Research map / Architectures and efficiency / Quantization
Quant-dLLM: Post-Training Extreme Low-Bit Quantization for Diffusion Large Language Models
Tianao Zhang, Zhiteng Li, Xianglong Yan, Haotong Qin, Yong Guo, Yulun Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Masked Calibration Simulation (ablation) | Quant-dLLM · MMLU 5-shot, LLaDA-8B-Base | accuracy: 52.10 -> 56.87 | Improved |
| add Masked Calibration Simulation (ablation) | Quant-dLLM · MMLU 5-shot, Dream-7B-Base | accuracy: 37.81 -> 40.22 | Improved |
| add ABMP with 5% reallocation (ablation) | Quant-dLLM · MMLU 5-shot, LLaDA-8B-Base | accuracy: 54.32 -> 56.87 | Improved |
| add ABMP with 10% reallocation (ablation) | Quant-dLLM · MMLU 5-shot, Dream-7B-Base | accuracy: 32.75 -> 40.22 | Improved |
| raise ABMP reallocation ratio from 5% to 15% (ablation) | Quant-dLLM · MMLU 5-shot, LLaDA-8B-Base | accuracy: 56.87 -> 53.01 | Hurt |
| raise ABMP reallocation ratio from 10% to 15% (ablation) | Quant-dLLM · MMLU 5-shot, Dream-7B-Base | accuracy: 40.22 -> 31.11 | Hurt |
| add RSR (without data-aware objective) to baseline quantizer (ablation) | DAQ · MMLU 5-shot, LLaDA-8B-Base | accuracy: 39.26 -> 48.32 | Improved |
| add data-aware objective (DOR) on top of RSR (ablation) | DAQ · MMLU 5-shot, LLaDA-8B-Base | accuracy: 48.32 -> 56.87 | Improved |
| add RSR and DOR to baseline quantizer (ablation) | DAQ · MMLU 5-shot, Dream-7B-Base | accuracy: 27.84 -> 40.22 | Improved |
| increase calibration samples from 128 to 256 (ablation) | Quant-dLLM · MMLU 5-shot, LLaDA-8B-Base | accuracy: 56.87 -> 55.59 | No clear effect |
| increase calibration samples from 64 to 128 (ablation) | Quant-dLLM · MMLU 5-shot, LLaDA-8B-Base | accuracy: 54.49 -> 56.87 | Improved |
| use Quant-dLLM instead of Slim-LLM | 2-bit PTQ · 7 general tasks avg, LLaDA-8B-Base | average accuracy: 42.39 -> 54.06 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Slim-LLM — SlimLLM: Accurate Structured Pruning for Large Language Models (ICML 2025) · better on 5 of 5 shared results
- GPTAQ — GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration (ICML 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.91)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.90)
- Uncertainty Quantification for Large Language Diffusion Models (NeurIPS 2026 · similarity 0.90)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.90)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · similarity 0.89)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
