Research map / Architectures and efficiency / Quantization
AutoQRA: Joint Optimization of Mixed-Precision Quantization and Low-rank Adapters for Efficient LLM Fine-Tuning
Changhai Zhou, Shiyang Zhang, Yuhua Zhou, Qian Qiao, Jun Gao, Cheng Jin, KAIZHOU QIN, WEIZHONG 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 |
|---|---|---|---|
| jointly search bit-width and LoRA rank (AutoQRA <=4-bit) instead of uniform 4-bit QLoRA | AutoQRA (<=4-bit) · Task average, LLaMA3.1-8B | Avg accuracy (%): 67.45 -> 69.83 | Improved |
| allow mixed precision above 4-bit average (AutoQRA Opt) vs FP16 LoRA | AutoQRA (Opt) · Task average, Qwen2.5-7B | Avg accuracy (%): 71.33 -> 73.19 | Improved |
| joint bit-rank search instead of decoupled AMQ then LoRA | AutoQRA (<=4-bit) · Task average, LLaMA3.2-3B | Avg accuracy (%): 63.31 -> 65.58 | Improved |
| joint search instead of AMQ+LoRA on code generation | AutoQRA (<=4-bit) · HumanEval | pass@1: 24.40 -> 41.96 | Improved |
| use surrogate-guided promotion instead of ranking by low-fidelity scores (ablation) | AutoQRA Phase I · Top-3 promotion hit rate | mean hit rate: 44.7% -> 67.3% | Improved |
| use AutoQRA search instead of random search (ablation) | AutoQRA · Evaluations to reach target accuracy | number of HF evaluations: 107 -> 6 | Improved |
| move allocation granularity from layer-level to tensor-level (ablation) | AutoQRA · Qwen2.5-3B | Avg accuracy (%): 66.33 -> 66.72 (search 55 min -> 3.6 h) | No clear effect |
| allocate per layer instead of uniform model-level (ablation) | AutoQRA · Qwen2.5-3B | Avg accuracy (%): 62.89 -> 66.33 | Improved |
| reduce memory budget to 0.7x (ablation) | AutoQRA · Qwen2.5-3B, combined avg with HumanEval | Avg: 64.55 -> 63.92 | Hurt |
| reuse configuration searched on Alpaca for HC3 instead of searching on HC3 (ablation) | AutoQRA · HC3, Qwen2.5-3B | Avg: 65.93 -> 65.54 | No clear effect |
| use tuned search hyperparameters (wider promotion set, more Phase II iterations) (ablation) | AutoQRA · Qwen2.5-3B 5-task average | 5-task Avg accuracy (%): 69.07 -> 70.02 | Improved |
| remove global Phase I search, importance warm start, multi-fidelity, or optimize only one axis (ablation) | AutoQRA · Accuracy vs memory (Figure 6) | accuracy-memory frontier | Hurt |
Closest papers by meaning
- LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits (ICML 2025 · similarity 0.90)
- FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · similarity 0.88)
- LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning (NeurIPS 2025 · similarity 0.88)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.88)
- RILQ: Rank-Insensitive LoRA-Based Quantization Error Compensation for Boosting 2-Bit Large Language Model Accuracy (AAAI 2025 · similarity 0.88)
- LoQT: Low-Rank Adapters for Quantized Pretraining (NeurIPS 2024 · similarity 0.88)
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.
