Research map / Architectures and efficiency / Quantization
XTC: Extreme Compression for Pre-trained Transformers Made Simple and Efficient
Xiaoxia Wu, Zhewei Yao, Minjia Zhang, Conglong Li, Yuxiong He
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 |
|---|---|---|---|
| Increase the training budget from Budget-A to Budget-C (ablation) | BERT-base with 1-bit quantization · GLUE | Avg. all: 82.38 -> 83.44 | Improved |
| Use one-stage KD with learning-rate search instead of multi-stage KD (ablation) | BERT-base with 1-bit quantization, Budget-A · GLUE | Avg all: 81.59 -> 82.38 | Improved |
| Remove data augmentation (ablation) | BERT-base with 1-bit quantization, Budget-A · GLUE | Avg. all: 82.38 -> 81.61; -0.77 | Hurt |
| Remove data augmentation (ablation) | BERT-base with 1-bit quantization, Budget-C · GLUE | Avg. all: 83.44 -> 82.78; -0.66 | Hurt |
| Use lightweight layer reduction initialized from the fine-tuned teacher instead of pre-training distillation (ablation) | Task-specific BERT student · GLUE | Avg. all: 83.11 -> 83.50 | Mixed |
| Reduce the Skip-BERT student from 6 layers to 5 (ablation) | Skip-BERT · GLUE | Avg. all: 83.50 -> 83.38 | Hurt |
| Reduce the Skip-BERT student from 6 layers to 4 (ablation) | Skip-BERT · GLUE | Avg. all: 83.50 -> 82.18 | Hurt |
| Reduce quantized student from 6 layers to 5 (ablation) | XTC with 2-bit quantization · GLUE | Avg.: 82.89 -> 82.46 | Hurt |
| Reduce quantized student from 6 layers to 4 (ablation) | XTC with 2-bit quantization · GLUE | Avg.: 82.89 -> 81.22 | Hurt |
| Reduce quantized student from 6 layers to 5 (ablation) | XTC with 1-bit quantization · GLUE | Avg.: 81.71 -> 81.34 | Hurt |
| Reduce quantized student from 6 layers to 4 (ablation) | XTC with 1-bit quantization · GLUE | Avg.: 81.71 -> 79.96 | Hurt |
| Add low-rank full-precision weight matrices (ablation) | XTC 1-bit layer-reduced models · GLUE | average GLUE score | Hurt |
Closest papers by meaning
- Extreme Compression of Large Language Models via Additive Quantization (ICML 2024 · similarity 0.88)
- PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (NeurIPS 2024 · similarity 0.87)
- Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression (NeurIPS 2025 · similarity 0.87)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.87)
- ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization (AAAI 2025 · similarity 0.86)
- BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments (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.
