A new method lets text-generating diffusion models borrow the tone of a reference passage, without retraining anything.
The technique, called Iterative Latent Representation Refinement (ILRR), targets masked diffusion language models (DLMs) like LLaDA and MDLM - models that generate text by denoising an entire sequence in parallel rather than writing token by token. ILRR pulls semantic signal out of a reference text and injects it into the model's activations as they evolve across denoising steps, giving developers a dial for coarse properties like sentiment. A companion technique, Spatially Modulated Steering, lets a short reference guide a much longer generated output by varying steering strength across the sequence. The cost is one extra parallel forward pass per denoising step, and the paper reports 10 to 60 point gains in attribute accuracy over baseline steering methods at similar compute.
Most steering research - prompting, activation steering, RLHF - was built for autoregressive models that commit to one token at a time. Diffusion LMs revise the whole sequence repeatedly, and this paper's core claim is that the repeated, global nature of that process makes it a better target for steering than the autoregressive alternative. If true, that's a real argument for diffusion LMs beyond their headline pitch of faster parallel generation.
Still, this is validated on research-scale models (LLaDA, MDLM) controlling a fairly blunt property, sentiment, not the nuanced style control marketers will eventually want. Whether it holds up on larger diffusion LMs, or on subtler attributes than "positive or negative," is the next question, not this paper's answer.