Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models

One-line takeaway

Quantization sensitivity in diffusion language models varies by task and denoising step; label-free T²Quant allocates activation precision across steps under a fixed bit budget, improving accuracy by up to 6.7 points across three DLMs and six benchmarks.

论文发现扩散语言模型的量化敏感度随任务和去噪步而变，提出无需标签的 T²Quant，在固定位宽预算下按去噪步分配激活精度，在 3 个模型 6 个基准上最多提升 6.7 个百分点。