New method improves text-model accuracy by up to 6.7 percentage points

One-Line Summary

For language models that refine text through repeated rounds, accuracy and computational cost depend on the task—not just the model or output length.

Research Papers

T2Quant reallocates precision to improve text accuracy

T2Quant improves the accuracy of text-generating models that use reduced numerical precision. Acrossthree models andsix benchmarks, the paper reports improvements in average accuracy of up to6.7 percentage points over existing methods while keeping the same precision budget. ¹Diffusion language models ** (DLMs) generate text through repeated cleanup rounds, a process calleddenoising**.Quantization reduces how many bits represent internal numbers, much like rounding figures to fewer decimal places. The paper finds that sensitivity to this loss of numerical detail changes with both the task and the generation step. ¹Mixed-precision quantization keeps more numerical detail in sensitive computations and less in others. T2Quant uses a sensitivity signal that requires neither labeled correct answers nor task-specific scoring rules to allocate precision for intermediate calculation values across denoising steps. Its reported gains come from allocation within a fixed budget, making accuracy on the same task under that same budget the relevant comparison. ¹

Text-generation costs change with the writing task

A second study finds that generating equally long answers can require very different amounts of computation depending on the writing task. Tests of two diffusion models across 10 tasks show that copying or lightly editing input is cheaper than open-ended writing. ForDiffusionGemma, story writing requiresmore than three times as many model runs per token—a small unit of text—as grammar correction. ²

Two properties explain much of the gap: how much of the output is new compared with the input, and how rare its words are within the task. The same pattern appears within a response: a final answer that repeats earlier reasoning costs less, while one that adds new information costs more. Languages that split into more tokens also require more model runs per token, adding computation beyond the cost of their larger token counts. ²

On the same math problem, wrong answers take more model runs per token than correct answers. Giving faster-generated candidate answers more weight improves answer selection over majority voting by up to 4.2 points, with no extra computational cost in the reported tests. Speed is therefore a useful signal in this study, not a guarantee that any particular answer is correct. ²

Why It Matters

Together, these studies separate two questions that a single accuracy or speed figure can hide: where a model needs numerical detail, and how much computation a particular writing task requires. One reallocates a fixed precision budget; the other shows why output length alone does not describe computational cost. The useful vocabulary is quantization—representing numbers with less detail—andmodel runs per token—the computational work needed for each text chunk—with task-matched comparisons connecting the two. ^{1 2}