Research map / Architectures and efficiency / Pruning and efficient tuning
BoRA: Towards More Expressive Low-Rank Adaptation with Block Diversity
Shiwei Li, Xiandi Luo, Haozhao Wang, Xing Tang, Ziqiang Cui, Dugang Liu, Yuhua Li, Yichen Li, Xiuqiang He, Ruixuan Li
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 LoRA r=8 with BoRA r=8, b=8 | LoRA · GLUE avg, RoBERTa-Base | accuracy (%): 82.55 -> 83.65 | Improved |
| Replace LoRA r=8 with BoRA r=8, b=16 | LoRA · GLUE avg, RoBERTa-Large | accuracy (%): 84.38 -> 86.51 | Improved |
| Replace LoRA r=8 with BoRA r=8, b=16 | LoRA · Math10K avg, LLaMA-3-8B | accuracy (%): 68.19 -> 72.24 | Improved |
| Replace LoRA r=8 with BoRA r=8, b=16 | LoRA · Commonsense170K avg, LLaMA-3-8B | accuracy (%): 85.45 -> 86.62 | Improved |
| Use BoRA r=8 b=16 instead of 4x larger LoRA r=32 | LoRA r=32 · Math10K avg, Qwen2.5-14B | accuracy (%): 80.15 -> 80.60 | Improved |
| Use BoRA r=8 b=16 instead of LoRA r=32 | LoRA r=32 · Commonsense170K avg, LLaMA-3-8B | accuracy (%): 86.72 -> 86.62 | No clear effect |
| Use MELoRA (zeroed off-diagonal blocks) instead of LoRA | LoRA r=8 · GLUE avg, RoBERTa-Base | accuracy (%): 82.55 -> 81.74 | Hurt |
| Increase blocks from 8 to 16 (ablation) | BoRA r=8 · GLUE avg, RoBERTa-Large | accuracy (%): 85.83 -> 86.51 | Improved |
| Add exponential function to sigma parametrization (ablation) | BoRA · Math10K | accuracy | Improved |
| Add mean-absolute-value normalization to sigma (ablation) | BoRA · Math10K | accuracy | Improved |
| Increase block number b at fixed rank (ablation) | BoRA · Math10K, LLaMA-3-8B | accuracy | Mixed |
| Apply BoRA vs LoRA at different tuned modules (Q, QV, QKVUD, QKVOGUD) (ablation) | LoRA · Math10K, LLaMA-3-8B | accuracy | Improved |
Closest papers by meaning
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.93)
- LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025 · similarity 0.93)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.92)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.92)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.92)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.92)
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.
