gap the authors identify.

Frequently asked questions

Does BYOD update the base model's parameters during conversion?

No. According to the paper, every base-model parameter remains frozen during BYOD conversion. Only the LoRA adapter parameters are trained, and those adapters comprise 4.1 to 5.8 percent of the final parameters in each converted model.

Which BYOD conversion was strongest on the nine-task evaluation?

BYOD-Gemma was the strongest BYOD conversion on the authors' nine-task evaluation. It reached 50.6 percent average accuracy, compared with 51.9 percent for LLaDA 8B Instruct in the same harness, according to the paper.

How long did each BYOD conversion take on one Google Colab GPU?

Each conversion took between 3.1 and 4.0 hours on one Google Colab GPU, according to the paper. Each 25,000-iteration conversion needed at most 102.4 million tokens, with the base model frozen throughout.

Source: arXiv Computation and Language, arXiv:2610.41574 — https://arxiv.org/abs/2610.41574

---

Source: arXiv Computation and Language — BYOD: Build Your Own DLM - Efficient Autoregressive-to-Diffusion Language Model Conversion