A new training recipe claims 5.17x better data efficiency than standard scaling
inductionheads · x · 2026-07-21
A research summary argues that as compute keeps growing and data becomes scarce, the standard recipe for training large language models may be leaving efficiency on the table.
Using a 200M-token budget, the authors report two key findings:
- raising weight decay to around 30× the usual 0.1 helps more than the standard setup
- training an ensemble of models with the same total parameter count can beat a single larger model
Under Chinchilla-style scaling assumptions, they compare several recipes and estimate that aggressively scaling model size and ensemble count together can match baseline performance with 5.17× less data. The chart in the post contrasts standard training with regularized and ensembling-based recipes, showing the joint scaling approach as the best data-efficiency strategy.
More from Research
- Cognition's SWE-2 uses a KKT duality argument in RL to shift the effort Pareto curve — YouJiacheng · 2026-09-11
- VidMap uses RoMa coarse matching on all frames, fine-scale only for keyframes — ducha_aiki · 2026-09-11
- Bug Hunt Bench author: leaderboard noise is about 2-3 points — PawelHuryn · 2026-09-11
- PNAS paper shows a tiny billiard-ball system is a universal computer — undecidability lives in two dimensions — eigensteve · 2026-09-11
- New paper: Absolute pose estimation from affine cues and gravity direction — ducha_aiki · 2026-09-11
- LoMa Paper Ships REALLY HardPairs Dataset, Accepted at ECCV 2026 — ducha_aiki · 2026-09-11