Paper: data schedules and learning rate schedules should be designed together
alifcoder · x · 2026-10-27
A new paper shows the optimal placement of limited math data in LLM pre-training flips with the learning rate schedule. At fixed compute and mid-training token budgets, earlier mixing yields lower MT loss when LR decays throughout training; with constant LR, placing all mid-training data at the end works better. Varying when decay begins, the best tested strategy keeps LR high after mid-training data enters and decays near the end — data and LR schedules should be co-designed.
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27