Research map / Architectures and efficiency / Quantization
STLA: Spatiotemporal Lookahead Alignment for Post-Training Quantization
Zuqi Zhang, Chenghe Sun, Xiangyi Chu, Wei-Han Yu, Ka-Fai Un, Rui P. Martins, Pui-In Mak, Jiawei Xu
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 |
|---|---|---|---|
| couple learning-based rounding with compensation (channel-wise) (ablation) | STLA · OPT-125M C4, channel-wise | perplexity: 31.21 -> 30.69 | Improved |
| couple learning and compensation (channel-wise, local loss) (ablation) | STLA · OPT-125M C4, channel-wise | perplexity: 31.21 -> 32.63 | Hurt |
| use lookahead global objective instead of local loss (ablation) | STLA coupled · OPT-125M C4, channel-wise | perplexity: 32.63 -> 30.69 | Improved |
| use lookahead global objective instead of local loss (ablation) | STLA coupled, group 256 · OPT-125M WikiText-2 | perplexity: 35.86 -> 32.00 | Improved |
| use global objective instead of local loss (ablation) | STLA coupled, group 256 · LLaMA2-7B WikiText-2 | perplexity: 6.11 -> 6.44 | Hurt |
| add SpinCluster ordering (ablation) | STLA coupled, global loss, group 256 · OPT-125M C4 | perplexity: 29.86 -> 29.26 | Improved |
| add SpinCluster ordering (ablation) | STLA coupled, global loss, group 256 · OPT-125M WikiText-2 | perplexity: 32.00 -> 31.01 | Improved |
| add SpinCluster ordering (ablation) | STLA coupled, global loss, group 256 · LLaMA2-7B WikiText-2 | perplexity: 6.44 -> 6.09 | Improved |
| couple learning and compensation in cluster-wise fashion (ablation) | STLA · LLaMA2-7B GPU runtime | quantization time: 2.3 HR -> 32.8 MIN | Improved |
| use group-wise (size 256) instead of channel-wise quantization (ablation) | STLA · OPT-125M C4 | perplexity: 30.69 -> 29.86 | Improved |
| replace GPTAQ with STLA in the rotation pipeline | QuaRot · LLaMA2-7B W2A4 WikiText-2 | perplexity: 16.90 -> 11.42 | Improved |
| replace GPTAQ with STLA in the rotation pipeline | SpinQuant · LLaMA3-8B W3A4 WikiText-2 | perplexity: 9.30 -> 8.91 | 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.
- GPTAQ — GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration (ICML 2025) · better on 9 of 9 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
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.88)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.87)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.87)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.87)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.87)
- ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs (NeurIPS 2026 · 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.
