Research map / Architectures and efficiency / Quantization
SineLoRA∆: Sine-Activated Delta Compression
Cameron Gordon, Yiping Ji, Hemanth Saratchandran, Paul Albert, Simon Lucey
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 post-quantization sine activation | LoRA rank 8, 2-bit · Commonsense Reasoning avg (LLaMA-3-8B) | accuracy (%): 75.2 -> 76.4 | Improved |
| add post-quantization sine activation | LoRA rank 2, 2-bit · Commonsense Reasoning avg (LLaMA-3-8B) | accuracy (%): 71.0 -> 73.7 | Improved |
| add post-quantization sine activation | LoRA rank 8, 5-bit · Commonsense Reasoning avg (LLaMA-3-8B) | accuracy (%): 76.7 -> 78.1 | Improved |
| add sine activation at full precision | LoRA rank 1, full · Commonsense Reasoning avg (LLaMA-3-8B) | accuracy (%): 73.7 -> 72.8 | Hurt |
| add sine activation at full precision | LoRA rank 4, full · Commonsense Reasoning avg (LLaMA-3-8B) | accuracy (%): 76.5 -> 78.5 | Improved |
| add sine activation, 2-bit quantization | LoRA · Commonsense Reasoning BD analysis | BD-Rate (%): -41.60% | Improved |
| add sine activation at 1-bit quantization | LoRA rank 16, 1-bit · CLIP few-shot avg 11 tasks | accuracy (%): 72.9 -> 68.3 | Hurt |
| add sine activation at 2-bit quantization | LoRA rank 16, 2-bit · CLIP few-shot avg 11 tasks | accuracy (%): 78.1 -> 77.4 | Hurt |
| add sine activation at 3-bit quantization | LoRA rank 16, 3-bit · CLIP few-shot avg 11 tasks | accuracy (%): 79.2 -> 79.5 | Improved |
| add sine activation at full precision | LoRA rank 16, full · CLIP few-shot avg 11 tasks | accuracy (%): 79.5 -> 80.3 | Improved |
| add sine activation, 1-bit | LoRA · DreamBooth text-to-image | DINO: 0.515 -> 0.554 | Improved |
| add sine activation, 1-bit BD analysis | LoRA · CLIP few-shot BD analysis | BD-Rate (%): 227.84% | Hurt |
Closest papers by meaning
- Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression (NeurIPS 2025 · similarity 0.88)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.86)
- SLiM: One-shot Quantization and Sparsity with Low-rank Approximation for LLM Weight Compression (ICML 2025 · similarity 0.86)
- Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models (NeurIPS 2024 · similarity 0.86)
- PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (NeurIPS 2024 · similarity 0.86)
- Benchmarking and Enhancing VLM for Compressed Image Understanding (ICML 2026 · similarity 0.85)
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.
