Research map / Architectures and efficiency / Pruning and efficient tuning
LoRA variants and theory
Improves low-rank adaptation through new parameterizations, optimization fixes, and expressiveness analysis.
121 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Pruning and efficient tuning, keyword cluster “fine tuning · lora · adaptation”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 3 tracked benchmarks.
21 papers in 2023–24 and 99 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 7.6% to 17.8%.
Open questions the data points to
- CIFAR-100 (accuracy): best went from 87.17 (up to 2024) to 91.96 in 2025 — From Weight-Based to State-Based Fine-Tuning: Further Memory Reduction on LoRA with Parallel Control (ICML 2025)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2022 · LoRA: Low-Rank Adaptation of Large Language Models (ICLR 2022)
- 2024 · PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024 · 72 citations)
- 2025 · LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · 10 citations)
- 2026 · Group Orthogonal Low-Rank Adaptation for RGB-T Tracking (AAAI 2026 · 1 citation)
Earliest papers here
- LoRA: Low-Rank Adaptation of Large Language Models (ICLR 2022)
- PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024 · 72 citations)
- DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024 · 38 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
CIFAR-100 · accuracy (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 87.17 | Parameter-Efficient Fine-Tuning with Controls (ICML 2024) |
| 2025 | 91.96 | From Weight-Based to State-Based Fine-Tuning: Further Memory Reduction on LoRA with Parallel Control (ICML 2025) |
E2E, GPT-2 Large · BLEU (higher is better)
4 papers, 2024–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 70.3 | VeRA: Vector-based Random Matrix Adaptation (ICLR 2024 · 8 citations) |
| 2025 | 70.4 | LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning (ICLR 2025 · 0 citations) |
E2E, GPT-2 Medium · BLEU (higher is better)
4 papers, 2024–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 70.1 | VeRA: Vector-based Random Matrix Adaptation (ICLR 2024 · 8 citations) |
What has been tried
60 papers here report 703 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 10%, hurt 21%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use principal instead of medium or minor singular values and vectors to initialize adapters | GSM8K and MATH · accuracy | Helped | PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024 · 72 citations) | |
| halve the LoRA rank in DoRA | Commonsense reasoning (8-task average) · accuracy | 74.7 -> 77.5 | Helped | DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024 · 38 citations) |
| Add task-specific low-rank modules to task-agnostic LoRA | PASCAL-Context MTL · accuracy-efficiency trade-off | Helped | MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · 33 citations) | |
| Use Init 1 instead of Init 2 | MMLU · accuracy | 0.6% | Helped | LoRA+: Efficient Low Rank Adaptation of Large Models (ICML 2024 · 18 citations) |
| Change subspace-switching frequency to be too frequent or too slow | C4 · convergence | Hurt | GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024 · 11 citations) | |
| Scale down top intruder dimensions with λ = 0.7 | QQP · pre-training loss change (%) | -33.2 | Helped | LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · 10 citations) |
| Add the b scaling vector | MRPC, RoBERTa-large · score | 89.7 -> 90.5 | Helped | VeRA: Vector-based Random Matrix Adaptation (ICLR 2024 · 8 citations) |
| replace the Fourier basis with a random basis | GLUE RTE, RoBERTa Base · accuracy | 79.1 -> 72.7 | Hurt | Parameter-Efficient Fine-Tuning with Discrete Fourier Transform (ICML 2024 · 5 citations) |
| Replace the structured mask with a random mask | Paintings, α=1 · HPSv2 score | 31.2 ± 1.7 -> 30.7 ± 1.9 | Hurt | Sparse High Rank Adapters (NeurIPS 2024 · 3 citations) |
| add orthogonality-aware magnitude weighting | GLUE · average score | 89.31 -> 89.79 | Helped | Unveiling LoRA Intrinsic Ranks via Salience Analysis (NeurIPS 2024 · 3 citations) |
| Add Occasional Low-cost SVD (Eqn. 7) | CIFAR-100 pre-training · Top-1 accuracy | 63.28 -> 70.39 | Helped | COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection (CVPR 2025 · 2 citations) |
| use the concatenated reference basis instead of the first task's basis | ViT-B/32 vision tasks · average accuracy | 60.4 -> 68.6 | Helped | Accurate and Efficient Low-Rank Model Merging in Core Space (NeurIPS 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 62 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GLUE · 15 of 62 papers
- MetaMathQA · 13 of 62 papers
- CIFAR-100 · 8 of 62 papers
- DreamBooth · 8 of 62 papers
- DTD · 7 of 62 papers
- CodeFeedback · 7 of 62 papers
- GSM8K · 7 of 62 papers
- CIFAR-10 · 6 of 62 papers
Baselines
- LoRA · 40 of 62 papers
- DoRA · 17 of 62 papers
- AdaLoRA · 14 of 62 papers
- PiSSA · 13 of 62 papers
- VeRA · 10 of 62 papers
- LoRA-GA · 8 of 62 papers
- rsLoRA · 7 of 62 papers
- Full fine-tuning · 7 of 62 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 120 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 4 papers
- Hong Kong University of Science and Technology · 4 papers
- Tsinghua University · 4 papers
- Fudan University · 3 papers
- National University of Singapore · 3 papers
- Tencent · 3 papers
- Meta · 3 papers
- Shanghai Jiao Tong University · 3 papers
- Peking University · 3 papers
- Amazon · 3 papers
Use it
Download BibTeX of 20 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Low-rank compression and merging (226 papers)
- Adapter-based parameter-efficient tuning (149 papers)
- Parameter-efficient transfer analysis (97 papers)
- Efficient ViT adaptation (62 papers)
Elsewhere in the topic
- Structured network pruning (162 papers)
- Sparse training from scratch (95 papers)
- Layer-wise pruning of LLMs (91 papers)
- Joint pruning and quantization (78 papers)
- Lottery ticket discovery (42 papers)
- Pruning at initialization (24 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).