Research map / Architectures and efficiency / Pruning and efficient tuning
Bi-LoRA: Efficient Sharpness-Aware Minimization for Fine-Tuning Large-Scale Models
Yuhang Liu, Tao Li, Zhehao Huang, Zuopeng Yang, Xiaolin Huang
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 auxiliary adversarial LoRA module (Bi-LoRA) to LoRA | LoRA, Llama 2-7B · GSM8K | accuracy (%): 58.21 -> 60.32 | Improved |
| add auxiliary adversarial LoRA module (Bi-LoRA) to LoRA | LoRA, Llama 2-7B · HumanEval | pass@1: 24.75 -> 27.20 | Improved |
| add auxiliary adversarial LoRA module (Bi-LoRA) to LoRA | LoRA, Llama 3.1-8B · DROP | score (%): 49.82 -> 51.53 | Improved |
| Bi-LoRA single-pass instead of LoRA-SAM two-step | LoRA-SAM, Llama 3.1-8B · training time per step | relative time: 215% -> 109% of LoRA | Improved |
| apply SAM to LoRA parameters (LoRA-SAM) | LoRA · Llama 2-7B math/code/chat | accuracy: GSM8K 58.21 -> 59.16; HumanEval 24.75 -> 26.59 | Mixed |
| random perturbation (ROP/RWP) instead of adversarial | LoRA · Llama 2-7B tasks | accuracy: GSM8K 58.21 -> 59.24 (ROP) | No clear effect |
| make auxiliary branch descend like primary (Dual-LoRA) instead of adversarial ascent (ablation) | Bi-LoRA · Llama 3.1-8B instruction following | average score: 49.79 -> 49.86 | No clear effect |
| add global clipping on auxiliary module (vs no clipping) (ablation) | Bi-LoRA · Llama 3.1-8B instruction following | average score: 47.71 -> 51.19 | Improved |
| use global instead of per-layer clipping (ablation) | Bi-LoRA · Llama 3.1-8B instruction following | average score: 50.03 -> 51.19 | Improved |
| vary auxiliary learning rate eta2 from 1e-4 to 1e-3 (ablation) | Bi-LoRA · Llama 3.1-8B instruction following | average score: 50.55 to 51.19 range | No clear effect |
| increase auxiliary rank r2 beyond 8 (ablation) | Bi-LoRA · Llama 3.1-8B instruction following | average score | No clear effect |
| combine Bi-LoRA with LoRA-GA, PiSSA, DoRA | LoRA variants · T5-base MRPC+CoLA | average score: DoRA 74.35 -> 75.16 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Flat-LoRA — Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.90)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- LoRA Training Provably Converges to a Low-Rank Global Minimum Or It Fails Loudly (But it Probably Won't Fail) (ICML 2025 · similarity 0.90)
- Implicit Regularization of Sharpness-Aware Minimization for Scale-Invariant Problems (NeurIPS 2024 · similarity 0.89)
- AuroRA: Breaking Low-Rank Bottleneck of LoRA with Nonlinear Mapping (NeurIPS 2025 · similarity 0.89)
- LoRA+: Efficient Low Rank Adaptation of Large Models (ICML 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.
