T2Quant Reallocates Precision Across Diffusion LLM Steps: Up to +6.7 pp

Diffusion language models generate text differently from conventional autoregressive LLMs. They do not write token by token. Instead, they gradually "clear up" a noisy sequence over several denoising steps. This design has a cost: the model has to be run many times in a row. Researchers behind a recent paper on Hugging Face Daily Papers propose cutting that cost through quantization and introduce a method called T2Quant. It allocates activation bit-width across denoising steps based on the specific task. Across three DLMs and six benchmarks, T2Quant improves the average accuracy of existing quantization methods by up to 6.7 percentage points at the same bit budget.

Why denoising steps break conventional quantization

Standard LLM quantization is tuned as if all of the model's forward passes were equivalent. For diffusion models, that assumption does not hold. A rounding error introduced at an early step, when the text is still almost pure noise, behaves differently from the same error at the final steps, when the model is refining details. The authors show that sensitivity to quantization depends on time, that is, on the step number.

The second dependency is tied to the task. Different downstream tasks need different precision at different stages of denoising. A profile that suits one type of benchmark can cost noticeable quality on another. A single precision scheme for all steps and all tasks ends up as a compromise that is optimal nowhere.

A signal that needs no labels

The paper's main practical finding is that these sensitivity patterns can be predicted without ground-truth labels and without task-specific metrics. Tuning a quantization configuration for a task normally means running evaluation on labeled data, which is expensive and not always possible. T2Quant uses a label-free sensitivity signal to decide which steps need more bits for activations and which can get by with fewer.

The constraint is strict: the total bit budget is fixed. The method does not add bits beyond the limit; it redistributes them across steps. If precision can be cut almost painlessly at some stages of generation, the freed-up resources go to the critical steps.

Key parameters of the work:

• subject: diffusion language models (DLMs);
• what is optimized: activation bit-width across denoising steps;
• constraint: fixed bit budget;
• selection signal: label-free, with no labels or task metrics;
• evaluation: three DLMs and six benchmarks;
• result: average accuracy of existing quantization methods rises by up to +6.7 pp.

What to change in the inference pipeline

One important detail: T2Quant does not replace existing quantization methods; it improves them. The authors report a consistent gain in average accuracy on top of already known approaches. A team deploying a DLM does not need to rewrite its whole stack. It is enough to add a layer that decides which activation bit-width to use at each step.

Inference engineers can take three points from the paper. First, the per-step precision profile deserves to be treated as a separate tuning parameter rather than a constant. Second, the profile can be tuned for a task without labeled data, which means it can also be tuned in production, where there are no reference answers. Third, under tight memory and compute limits, a gain of several percentage points comes without increasing the bit budget.

For the industry, this is a sign that diffusion language models are moving toward the same optimization tools long applied to autoregressive LLMs, with an adjustment for the time dimension. The cheaper iterative denoising gets, the more realistic it becomes for DLMs to compete with conventional models on cost per token. For now, this is an academic result on three models and six benchmarks. But the principle that precision depends on both step and task will almost certainly be tested in other DLM implementations.