Research map / Architectures and efficiency / Pruning and efficient tuning
ReFTA: Breaking the Weight Reconstruction Bottleneck in Tensorized Parameter-Efficient Fine-Tuning
Jingjing Zheng, Anda Tang, Qiangqiang Mao, Zhouchen Lin, Yankai Cao
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 only principal tensor components (T-SVD) instead of LoRA-style adapters | ReFTA on ViT-Large · Image classification avg (5 datasets) | average accuracy: 79.99 -> 85.67 | Improved |
| ReFTA vs LoRA (r = 8) with far fewer parameters | ViT-Huge · Image classification avg (OxfordPets, StanfordCars, FGVC) | average accuracy: 75.23 -> 76.32 | Improved |
| ReFTA vs LoRA on RoBERTa-Base GLUE | ReFTA · GLUE avg (6 datasets) | average score: 85.2 (LoRA) vs 84.8 | No clear effect |
| ReFTA vs PiSSA/LoRA-PRO/LoRETTA at small rank on RoBERTa-Large | ReFTA · GLUE avg (6 datasets) | average score: 83.2 -> 88.2 (vs LoRA-PRO) | Improved |
| ReFTA (r=15) vs LoRA (r=8) on CommonsenseQA with ~94% fewer parameters | ReFTA · CommonsenseQA | accuracy: 86.11 -> 85.91 (avg) | No clear effect |
| use DCT or LSM-3 invertible transform instead of identity (ablation) | ReFTA · StanfordCars, FGVC (ViT-Large), CoLA, RTE (RoBERTa-Large) | accuracy: up to 3-5% on vision tasks | Improved |
| make the transform U0 trainable (init DCT) (ablation) | ReFTA · FGVC (ViT-Large) | accuracy | No clear effect |
| fine-tune only principal components (residual quantized) vs Gaussian-Zero init variant (ablation) | ReFTA · Quantization error at init (ViT-Large query weights) | quantization error | Improved |
| increase tensor rank R from 5 to 15 (ablation) | ReFTA on ViT-Huge · Image classification avg (OxfordPets, StanfordCars, FGVC) | average accuracy: 74.52 -> 76.32 | Improved |
| ReFTA vs LoRETTA on ViT-Large | ReFTA · Image classification avg (5 datasets) | average accuracy: 78.51 -> 85.67 | 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.
- LoRA-PRO — LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025) · better on 3 of 3 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
- Low-Rank Tensor Transitions (LoRT) for Transferable Tensor Regression (ICML 2025 · similarity 0.87)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.86)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (NeurIPS 2025 · similarity 0.86)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.86)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.86)
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.
