Research map / Architectures and efficiency / Quantization
Qronos: Correcting the Past by Shaping the Future... in Post-Training Quantization
Shihao Zhang, Haoyu Zhang, Ian Colbert, Rayan Saab
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 |
|---|---|---|---|
| replace OPTQ rounding with Qronos (corrects activation/previous-layer mismatch) | HIP+MagR 2-bit weight-only · WikiText2, Llama3.2-1B | perplexity: 24.6 -> 17.8 | Improved |
| replace OPTQ rounding with Qronos | HIP+MagR 1.58-bit weight-only · WikiText2, Llama3-8B | perplexity: 43.3 -> 18.0 | Improved |
| use Qronos instead of GPTAQ at 2-bit with unfavorable grid beta=1 (ablation) | HIP+MagR · WikiText2, Llama3-8B | perplexity: 16.4 -> 12.4 | Improved |
| grid scale beta=0.8 instead of beta=1 for 2-bit (ablation) | Qronos HIP+MagR · WikiText2, Llama3-8B | perplexity: 12.4 -> 9.3 | Improved |
| replace OPTQ with Qronos in W3 HIP weight-only | Llama3 · WikiText2, 8B | perplexity: 8.6 -> 7.8 | Improved |
| replace GPTAQ with Qronos at W4 HIP | Llama3-8B · 0-shot avg, W4 HIP | accuracy: 72.4 -> 71.0 | Hurt |
| replace OPTQ with Qronos in W4A4KV4 SpinQuant | Llama3-3B · WikiText2 | perplexity: 15.0 -> 11.1 | Improved |
| replace GPFQ with Qronos in W3A3 QuaRot | Llama3-8B · WikiText2 | perplexity: 27.9 -> 20.4 | Improved |
| add Qronos instead of GPTAQ in W4A4KV4 QuaRot at 8B | Llama3-8B · WikiText2 | perplexity: 9.3 -> 9.3 | No clear effect |
| use Qronos instead of OPTQ on instruct Llama3 W3 no transform | Llama3-8B-Instruct · WikiText2 | perplexity: 12.2 -> 10.6 | Improved |
| SmoothQuant gamma 0.3 vs 0.8 (RTN, 4-bit) (ablation) | Llama3.2-1B-Instruct · WikiText2 | perplexity: 3e4 -> 18.6 | Improved |
| MagR theta 0.01 vs 0.001 (ablation) | Llama3.2-1B-Instruct RTN 4-bit · WikiText2 | perplexity: 105.0 -> 25.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GPTAQ — GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration (ICML 2025) · better on 7 of 9 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
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.87)
- Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization (NeurIPS 2025 · similarity 0.87)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.86)
- FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · similarity 0.86)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.86)
- MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization (NeurIPS 2024 · similarity 0.86)
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.
