Research map / Architectures and efficiency / Pruning and efficient tuning
SALAAD: Sparse And Low-Rank Adaptation via ADMM for Large Language Model Inference
Hao Ma, Melis Ilayda Bal, Liang Zhang, Bingcong Li, Niao He, Melanie Zeilinger, Michael Muehlebach
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 SALAAD instead of full-rank training | LLaMA-based 1B · C4 | perplexity: 15.56 -> 14.74 | Improved |
| use SALAAD instead of SLTrain at similar budget | LLaMA-based 350M · C4 | perplexity: 19.42 -> 18.04 (L+S) | Improved |
| use SALAAD instead of LOST | LLaMA-based 350M, HPA-compressed · C4 | perplexity: 18.95 -> 19.70 | Hurt |
| add HPA compression of SALAAD model to 646M | LLaMA-based 1B · C4 | perplexity: 14.72 -> 15.07 (L+S to eL+eS) | Mixed |
| include embedding layer in SALAAD training (ablation) | LLaMA-based 350M · C4 | training loss | No clear effect |
| include LM head in SLR induction (ablation) | SALAAD · LLaMA-based | SLR structure | Hurt |
| apply SLR induction at training time instead of post-hoc RPCA on vanilla | HPA elastic deployment · C4, 60M-1B | perplexity vs parameters | Improved |
| assign larger share of budget to low-rank component (kappa > 0.5) (ablation) | HPA allocation · C4, 130M/350M/1B | perplexity | Improved |
| increase delta_beta from 0.003 to 0.1 (ablation) | SALAAD · C4, LLaMA 350M | perplexity (X): 18.62 -> 19.97 | Hurt |
| increase delta_alpha from 0.08 to 0.2 (ablation) | SALAAD · C4, LLaMA 350M | perplexity (X): 18.25 -> 18.97 | Hurt |
| increase rho from 5e-8 to 1e-7 (ablation) | SALAAD · C4, LLaMA 350M | perplexity (X): 18.45 -> 19.53 | Hurt |
| compress SALAAD 1B to 646M | SALAAD · zero-shot downstream, 1B | accuracy: COPA 71.0 -> 69.0; BoolQ 54.0 -> 52.0 | No clear effect |
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.
- SLTrain — SLTrain: a sparse plus low rank approach for parameter and memory efficient pretraining (NeurIPS 2024) · better on 4 of 4 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
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.89)
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.89)
- LDAdam: Adaptive Optimization from Low-Dimensional Gradient Statistics (ICLR 2025 · similarity 0.88)
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML 2024 · similarity 0.88)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.88)
- La RoSA: Enhancing LLM Efficiency via Layerwise Rotated Sparse Activation (ICML 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.
