Research map / Architectures and efficiency / Quantization
Preserve-Then-Quantize: Balancing Rank Budgets for Quantization Error Reconstruction in LLMs
Yoonjun Cho, Dongjae Jeon, Soeun Kim, Moongyu Jeon, Albert No
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 |
|---|---|---|---|
| preserve top-k singular subspace before quantizing (SRR) on QERA-exact | QERA-exact, 3-bit MXINT r=64 · WikiText2 Gemma-2 2B | perplexity: 19.36 -> 18.30 | Improved |
| add SRR to LQER | LQER, 3-bit MXINT r=32 · WikiText2 LLaMA-2 7B | perplexity: 14.51 -> 11.22 | Improved |
| add SRR to QERA-exact | QERA-exact, 3-bit MXINT r=64 · WikiText2 LLaMA-3.1 70B | perplexity: 6.55 -> 6.46 | Improved |
| add SRR to QERA-exact | QERA-exact, 3-bit MXINT r=64 · 5-task zero-shot average, LLaMA-3.1 8B | accuracy: 59.05 -> 60.79 | Improved |
| apply SRR with 2-bit QuIP# quantizer | QERA-exact · WikiText2 LLaMA-3.1 8B | perplexity: 14.84 -> 13.73 | Improved |
| apply SRR with 3-bit GPTQ quantizer | QERA-exact · WikiText2 LLaMA-2 7B | perplexity: 10.06 -> 9.98 | Improved |
| use SRR initialization for QPEFT instead of QERA | RoBERTa-base 2-bit QPEFT · GLUE average | average score: 72.51 -> 78.43 | Improved |
| use SRR initialization for QPEFT instead of QERA | LLaMA-3.1 8B 2-bit QPEFT · GSM8K | accuracy: 25.11 -> 26.87 | Improved |
| use SRR with 4-bit QPEFT on RTE-sized small gains | LLaMA-3.1 8B 4-bit QPEFT · GSM8K | accuracy: 54.20 -> 54.34 vs QLoRA | No clear effect |
| no gradient attenuation on preserved directions (gamma=1 vs 0.1) (ablation) | SRR QPEFT, 2-bit · GLUE average | average score: 73.38 -> 78.43 when adding gamma=0.1 | Improved |
| freeze preserved directions entirely (gamma=0) vs gamma=0.1 (ablation) | SRR QPEFT, 2-bit · GLUE average | average score: 71.10 -> 78.43 when using gamma=0.1 | Improved |
| use SGP instead of fixed gamma=0.1 (ablation) | SRR QPEFT, 2-bit · GLUE average | average score: 78.43 -> 78.88 | No clear effect |
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.
- LQER — LQER: Low-Rank Quantization Error Reconstruction for LLMs (ICML 2024) · better on 1 of 1 shared result
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
- LQER: Low-Rank Quantization Error Reconstruction for LLMs (ICML 2024 · similarity 0.91)
- QERA: an Analytical Framework for Quantization Error Reconstruction (ICLR 2025 · similarity 0.90)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.88)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.87)
- ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization (AAAI 2025 · similarity 0.87)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.87)
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.
