Research map / Architectures and efficiency / Pruning and efficient tuning
Efficient Orthogonal Fine-Tuning with Principal Subspace Adaptation
Fei Wu, Jia Hu, Geyong Min, Shiqiang Wang
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 |
|---|---|---|---|
| confine orthogonal fine-tuning to the principal subspace (PSOFT) instead of full-space GOFTv2 | PSOFT · DeBERTaV3-base GLUE | peak memory (GB): 18.5 -> 4.1 | Improved |
| use PSOFT instead of LoRA | LLaMA-3.2-3B · GSM-8K | accuracy: 60.80 -> 63.08 | Improved |
| use PSOFT instead of FFT on MATH | LLaMA-3.2-3B · MATH | accuracy: 16.84 -> 15.98 | Hurt |
| use PSOFT instead of LoRA | LLaMA-3.1-8B · Commonsense reasoning avg | average accuracy: 81.11 -> 82.54 | Improved |
| use PSOFT instead of LoRA on ViT-B/16 | ViT-B/16 · VTAB-1K | average accuracy: 71.8 -> 73.4 | Improved |
| orthogonal initialization of A (A_orth R B) instead of PiSSA initialization (ablation) | PSOFT · RTE (DeBERTaV3-base) | accuracy: 71.11 -> 85.92 | Improved |
| enforce orthogonality on B instead of A at initialization (ablation) | PSOFT · RTE (DeBERTaV3-base) | accuracy: 85.92 -> 52.71 | Hurt |
| enable both tunable vectors alpha and beta (ablation) | PSOFT · LLaMA-3.2-3B rank 64, GSM-8K and MATH | accuracy | Improved |
| strict orthogonality via Cayley parameterization versus unconstrained R (gamma=0) (ablation) | PSOFT · ablation on orthogonality | accuracy | Improved |
| add DoRA weight decomposition to LoRA | LoRA · ViT-B/16 VTAB-1K | peak memory (GB): 9.9 -> 17.8 | Hurt |
Closest papers by meaning
- LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection (NeurIPS 2026 · similarity 0.92)
- Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation (ICML 2024 · similarity 0.89)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.88)
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.87)
- Large Convolutional Model Tuning via Filter Subspace (ICLR 2025 · similarity 0.87)
- Parameter-Efficient Orthogonal Finetuning via Butterfly Factorization (ICLR 2024 · similarity 0.87)
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.
