Research map / Architectures and efficiency / Pruning and efficient tuning
DiaBlo: Diagonal Blocks Are Sufficient For Finetuning
Selcuk Gurses, Aozhong Zhang, Yanxia Deng, Xun Dong, Xin Li, Naigang Wang, Penghang Yin, Zi Yang
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 |
|---|---|---|---|
| Fine-tune diagonal blocks (DiaBlo) instead of LoRA | LoRA r=32 · commonsense reasoning, Llama2-7B | avg accuracy: 77.6 -> 83.5 | Improved |
| Use DiaBlo instead of SMT | SMT · commonsense reasoning, Llama2-7B | avg accuracy: 81.8 -> 83.5 | Improved |
| Use DiaBlo instead of full fine-tuning | full FT · commonsense reasoning, Llama3-8B | avg accuracy: 87.5 vs 87.3 | No clear effect |
| Use DiaBlo instead of LoRA on arithmetic reasoning | LoRA r=64 · GSM8K/MATH avg, Llama2-7B | avg exact match: 38.7 -> 43.4 | Improved |
| Use N=64 instead of N=32 diagonal blocks (ablation) | DiaBlo · MATH, Llama2-7B | exact match: 20.4 -> 17.6 | Hurt |
| Use DiaBlo instead of LoRI | LoRI · HumanEval Pass@1, Llama3-8B | Pass@1: 43.2 vs 43.2 | No clear effect |
| Use DiaBlo instead of DoRA/LoRA for safety alignment | LoRA · HEx-PHI, Llama3-8B | refusal rate: 91.6 -> 97.6 | Improved |
| Replace LoRA with DiaBlo on MagR-quantized 2-bit model | MagR-LoRA · arithmetic avg, Llama2-7B 2-bit | avg accuracy: 46.2 -> 48.7 | Improved |
| Replace LoRA with DiaBlo on 4-bit Llama2-7B | QLoRA · arithmetic avg, 4-bit | avg accuracy: 53.7 -> 54.8 | Improved |
| Use diagonal blocks instead of random entries as sparsity pattern (ablation) | DiaBlo · GSM8K fine-tuning, Llama3-8B | accuracy: 65.35 -> 67.68 | Improved |
| Use diagonal blocks instead of random rows (ablation) | DiaBlo · GSM8K retained update at 1/128 | accuracy: 2.85 -> 63.11 | Improved |
| Use DiaBlo instead of DoRA | DoRA · training time per epoch | minutes: 480 -> 170 | Improved |
Closest papers by meaning
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.90)
- Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.89)
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.89)
- RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models (NeurIPS 2025 · similarity 0.89)
- Loquetier: A Virtualized Multi-LoRA Framework for Unified LLM Fine-tuning and Serving (NeurIPS 2025 · similarity 0.89)
- GoRA: Gradient-driven Adaptive Low Rank Adaptation (NeurIPS 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.
