A training-free swap of one decoding rule pushes diffusion language model throughput up to 37% higher without touching the underlying model.
Diffusion LLMs generate text by unmasking tokens in parallel batches rather than producing them left-to-right, but choosing which tokens to commit each step has been a persistent speed bottleneck. The existing Fast-dLLM framework handled this with a rule that judged an entire candidate batch by its single least-confident token — a conservative stance that discards real information about the rest of the batch. Fast-dLLM++ replaces that with what the authors call Fréchet profile decoding, which evaluates the full sorted confidence distribution of a candidate set instead of collapsing it to its weakest element. Benchmarks on four math and coding tasks using the LLaDA-8B model showed gains of up to 37% in throughput at comparable accuracy, with no changes to model weights, training, or cache infrastructure.
Diffusion LLMs' primary argument against autoregressive models is parallel generation speed, so the actual throughput number matters more than most architectural claims. A training-free drop-in that demonstrably improves that number removes the usual "but it needs a full retrain" objection, and the code is already public on GitHub. If the gains replicate on larger models, the case for diffusion LLMs as a production deployment option becomes harder to dismiss.
Four benchmarks and one 8B model are a narrow foundation for a 37% headline, and wider replication on larger hardware will determine whether the improvement actually holds at scale.