Researchers have found that diffusion language models respond very differently depending on where you place the input query — a problem that AR-style copy-paste prompting has been quietly creating.
Most people building with large language models have inherited prompt templates designed for autoregressive models, which process text left-to-right with causal masking. Diffusion LLMs work differently: they use bidirectional attention, meaning the model sees the full context at once and has no structural reason to privilege text at the end. A new paper from researchers posted to arXiv shows that this structural difference is not academic — positional variance in diffusion LLMs affects output quality roughly as much as choosing better or worse examples in the prompt. The team traced the root cause to a spatial "recency effect" in how attention flows and to shifts in the iterative decoding trajectory depending on where the query sits.
That finding matters because diffusion LLMs are gaining traction as a faster, parallelizable alternative to autoregressive generation, with models like Gemini Diffusion and Inception Labs' Mercury drawing real commercial interest. If the community is unknowingly bottlenecking these models by using the wrong prompt layout, published benchmarks may be understating what they can actually do.
The paper also introduces a training-free remedy called Auto-ICL, which dynamically routes queries to better positions using a new confidence metric — Average Confidence — designed to track the multi-step denoising process rather than relying on a single-step snapshot, which the authors show fails in this setting. The honest caveat: results here are from controlled experiments on reasoning and perception tasks, and real-world prompt engineering is messier than a lab setup suggests.