Research map / Architectures and efficiency / Pruning and efficient tuning
MoSA: Mosaic Shared Adaptation of Large Language Models
Xiequn Wang, Zhan Zhuang, Shengda Luo, Yu Zhang
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 balanced random tessellation instead of row-stripe grouping (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 87.05 -> 87.63 | Improved |
| use balanced random tessellation instead of column-stripe grouping (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 86.65 -> 87.63 | Improved |
| use balanced group sizes instead of skewed group sizes (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 74.36 -> 87.63 | Improved |
| adapt FFN projections in addition to QKV (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 87.16 -> 87.63 | Improved |
| adapt attention QKV in addition to FFN (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 87.35 -> 87.63 | Improved |
| adapt value projection instead of key projection only (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 82.12 -> 86.29 | Improved |
| reduce budget to LoRA r=1 equivalent (0.022% params) (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 80.79 -> 86.81 | No clear effect |
| raise learning rate from 1e-5 to 5e-5 (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: 87.63 -> 86.21 | Hurt |
| vary random seed of tessellation (ablation) | MoSA · Commonsense reasoning (Llama-3-8B) | average accuracy: mean 87.45, std 0.14 | No clear effect |
| use MoSA instead of HiRA | Llama-3-8B PEFT · GSM8K (trained on MetaMathQA) | accuracy: 70.81 -> 78.00 | Improved |
| use MoSA instead of HiRA | Llama-2-7B PEFT · Commonsense reasoning (8 tasks avg) | average accuracy: 81.42 -> 83.83 | Improved |
| use MoSA instead of HiRA | Llama-3-8B PEFT · ConvAI2 (average of metrics) | average score: 47.80 -> 50.14 | 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.
- LoRA-Pro — LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025) · better on 2 of 2 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
- Loquetier: A Virtualized Multi-LoRA Framework for Unified LLM Fine-tuning and Serving (NeurIPS 2025 · similarity 0.89)
- KaSA: Knowledge-Aware Singular-Value Adaptation of Large Language Models (ICLR 2025 · similarity 0.89)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.89)
- LoRASuite: Efficient LoRA Adaptation Across Large Language Model Upgrades (NeurIPS 2025 · similarity 0.89)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.89)
- GoRA: Gradient-driven Adaptive Low Rank Adaptation (NeurIPS 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.
