Research map / Architectures and efficiency / Pruning and efficient tuning
Multi-Task Low-Rank Model Adaptation
Zichen Tian, Antoine Ledent, Qianru Sun
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 block-level adaptation (ablation) | HydraLoRA · average of DOTA, iNat2018, Dolly-15k, BBH | average accuracy (%): 61.1 -> 63.0 | Improved |
| apply block-level adapters to attention and FFN instead of Wq, Wv (ablation) | HydraLoRA · average of DOTA, iNat2018, Dolly-15k, BBH | average accuracy (%): 61.1 -> 63.1 | Improved |
| add spectral-aware regularization (ablation) | mtLoRA block-level · average of DOTA, iNat2018, Dolly-15k, BBH | average accuracy (%): 63.2 -> 63.8 | Improved |
| add fine-grained routing (ablation) | mtLoRA block-level · average of DOTA, iNat2018, Dolly-15k, BBH | average accuracy (%): 63.2 -> 63.1 | Mixed |
| add all three designs (ablation) | HydraLoRA · average of DOTA, iNat2018, Dolly-15k, BBH | average accuracy (%): 61.1 -> 63.9 | Improved |
| increase routing groups from g=1 to g=32 (ablation) | HydraLoRA · BBH | accuracy (%): 35.5 -> 37.7 | Improved |
| increase routing groups from g=1 to g=2 (ablation) | HydraLoRA · BBH | accuracy (%): 35.5 -> 37.0 | Improved |
| add orthogonal regularization without routing (ablation) | HydraLoRA uniform routing · DOTA | accuracy (%): 20.5 | Hurt |
| add orthogonal regularization with dynamic routing (ablation) | HydraLoRA · DOTA | accuracy (%): 89.0 -> 89.8 | Improved |
| add sparsity regularization with dynamic routing (ablation) | HydraLoRA · DOTA | accuracy (%): 89.0 -> 87.9 | Hurt |
| raise orthogonal regularization strength from 0.25 to 1.0 (ablation) | HydraLoRA with routing · Flan-v2 to BBH | accuracy (%): -1.8% | Hurt |
| use block-level instead of component-level adapters (ablation) | LoRA · gradient cosine similarity between tasks | cosine similarity: -0.054 -> -0.013 | Improved |
Closest papers by meaning
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.91)
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.90)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.90)
- RepLoRA: Reparameterizing Low-rank Adaptation via the Perspective of Mixture of Experts (ICML 2025 · similarity 0.89)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
- LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
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.
