T²Quant: task- and timestep-aware mixed precision for quantizing diffusion LMs
Ofek Avivi · hf · 2026-10-24
Researchers from Technion and Ben-Gurion University propose T²Quant for quantizing diffusion language models. They show quantization sensitivity in DLMs is both task- and timestep-dependent — tasks like GSM8K, HumanEval, and MMLU-Pro exhibit distinct precision requirements across 128 denoising steps — and that these patterns can be predicted without ground-truth labels or task-specific metrics. T²Quant exploits this to allocate activation precision dynamically across denoising steps.
More from Research
- Teaching intro theory of computation in the AI era: NP-completeness isn't the end anymore — fortnow · 2026-10-24
- AgentGUI at EMNLP 2026: a pixel-art interface for observing and steering AI agents — Michael_D_Moor · 2026-10-24
- DISARM paper: disclosure cuts manipulative LLM's belief shift from 15.3 to 3.3 points — MaartenSap · 2026-10-24
- mini-AGI: continual-learning model trained on 8GB GPU for $56 beats GPT-2 Small on enwik9 — Another__one · 2026-10-24
- Cisco ships Agentic FAITH: one workflow for cybersecurity agent evals — aminkarbasi · 2026-10-24
- 1979 paper dismissed formal verification; 2026 forecast says it'll be as common as fuzzing — danbri · 2026-10-24