Research map / Architectures and efficiency / RNNs and normalization
The Stability-Efficiency Dilemma: Investigating Sequence Length Warmup for Training GPT Models
Conglong Li, Minjia Zhang, 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 |
|---|---|---|---|
| add sequence length warmup (SLW) at batch size 4K | GPT-2 1.5B · WikiText-103 zero-shot | PPL: 14.76 -> 13.72 | Improved |
| add SLW at batch size 4K | GPT-2 117M · WikiText-103 zero-shot | PPL: 28.09 -> 27.15 | Improved |
| add SLW at batch size 512 | GPT-2 1.5B · WikiText-103 zero-shot | PPL: 13.89 -> 13.88 | No clear effect |
| add SLW with sequence length 2K | GPT-2 117M · WikiText-103 zero-shot | PPL: 28.19 -> 26.03 | Improved |
| replace Shortformer two-stage short-then-full sequences with SLW | GPT-2 1.5B bsz4K · WikiText-103 zero-shot | PPL: 14.14 -> 13.72 | Improved |
| replace batch size warmup with SLW | GPT-2 1.5B bsz4K · WikiText-103 zero-shot | PPL: 14.21 -> 13.72 | Improved |
| add SLW enabling 40x LR, 30B tokens | GPT-3 125M · 11-task zero-shot average | average accuracy: 29.8 -> 31.1 | Improved |
| add SLW to GPT-3 1.3B at 300B tokens | GPT-3 1.3B · zero-shot average accuracy | average accuracy: 41.6 -> 41.9 | Improved |
| use discrete 2-stage pacing function instead of linear step-wise (ablation) | SLW · GPT-2 pretraining | stability/convergence | Hurt |
| use root or adaptive pacing function instead of linear (ablation) | SLW · GPT-2 pretraining | convergence | No clear effect |
| vary SLW duration T within reasonable range (ablation) | SLW on GPT-2 117M · WikiText/LAMBADA | accuracy | No clear effect |
| increase gradient clipping strength instead of SLW (ablation) | GPT-2 baseline · GPT-2 pretraining | training stability | No clear effect |
Closest papers by meaning
- PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs (ICLR 2025 · similarity 0.89)
- Analyzing & Reducing the Need for Learning Rate Warmup in GPT Training (NeurIPS 2024 · similarity 0.88)
- Critical Batch Size Revisited: A Simple Empirical Approach to Large-Batch Language Model Training (NeurIPS 2025 · similarity 0.88)
- Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training (NeurIPS 2025 · similarity 0.88)
- Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation is Wasteful (NeurIPS 2025 · similarity 0.87)
- Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 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.
