Research map / Architectures and efficiency / Quantization
Rethinking 1-bit Optimization Leveraging Pre-trained Large Language Models
Zhijun Tu, Jian Li, Yuanyuan Xi, Siqi Liu, Chuanjian Liu, Hanting Chen, Jie Hu, Yunhe Wang
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 consistent progressive training to vanilla 1-bit (ablation) | 1-bit SmolLM-135M · Wikitext2 | perplexity: 30.4 -> 27.4 | Improved |
| add consistent progressive training to vanilla 1-bit (ablation) | 1-bit SmolLM-135M · Zero-shot 7-task avg | accuracy: 41.0 -> 42.0 | Improved |
| add IR-Net style progressive training to vanilla 1-bit (ablation) | 1-bit SmolLM-135M · Wikitext2 | perplexity: 30.4 -> 29.8 | Improved |
| use dual scaling S_l x S_a instead of S_a only (ablation) | 1-bit Pythia-70M · Wikitext2 | perplexity: 73.1 -> 68.7 | Improved |
| use dual scaling S_l x S_a instead of S_a only (ablation) | 1-bit Pythia-70M · Average of Wiki2, C4, PTB | perplexity: 75.5 -> 70.9 | Improved |
| replace fixed S_a with learnable S_l initialized from S_a (ablation) | 1-bit Pythia-70M · Average of Wiki2, C4, PTB | perplexity: 75.5 -> 77.0 | Hurt |
| start from stronger full-precision model (SmolLM-135M instead of Pythia-160M) (ablation) | BinaryLLM · Zero-shot 7-task avg | accuracy: Pythia +0.3 vs SmolLM -6.7 average change versus FP16 | Mixed |
| use quantization-aware BinaryLLM instead of 1-bit PTQ ARB-LLM-X | LLaMA3-1B · Wikitext2 | perplexity: 115.4 -> 14.7 | Improved |
| use quantization-aware BinaryLLM instead of 1-bit PTQ ARB-LLM-X | LLaMA3-3B · Wikitext2 | perplexity: 49.2 -> 12.4 | Improved |
| initialize 1-bit model from pre-trained SmolLM (20B tokens) instead of training FBI-LLM from scratch (1.26T tokens) | 130M-scale 1-bit LLM · Zero-shot 7-task avg | accuracy: 40.4 -> 42.8 | 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.
- OneBit — OneBit: Towards Extremely Low-bit Large Language Models (NeurIPS 2024) · better on 1 of 1 shared result
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
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.93)
- OneBit: Towards Extremely Low-bit Large Language Models (NeurIPS 2024 · similarity 0.92)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.92)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.92)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.92)
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.91)
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.
