Research map / Architectures and efficiency / Pruning and efficient tuning
A3: an Analytical Low-Rank Approximation Framework for Attention
Jeffrey T. H. Wong, Cheng Zhang, Xinye Cao, Pedro Gimenes, Christos-Savvas Bouganis, George Anthony Constantinides, Wayne Luk, Yiren Zhao
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 A3 instead of SVD-LLM | A3 · WikiText-2 perplexity, LLaMA-3.1-70B (10% compression) | perplexity: 7.87 -> 4.69 | Improved |
| use A3 instead of SVD-LLM | A3 · WikiText-2 perplexity, LLaMA-3.1-8B (20% compression) | perplexity: 42.28 -> 11.36 | Improved |
| use A3 instead of SVD-LLM | A3 · Downstream average accuracy, LLaMA-3.1-70B (10% compression) | accuracy: 0.6797 -> 0.7508 | Improved |
| use A3 instead of SVD-LLM at higher compression | A3 · C4 perplexity, LLaMA-3.1-70B (20% compression) | perplexity: 13.77 -> 13.94 | Hurt |
| use A3 instead of SVD-LLM at 20% compression | A3 · Downstream accuracy, LLaMA-3.1-70B (20% compression, ARC-c) | accuracy: 0.4957 -> 0.4667 | Hurt |
| use A3 instead of Clover (MPT-7B, 20%) | A3 · WikiText-2 perplexity, MPT-7B | perplexity: 77.78 -> 8.05 | Improved |
| use A3 instead of Palu (MPT-7B, 60%) | A3 · WikiText-2 perplexity, MPT-7B (60% compression) | perplexity: 25.09 -> 15.58 | Improved |
| use A3 instead of Palu (MPT-7B, 80%) | A3 · WikiText-2 perplexity, MPT-7B (80% compression) | perplexity: 13714 -> 849 | Improved |
| use A3 instead of SVD-LLM with SlimPajama calibration (LLaMA-7B, 40% compression) | A3 · Average WikiText-2/C4 perplexity, LLaMA-7B (40%) | perplexity: 30.84 -> 25.22 | Improved |
| use A3 instead of SVD-LLM with WikiText-2 calibration (LLaMA-7B, 40%) | A3 · Average WikiText-2/C4 perplexity, LLaMA-7B (40%) | perplexity: 44.58 -> 12.82 | Improved |
| use activation autocorrelation weighting (A3-QK) instead of identity (CLOVER) (ablation) | A3-QK · MPT-7B QK component (20% compression) | delta PPL | Improved |
| use A3-OV instead of Palu (ablation) | A3-OV · MPT-7B OV component (20% compression) | delta PPL | 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.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 16 of 16 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
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.91)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.90)
- LoRAP: Transformer Sub-Layers Deserve Differentiated Structured Compression for Large Language Models (ICML 2024 · similarity 0.90)
- Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix (ICLR 2025 · similarity 0.89)
- LQ-LoRA: Low-rank plus Quantized Matrix Decomposition for Efficient Language Model Finetuning (ICLR 2024 · similarity 0.88)
- SALS: Sparse Attention in Latent Space for KV Cache Compression (NeurIPS 2025 · similarity 0.88)
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.
