Research map / Architectures and efficiency / RNNs and normalization
Predicting Large Model Test Losses with a Noisy Quadratic System
Chuning Li, Chris J. Maddison
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 |
|---|---|---|---|
| replace Chinchilla loss model with NQS | loss predictor · Pythia + OWT2 holdout IsoFLOPs, x1024 gap | Huber loss x1e-5: 9.0 -> 2.5 | Improved |
| replace Chinchilla loss model with NQS | loss predictor · Pythia + OWT2 holdout B-K plane, x1024 gap | Huber loss x1e-5: 9.8 -> 5.6 | Improved |
| replace Chinchilla loss model with NQS | loss predictor · Llama + LM1B holdout B-K plane, x6 gap | Huber loss x1e-5: 8.7 -> 8.2 | Improved |
| replace Chinchilla loss model with NQS on training data | loss predictor · Pythia + OWT2 train IsoFLOPs | Huber loss x1e-5: 0.8 -> 2.8 | Hurt |
| add LayerNorm effective learning rate adjustment (gamma ~ 1/||w||^2) (ablation) | NQS · small batch sizes, large models | loss fit | Improved |
| add LayerNorm adjustment (ablation) | NQS · large batch sizes | loss fit | No clear effect |
| increase extrapolation compute gap to x100 and beyond | Chinchilla · Pythia + OWT2 | extrapolation error | Hurt |
| increase extrapolation gap up to 1/4000 compute | NQS · Pythia + OWT2 | extrapolation error | No clear effect |
| mix critical and non-critical batch sizes in Chinchilla training data (ablation) | Chinchilla · Pythia + OWT2 | prediction accuracy | Hurt |
| use NQS to choose N, B, K under compound constraints | NQS · compute plus time/memory/data constraints | optimality of chosen config | Improved |
Closest papers by meaning
- A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules (ICLR 2025 · similarity 0.85)
- Power Lines: Scaling laws for weight decay and batch size in LLM pre-training (NeurIPS 2025 · similarity 0.84)
- Bayes-optimal learning of an extensive-width neural network from quadratically many samples (NeurIPS 2024 · similarity 0.83)
- Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize (AAAI 2024 · similarity 0.83)
- Critical Batch Size Revisited: A Simple Empirical Approach to Large-Batch Language Model Training (NeurIPS 2025 · similarity 0.83)
- Debiasing Mini-Batch Quadratics for Applications in Deep Learning (ICLR 2025 · similarity 0.82)
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.
