Research map / Architectures and efficiency / Quantization
FAIR-Calib: Frontier-Aware Instability-Reweighted Calibration for Post-Training Quantization of Diffusion Large Language Models
Haoyu Huang, Linlin Yang, Sheng Xu, Boyu Liu, Guodong Guo, Zhongqian Fu, Hang Zhou, Baochang 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 frontier-hit weighting to uniform PTQ (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 61.76 -> 63.12 | Improved |
| add masked-stage reliability weighting to uniform PTQ (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 61.76 -> 62.89 | Improved |
| combine both signals (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 63.12 -> 64.64 | Improved |
| raise probing budget from 128 to 512 (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 63.15 -> 64.64 | Improved |
| raise probing budget from 512 to 1024 (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 64.64 -> 64.63 | No clear effect |
| use early-boost time weighting instead of late-boost (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 62.58 -> 64.64 | Improved |
| use early-boost time weighting instead of uniform (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 63.21 -> 64.64 | Improved |
| add lambda0 time weighting (ablation) | FAIR-Calib on Dream-Base · W4A4 average | avg accuracy (%): 62.89 -> 64.64 | Improved |
| use FAIR-Calib instead of FlatQuant | Dream-Base · W4A4 average | avg accuracy (%): 62.08 -> 64.64 | Improved |
| use FAIR-Calib instead of FlatQuant | LLaDA-Base · W4A4 average | avg accuracy (%): 59.37 -> 61.09 | Improved |
| use FAIR-Calib instead of FlatQuant | LLaDA-Instruct · W3A4 average | avg accuracy (%): 69.58 -> 70.41 | Improved |
| add position-weighted MSE to QDrop-style calibration | LLaDA-Instruct · W8A8 average | avg accuracy (%): 70.66 -> 73.17 | 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.
- FlatQuant — FlatQuant: Flatness Matters for LLM Quantization (ICML 2025) · better on 7 of 7 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · 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
- Uncertainty Quantification for Large Language Diffusion Models (NeurIPS 2026 · similarity 0.91)
- Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (NeurIPS 2026 · similarity 0.90)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.89)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models (ICLR 2026 · similarity 0.89)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · 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.
