Research map / Architectures and efficiency / Pruning and efficient tuning
LAVA: A Unified Framework for Finetuning Language and Vision Models
Daorui Ding, Fanhua Shang, Tiancan Feng, Junkang Liu, Hongying Liu
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 orthogonal regularization to attention LoRA (ablation) | SDXL attention LoRA · pokemon-blip-captions | CLIP score: 29.28 -> 29.42 | Improved |
| add orthogonal regularization to reshape-free conv adapter (ablation) | LAVA on SDXL · pokemon-blip-captions | CLIP score: 29.62 -> 29.64 | Improved |
| add orthogonal regularization to both attention and conv (ablation) | LAVA on SDXL · pokemon-blip-captions | CLIP score: 29.62 -> 29.96 | Improved |
| tune conv layers with reshape-involved LoRA instead of leaving them frozen (ablation) | SDXL attention LoRA · pokemon-blip-captions | CLIP score: 29.28 -> 28.10 | Hurt |
| replace reshape-involved LoRA with reshape-free tensor decomposition for conv (ablation) | SDXL full U-Net · pokemon-blip-captions | CLIP score: 28.10 -> 29.62 | Improved |
| tune conv layers reshape-free in addition to attention (ablation) | SDXL attention LoRA · pokemon-blip-captions | CLIP score: 29.28 -> 29.62 | Improved |
| reshape-free instead of reshape-involved low-rank conv finetuning (ablation) | six CNNs · CIFAR10 | accuracy: around 10% higher with fewer trainable parameters | Improved |
| use LAVA instead of LoRA | LLaMA2-7b · commonsense reasoning | average accuracy: 77.6 -> 79.5 | Improved |
| use LAVA instead of LoRA | RoBERTa-base · GLUE | average score: 84.7 -> 85.9 | Improved |
| use LAVA instead of DoRA | LLaMA3-3b · commonsense reasoning | average accuracy: 79.5 -> 79.8 | No clear effect |
| use LAVA instead of LoRA for convolution layers of Depth-Anything | Depth-Anything-large · NYU Depth v2 | delta1: 0.937 -> 0.972 | Improved |
| use LAVA instead of full finetuning | Depth-Anything-large · NYU Depth v2 | delta1: 0.984 -> 0.972 | Hurt |
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.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · 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
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.89)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.89)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.88)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · 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.
