arxiv:2610.31198

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

Published on Oct 22

· Submitted by

Eliron Rahimi on Oct 23

Upvote

Authors:

Ofek Avivi ,

Eliron Rahimi ,

Abstract

T2Quant improves diffusion language model quantization by dynamically allocating activation precision across denoising steps based on predicted task- and time-dependent sensitivity.

Generated by thinkingmachines/Inkling-Small

Diffusion Language Models (DLMs) introduce unique quantization challenges due to their iterative denoising process. In this work, we reveal that quantization sensitivity is both taskand time-dependent, with different downstream tasks exhibiting distinct precision requirements across denoising steps. We further show that these patterns can be predicted without groundtruth labels or task-specific evaluation metrics. Building on these findings, we introduce T 2Quant, a task and time-aware mixedprecision quantization framework that uses a label-free sensitivity signal to optimally allocate activation precision across denoising steps under a fixed bit-width budget. Extensive experiments across three DLMs and six benchmarks demonstrate that T 2Quant consistently improves the average accuracy of existing quantization methods, achieving gains of up to 6.7 percentage points under the same budget.

View arXiv page View PDF Project page GitHub 3 Add to collection

Community

ElironRahimi

Paper author Paper submitter 4 days ago

Excited to share our new paper: Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models! 🚀

We reveal that quantization sensitivity is both task and time-dependent, with different downstream tasks exhibiting distinct precision requirements across denoising steps.

Building on this finding, we introduce T²Quant, a mixed-precision framework that:

🔍 Predicts precision sensitivity using only a few unlabeled calibration prompts.
⚙️ Allocates activation precision across denoising steps under a fixed bit-width budget.
📈 Improves existing quantization methods by up to 6.7 percentage points, without increasing the average precision budget.

Evaluated across 3 diffusion language models and 6 benchmarks.

📄 Paper: https://arxiv.org/abs/2610.31198

We welcome feedback, questions, and discussions!

librarian-bot

4 days ago

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

guicybercode

1 day ago

The label-free sensitivity signal is the part I care about, since it means the precision schedule could be picked per task at deployment. How expensive is computing that signal compared to one generation pass? And does the task and timestep dependence also show up with weight-only quantization, the setup most people use for local inference on CPU, or only when activations are quantized as in your experiments?

ElironRahimi

Paper author about 7 hours ago

Hi @guicybercode , thank you very much for reading our paper and for your questions and discussion!

Regarding the sensitivity signal, the time required to construct it depends on the number of prompts in the calibration set, and we showed that effective results can be achieved with as few as 16 prompts!

Nevertheless, our setting follows the standard task-aware quantization approach, which includes an offline calibration phase, and signal construction is part of this phase. Therefore, it does not introduce additional overhead during inference, and we did not explicitly measure its construction time.

A setting without a calibration phase, where the signal is predicted dynamically for each new prompt during inference, is a different and much more challenging setup that was beyond the scope of this work. That said, we believe our paper provides a foundation for such future research. Based on our novel findings that temporal mixed precision can improve DLM quantization, and that precision sensitivity depends on the task and can be predicted, we believe such an approach may be possible, or at least worth exploring!

Regarding the weight quantization question, our method focuses on activation quantization and can be effectively integrated with weight quantization, as we demonstrated in the paper, where this integration successfully improved quantization results.

Thank you again for your questions and for the discussion about our paper!