Research map / Architectures and efficiency / Pruning and efficient tuning
Composing Parameter-Efficient Modules with Arithmetic Operation
Jinghan Zhang, Shiqi Chen, Junteng Liu, Junxian He
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 |
|---|---|---|---|
| merge two LoRA modules trained on different subsets | LoRA RoBERTa-base · GLUE RTE | accuracy: +5.2 | Improved |
| merge two (IA)3 modules trained on different subsets | (IA)3 RoBERTa-base · GLUE RTE | accuracy: +4.0 | Improved |
| merge two LoRA modules | LoRA RoBERTa-base · GLUE SST-2 | accuracy: +0.2 | Improved |
| negate toxic LoRA (lambda=1) instead of toxic LoRA | GPT-2 large · Civil Comments toxicity | toxicity score: 0.43 -> 0.01 | Improved |
| negate toxic (IA)3 | GPT-2 large · toxicity | toxic generation (%): 20.5 -> 0.9 | Improved |
| negate toxic LoRA | GPT-2 large · WikiText-103 | perplexity: 16.44 -> 16.67 | Hurt |
| simply negate all LoRA parameters (ablation) (ablation) | LoRA GPT-2 · toxicity | toxicity score: 0.43 -> 0.43 | No clear effect |
| simply negate all (IA)3 parameters (lambda=1) (ablation) | (IA)3 GPT-2 · WikiText-103 | perplexity: 16.44 -> 843.83 | Hurt |
| merge LoRA on multi-task MNLI+RTE | LoRA T5-base · MNLI/RTE | individual task accuracy | Hurt |
| domain transfer by arithmetic merge for LoRA | LoRA T5-small · Yelp test | accuracy: 94.76 -> 95.83 | Improved |
| domain transfer by arithmetic merge for LoRA | LoRA T5-small · Amazon test | accuracy: 91.03 -> 91.94 | Improved |
| domain transfer by arithmetic merge for LoRA | LoRA T5-base · Amazon test | accuracy: 94.50 -> 94.50 | No clear effect |
Closest papers by meaning
- 3-in-1: 2D Rotary Adaptation for Efficient Finetuning, Efficient Batching and Composability (NeurIPS 2024 · similarity 0.88)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.88)
- HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning (NeurIPS 2024 · similarity 0.87)
- VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks (NeurIPS 2024 · similarity 0.87)
- Gradient Weight-normalized Low-rank Projection for Efficient LLM Training (AAAI 2025 · similarity 0.87)
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (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.
