Research map / Architectures and efficiency / RNNs and normalization
Pre-training under infinite compute
Konwoo Kim, Suhas Kotha, Percy Liang, Tatsunori Hashimoto
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 |
|---|---|---|---|
| tune weight decay about 30x larger than default and jointly tune lr and epochs | standard recipe 200M tokens · DCLM validation loss | loss asymptote: 3.75 -> 3.43 | Improved |
| increase epoch count too far | 300M model, 200M tokens · DCLM validation loss | loss | Hurt |
| increase parameter count to 1.4B without regularization | standard recipe · DCLM validation loss | loss | Hurt |
| ensemble K models instead of scaling parameters | regularized recipe · DCLM validation loss | loss asymptote: 3.43 -> 3.34 | Improved |
| tune member hyperparameters for K to infinity (more epochs, less weight decay) (ablation) | ensemble of 300M models · DCLM validation loss | loss asymptote: 3.34 -> 3.27 | Improved |
| compose ensembling and parameter scaling (joint scaling) | ensemble recipe · DCLM validation loss | loss asymptote: 3.43 -> 3.17 | Improved |
| fix weight decay at 0.8 tuned only at 150M and transfer to larger models (ablation) | regularized scaling · DCLM validation loss | scaling behavior | Hurt |
| fix epoch count 16 tuned only at 150M (ablation) | regularized scaling · DCLM validation loss | scaling behavior | Hurt |
| distill 8-ensemble into 300M student | regularized 300M model · DCLM validation loss | loss: 3.57 -> 3.36 | Improved |
| self-distill 300M teacher into same-size student | 300M model · DCLM validation loss | loss | Improved |
| reduce batch size in continued pre-training (ablation) | CPT Llama 3B · MegaMath CPT average accuracy | accuracy: 30.59 -> 34.48 | Improved |
| add epoching (K=1) to lower batch size (ablation) | CPT Llama 3B · MegaMath CPT average accuracy | accuracy: 34.48 -> 35.82 | Improved |
Closest papers by meaning
- Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws (NeurIPS 2026 · similarity 0.92)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.91)
- Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training (NeurIPS 2025 · similarity 0.91)
- Bridging Compute- and Data-Optimal Pretraining (NeurIPS 2026 · similarity 0.90)
- Scaling Laws for Precision (ICLR 2025 · similarity 0.90)
- Resolving Discrepancies in Compute-Optimal Scaling of Language Models (NeurIPS 2024 · similarity 0.89)
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.
