Research map / Architectures and efficiency / Pruning and efficient tuning
Exploring extreme parameter compression for pre-trained language models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply tensor-train decomposition at rank 384 | BERT-base · GLUE test | all-task average score: 82.7 -> 83.2 | Helped |
| Apply Tucker decomposition at ranks 72-384 | BERT-base · GLUE test | all-task average score: 82.7 -> 82.6 | No clear effect |
| Use Tucker decomposition with ranks 36-128 instead of 72-384 | BERT-base · GLUE test | all-task average score: 82.6 -> 80.8 | Hurt |
| Reduce tensor-train dimension rank from 384 to 64 | BERT-base · GLUE test | RPS: 420.1 -> 1143.6 | Helped |
| Apply Tucker decomposition at ranks 72-384 | TinyBERT-6layer · GLUE test | all-task average score: 83.0 -> 81.6 | Hurt |
| Remove task-specific distillation (ablation) | BERT-IV-72-384 · GLUE test | all-task average score: 82.6 -> 80.0 | Hurt |
| Remove knowledge distillation (ablation) | BERT-IV-72-384 · GLUE test | all-task average score: 82.6 -> 47.3 | Hurt |
| Decompose FFN only rather than jointly decomposing SAN and FFN (ablation) | BERT-IV · GLUE test | all-task average score: 82.6 -> 83.1 | Helped |
| Decompose SAN only rather than jointly decomposing SAN and FFN (ablation) | BERT-IV · GLUE test | all-task average score: 82.6 -> 82.4 | Hurt |
| Add a regularizer to make reconstructed weights match raw weights (ablation) | BERT-IV · GLUE | performance | Hurt |
| Increase Tucker layer and dimension ranks (ablation) | BERT-IV · GLUE | performance | Helped |
Closest papers by meaning
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · similarity 0.88)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.88)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.87)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.87)
- Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models (NeurIPS 2024 · similarity 0.87)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 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.