Research map / Architectures and efficiency / Pruning and efficient tuning
PiCa: Parameter-Efficient Fine-Tuning with Column Space Projection
Junseo Hwang, Wonguk Cho, Taesup Kim
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 |
|---|---|---|---|
| project gradients onto the column space of pre-trained weights instead of a random space (ablation) | PiCa · Commonsense Reasoning (Gemma-2B) | avg accuracy: 63.18 -> 67.60 | Improved |
| add weight-sharing across layers (rank 128, 5.1M params) versus no sharing (rank 16, 35.8M params) (ablation) | PiCa · Commonsense Reasoning (Gemma-7B) | accuracy: 35.8M -> 5.1M trainable parameters | No clear effect |
| add weight-sharing under varying ranks (ablation) | PiCa · GSM-8K (Gemma-2B) | accuracy | Improved |
| increase PiCa rank from 32 to 256 | PiCa · GSM-8K (Gemma-2B) | accuracy: 41.32 -> 52.77 | Improved |
| increase PiCa rank from 32 to 256 | PiCa · GSM-8K (LLaMA-3-8B) | accuracy: 73.54 -> 76.12 | Improved |
| increase PiCa rank from 16 to 128 | PiCa · Commonsense Reasoning (Gemma-7B) | avg accuracy: 82.83 -> 84.47 | Improved |
| use PiCa r=256 instead of SVFT-R | SVFT-R · GSM-8K (Gemma-2B) | accuracy: 50.03 -> 52.77 | Improved |
| use PiCa r=32 instead of SVFT-P | SVFT-P · GSM-8K (Gemma-2B) | accuracy: 40.34 -> 41.32 | Improved |
| use PiCa r=128 instead of LoRA r=32 | LoRA · Commonsense Reasoning (Gemma-7B) | avg accuracy: 83.69 -> 84.47 | Improved |
| use PiCa r=16 instead of SVFT-P | SVFT-P · Commonsense Reasoning (Gemma-7B) | avg accuracy: 80.85 -> 82.83 | Improved |
| use PiCa instead of BOFT | BOFT · GLUE (DeBERTaV3-base) | avg score: 89.89 -> 89.69 | No clear effect |
| use PiCa r=16 instead of LoRA r=8 | LoRA · GLUE (DeBERTaV3-base) | avg score: 88.50 -> 89.69 | Improved |
Closest papers by meaning
- LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection (NeurIPS 2026 · similarity 0.87)
- Increasing Model Capacity for Free: A Simple Strategy for Parameter Efficient Fine-tuning (ICLR 2024 · similarity 0.87)
- PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization (NeurIPS 2024 · similarity 0.87)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.87)
- Parameter Efficient Fine-tuning via Explained Variance Adaptation (NeurIPS 2025 · similarity 0.86)
- Maintaining Structural Integrity in Parameter Spaces for Parameter Efficient Fine-tuning (ICLR 2025 · similarity 0.86)
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.
