Research map / Architectures and efficiency / Quantization
LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization
Haoyu Wang, Xingyu Yu, Haiyan Zhao, Fengxiang Wang, Xu Han
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 |
|---|---|---|---|
| initialize QAT from the LC-PTQ result instead of GPTQ or random init (ablation) | LC-QAT · Zero-shot QA avg, Qwen-3-8B | average accuracy: 37.53 (GPTQ init) -> 67.95 (LC-PTQ init); random init 39.51 | Improved |
| train end-to-end with LC-QAT instead of PV-Tuning coordinate descent from the same LC-PTQ init (ablation) | LC-PTQ model · Zero-shot QA avg, Qwen-3-1.7B | average accuracy: 55.35 -> 59.75 | Improved |
| add integer weight preprocessing (Xavier-aligned proxy weight init) (ablation) | LC-QAT · training loss | training loss | Improved |
| use Differentiable Gradient Estimator instead of STE (ablation) | LC-QAT · training loss | loss spikes | Improved |
| use LC-QAT instead of PV-Tuning (same data) | 2-bit VQ-QAT · Zero-shot QA avg, Qwen-3-1.7B | average accuracy: 58.05 -> 61.43 | Improved |
| use LC-QAT with 4B tokens instead of ParetoQ with 30B tokens | LLaMA-3-8B 2-bit · Zero-shot QA avg | average accuracy: 72.07 -> 73.42 | Improved |
| LC-QAT vs ParetoQ on perplexity | LLaMA-3-8B 2-bit · WikiText-2 test | perplexity: 8.00 (ParetoQ) vs 9.38 (LC-QAT) | Hurt |
| LC-QAT 1.7B instruct vs BitNet 2B4T | instruction-following 2-bit model · OpenBookQA/IFEval/MMLU/GSM8K/MATH/HumanEval avg | average score: 48.07 -> 50.35 average; MMLU 53.17 -> 46.77 | Mixed |
| increase training budget from 1B to 10B tokens (ablation) | LC-QAT · zero-shot QA | accuracy | Improved |
| vector dimension d=8 instead of d=4 (ablation) | LC-PTQ initialization · zero-shot QA | accuracy | 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.
- PV-Tuning — PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (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
- HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs (NeurIPS 2026 · similarity 0.91)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.90)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.89)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.89)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.88)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · 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.
