T2Quant tunes DLM quantization by task and denoising step
TL;DR
- Quantization sensitivity in diffusion language models is both task-dependent and time-dependent across denoising steps, the paper reports.
- T2Quant uses a label-free sensitivity signal to allocate activation precision across steps under a fixed bit-width budget.
- Across three DLMs and six benchmarks, T2Quant reports gains of up to 6.7 percentage points over existing quantization methods.
Quantizing a diffusion language model well on one task does not mean quantizing it well on all of them, according to a new arXiv preprint. The authors, Ofek Avivi, Eliron Rahimi, Avi Mendelson and Chaim Baskin, report that sensitivity to lower-precision activations shifts both with the downstream task and with where the model is inside its iterative denoising loop.
Their framework, T2Quant, uses a label-free signal to decide how many bits each denoising step should get under a fixed budget. "T2Quant consistently improves the average accuracy of existing quantization methods, achieving gains of up to 6.7 percentage points under the same budget," the abstract reports, across three DLMs and six benchmarks. The authors also claim the sensitivity patterns "can be predicted without groundtruth labels or task-specific evaluation metrics."
The abstract publishes no per-task breakdown, no baseline methods by name, and no cost for computing the sensitivity signal itself.
Originally reported by paper
Read the original article →Original headline: T²Quant Finds DLM Quantization Sensitivity Is Task- and Timestep-Dependent, Gains Up to +6.7 pp