on text diffusion but it does come with some intriguing promises. One promise is faster inference. This is particularly relevant for latency sensitive applications like coding, but also convenient for LLM providers because a 10x speed up means 10x cost savings. ARMS perform one iteration per output token until the special end of sequence token is produced. Depending on your prompt, that can be in the thousands or more. Diffusion models can be faster because they require fewer refinement operations. The exact number is a hyperparameter that acts as a knob between speed and quality. For good outputs, a reasonable number of refinements today is in the tens or low hundreds. But new research keeps bringing this number down. Now this reduction in the number of iterations is only meaningful as long as we can generate all the fusion tokens in parallel. Inception labs report they built their own proprietary inference engine to maximize GPU utilization and ensure high token parallelization and their efforts are paying off. A recent study from Bite Dance, which is Tik Tok's parent company, shows that diffusion LLMs achieve speeds between 1,000 to 2,000 tokens per second on an age 20 GPU, while a comparable auto reggressive model will cap somewhere around 200. That's a 10x speed up already. The Fusion LLMs also promise higher quality outputs. The inherent left to right bias in auto reggression limits the behavior of the model in several ways. An obvious one is once a token is produced, it can't be revised. In this example, I asked the GPT40 model a yes or no question. Its response contradicted itself, going from a yes to a no after some intermediate thinking. But diffusion could have handled this change apart more naturally by iteratively revising the entire response. Left right generation can also lead to exposure bias or sampling drift. The model is trained to predict the next token for a ground truth prefix from the training set. But during inference, it has to complete its own previous output. So a small error can snowball as text gets longer leading to drift. Currently, ARMS mitigate this with hacks like temperature or nuclear sampling, but the fusion would avoid this altogether by working on the entire sequence holistically. Finally, the fusion offers flexible prompting. While auto reggressive models expect the prompt to be a prefix, the fusion LLMs allow the prompt to sit at any arbitrary position. This addresses certain use cases more naturally. For instance, filling in missing spans in a PDF, rewriting a paragraph in the middle of an essay, or refactoring a block of code. ARMS can also do this already, but more awkwardly, often by regenerating parts of the document that shouldn't be touched. I suspect that AI coding tools like Cursor currently need a lot of special logic to incorporate the LLM output back into the existing code in the right place. So diffusion LLMs make interesting promises, but to me what's even more interesting is their