Research map / Large language models / Instruction and fine-tuning
QWHA: Quantization-Aware Walsh-Hadamard Adaptation for Parameter-Efficient Fine-Tuning on Large Language Models
Hyesung Jeon, Seojune Lee, Beomseok Kang, Yulhwa Kim, Jae-Joon Kim
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 AdaAlloc parameter selection (vs random with QA init) (ablation) | QWHA (WHA) · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 24.48 -> 29.11 | Improved |
| add AdaAlloc parameter selection (vs magnitude) (ablation) | QWHA (WHA) · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 28.12 -> 29.11 | Improved |
| add AdaAlloc parameter selection (vs LoCA selection) (ablation) | QWHA (WHA) · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 21.15 -> 29.11 | Improved |
| add AdaAlloc parameter selection (vs SSH selection) (ablation) | QWHA (WHA) · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 27.14 -> 29.11 | Improved |
| add quantization-aware initialization and refinement to WHA with random selection (ablation) | WHA · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 24.41 -> 24.48 | No clear effect |
| use WHA adapter instead of DCA (ablation) | QWHA with AdaAlloc · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 27.29 -> 29.11 | Improved |
| use WHA adapter instead of DHA (ablation) | QWHA with AdaAlloc · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 27.52 -> 29.11 | Improved |
| use WHA adapter instead of plain sparse adapter (ablation) | QWHA with AdaAlloc · GSM8k 2-bit LLaMA-3.2-3B | accuracy (%): 26.54 -> 29.11 | Improved |
| add QWHA at 2-bit instead of CLoQ low-rank adapter | LLaMA-3.1-8B · GSM8k 2-bit | accuracy (%): 33.89 -> 37.83 | Improved |
| add QWHA at 2-bit instead of CLoQ low-rank adapter | LLaMA-3.1-8B · CSQA 2-bit | accuracy (%): 56.49 -> 60.98 | Improved |
| add quantization-aware init (QWHA vs SSH without it) | LLaMA-3.1-8B · GSM8k 3-bit | accuracy (%): 50.34 -> 55.34 | Improved |
| use QWHA at 4-bit vs 16-bit fine-tuned model | LLaMA-3.1-8B · GSM8k 4-bit | accuracy (%): 59.74 -> 56.10 | Hurt |
Closest papers by meaning
- LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models (ICLR 2024 · similarity 0.91)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.90)
- RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models (ICML 2025 · similarity 0.90)
- QERA: an Analytical Framework for Quantization Error Reconstruction (ICLR 2025 · similarity 0.89)
- OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting (ICLR 2025 · 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.
