Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models
T²Quant reports accuracy gains of up to 6.7 percentage points at the same quantization bit-width budget.
The preprint finds that diffusion language models’ sensitivity to quantization varies by task and denoising step. T²Quant uses a label-free sensitivity signal to allocate activation precision across those steps, without ground-truth labels or task-specific evaluation metrics. Tests across three models and six benchmarks showed consistent improvements in average accuracy over existing quantization methods.
HF Daily Papers' note
The preprint finds that diffusion language models’ sensitivity to quantization varies by task and denoising step. T²Quant uses a label-free sensitivity signal to allocate activation precision across those steps, without ground-truth labels or task-specific evaluation metrics. Tests across three models and six benchmarks showed consistent improvements in average accuracy over existing quantization methods.
HF Daily Papers' note
score 5