Research map / Architectures and efficiency / Pruning and efficient tuning
On-the-Fly Adaptation to Quantization: Configuration-Aware LoRA for Efficient Fine-Tuning of Quantized LLMs
Rongguang Ye, Ming Tang, Edith C. H. Ngai
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 |
|---|---|---|---|
| use a configuration-aware adjustment model instead of one shared LoRA adapter | Shared-LoRA · SST-2 (RoBERTa-Large) | avg accuracy gap vs QLoRA: -5.06% -> +1.74% | Improved |
| use CoA-LoRA instead of per-configuration QLoRA fine-tuning | QLoRA · QNLI/MNLI/SST-2/QQP | total training time: QNLI 119m -> 57m | Improved |
| use CoA-LoRA instead of QLoRA | QLoRA · MNLI (RoBERTa-Large) | avg accuracy gap: +8.89% | Improved |
| add Pareto-based configuration search to training configuration set (ablation) | CoA-LoRA · QNLI and SST-2 | accuracy: nearly 2% | Improved |
| randomly add ten new configurations per epoch instead of guided search (ablation) | CoA-LoRA · configuration search ablation | accuracy | No clear effect |
| use segmented Pareto selection (U=20 or 40) instead of none (U=0) (ablation) | CoA-LoRA · QNLI/SST-2/QQP | accuracy | Improved |
| increase segments from U=20 to U=40 (ablation) | CoA-LoRA · SST-2 vs QNLI/QQP | accuracy | Mixed |
| increase LoRA rank (32, 64, 128) (ablation) | CoA-LoRA · four GLUE tasks | HV: QNLI 0.619 / 0.629 / 0.623 | No clear effect |
| use CoA-LoRA on integer mixed-precision quantization | QLoRA · QNLI/MNLI/SST-2/QQP | accuracy | Mixed |
| use CoA-LoRA on LLaMA-2-13B at 3 to 4 bits | QLoRA · C4 perplexity, rank 64 | perplexity: 4-bit: 6.89 -> 6.99 | No clear effect |
| use CoA-LoRA at 2.5 bits on LLaMA-2-13B | QLoRA · C4 perplexity, 2.5-bit | perplexity: 9.11 -> 8.65 | Improved |
| use CoA-LoRA at zero-shot on Qwen2.5-3B | LQ-LoRA · zero-shot 3-bit | average accuracy: 51.90 -> 53.35 | 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.
- LQ-LoRA — LQ-LoRA: Low-rank plus Quantized Matrix Decomposition for Efficient Language Model Finetuning (ICLR 2024) · better on 9 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
- LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits (ICML 2025 · similarity 0.89)
- Accurate LoRA-Finetuning Quantization of LLMs via Information Retention (ICML 2024 · similarity 0.88)
- RILQ: Rank-Insensitive LoRA-Based Quantization Error Compensation for Boosting 2-Bit Large Language Model Accuracy (AAAI 2025 · similarity 0.87)
- Efficient Multi-task LLM Quantization and Serving for Multiple LoRA Adapters (NeurIPS 2024 · similarity 0.87)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · 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.
