Research map / Architectures and efficiency / Pruning and efficient tuning
SumRA: Parameter Efficient Fine-tuning with Singular Value Decomposition and Summed Orthogonal Basis
Chin Yuen Kwok, Yongsen Zheng, Jia Qi Yip, Kwok-Yan Lam, Eng Siong Chng
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 |
|---|---|---|---|
| use SumRA summed-singular-vector init of A instead of LoRA | LoRA rank 32, whisper-large-v2 · Common Voice eo | WER: 14.42 -> 12.41 | Improved |
| use SumRA init instead of LoRA | LoRA rank 32, whisper-small · Common Voice eo | WER: 23.39 -> 20.77 | Improved |
| use SumRA instead of PiSSA | rank 32, whisper-large-v2 · Common Voice ia | WER: 8.82 -> 8.17 | Improved |
| use SumRA instead of CorDA | rank 32, whisper-large-v2 · Common Voice ia | WER: 9.18 -> 8.17 | Improved |
| interleave sum instead of block sum for summing singular vectors (ablation) | SumRA, whisper-small rank 32 · Common Voice eo | WER: 21.68 -> 20.77 | Improved |
| interleave sum instead of block sum for summing singular vectors (ablation) | SumRA, whisper-small rank 32 · Common Voice kmr | WER: 47.30 -> 44.47 | Improved |
| greedy sum instead of interleave sum (ablation) | SumRA, whisper-small rank 32 · Common Voice fy-NL and mhr | WER: 33.37 -> 33.91 (fy-NL), 36.30 -> 37.53 (mhr), 13.38 -> 13.16 (ia) | Mixed |
| train matrix A as well as B (ablation) | SumRA, whisper-small · Esperanto 10h | WER: 20.77 -> 20.14 | Improved |
| train matrix A as well as B (ablation) | SumRA, whisper-small · Esperanto 50h | WER: 14.49 -> 13.75 | Improved |
| train matrix A as well as B (ablation) | SumRA, whisper-small · Esperanto 100h | WER: 13.39 -> 13.02 | Improved |
| SumRA instead of LoRA with more data | whisper-small · Esperanto 100h | WER: 13.28 -> 13.39 | No clear effect |
| SumRA instead of LoRA with 50h data | whisper-small · Esperanto 50h | WER: 15.20 -> 14.49 | 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.
- PiSSA — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 5 of 5 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
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.89)
- PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024 · similarity 0.89)
- Batched Low-Rank Adaptation of Foundation Models (ICLR 2024 · similarity 0.89)
- Parameter-Efficient Fine-Tuning with Discrete Fourier Transform (ICML 2024 · similarity 0.89)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
- LoRA Training in the NTK Regime has No Spurious Local Minima (ICML 2024 · similarity 0.88)
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.
