Research map / Architectures and efficiency / Pruning and efficient tuning
Efficiently Seeking Flat Minima for Better Generalization in Fine-Tuning Large Language Models and Beyond
Jiaxin Deng, Qingcheng Zhu, Junbiao Pang, Linlin Yang, Zhongqian Fu, Baochang 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 |
|---|---|---|---|
| add transferred full-space SAM perturbation (FMLoRA) to LoRA | LoRA · RoBERTa-large few-shot avg | average score: 80.0 -> 83.1 | Improved |
| add EMA-estimated perturbation (EFMLoRA) to LoRA | LoRA · RoBERTa-large few-shot avg | average score: 80.0 -> 82.3 | Improved |
| use EFMLoRA instead of FMLoRA with exact second pass (ablation) | FMLoRA · RoBERTa-large few-shot avg | average score: 83.1 -> 82.3 | Hurt |
| add EFMLoRA to LoRA | LoRA · GLUE (8 tasks avg, RoBERTa-large) | average score: 88.4 -> 89.4 | Improved |
| use EFMLoRA instead of full fine-tuning | full fine-tuning · GLUE (RoBERTa-large avg) | average score: 88.9 -> 89.4 | Improved |
| add EFMLoRA to LoRA | LoRA · E2E NLG (GPT-2 Medium) | BLEU: 69.2 -> 69.7 | Improved |
| add FMLoRA to LoRA | LoRA · E2E NLG (GPT-2 Medium) | BLEU: 69.2 -> 69.2 | No clear effect |
| add EFMLoRA to LoRA | LoRA · E2E NLG (GPT-2 Large) | CIDEr: 2.53 -> 2.52 | No clear effect |
| add EFMLoRA to LoRA | LoRA · CLIP EuroSAT 1-shot | accuracy (%): 72.3 -> 78.3 | Improved |
| add EFMLoRA to LoRA | LoRA · CLIP Flowers102 1-shot | accuracy (%): 83.2 -> 81.0 | Hurt |
| add EFMLoRA to LoRA | LoRA · CLIP EuroSAT 16-shot | accuracy (%): 92.1 -> 91.6 | No clear effect |
| increase perturbation radius rho from 0.2 to 0.6 (ablation) | EFMLoRA · ScienceQA (Qwen-VL-Chat) | accuracy (%): 91.6 -> 89.6 | Hurt |
Closest papers by meaning
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.93)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.90)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.90)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.89)
- NOLA: Compressing LoRA using Linear Combination of Random Basis (ICLR 2024 · 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.
