Paper tests if universal MLIP scaling laws hold — sampling policy and test set shift flip rankings
CatAstro_Piyush · x · 2026-11-16
The arXiv paper "Are Scaling Laws of Universal MLIPs Universal?" systematically tests whether data scaling laws hold universally for machine-learned interatomic potentials under fixed model configurations.
- Setup: M3GNet, MACE, and PET trained on subsets (10K–640K structures) of a combined MatPES-MP-ALOE pool plus the full 1.16M-structure set
- Sampling policies compared: uniform random, three chemistry/structure-informed stratification schemes, iterative DIRECT, and furthest-point sampling using embeddings from three pretrained MLIPs
- Findings: error follows power-law scaling with more data, but the fitted exponent, apparent error floor, and sample efficiency depend on architecture, target quantity, sampling policy, and test distribution
- Policy rankings reverse across test distributions: chemical stratification helps cross-dataset performance while hurting in-distribution; embedding-diversity sampling often underperforms random sampling
- Takeaway: scaling gains hinge on test-train chemical similarity — understanding the source data matters more than blindly adding more
More from Research
- Open-source TotalSegmentator spots a 42mm dilated aorta on a routine chest CT — MaziyarPanahi · 2026-11-16
- Tivadar Danka Publishes 4,000-Word Visual Guide to Graph Theory for Machine Learning — TivadarDanka · 2026-11-16
- METR time-horizon chart updated: doubling every 1.7 months since Opus 4.5, 'Model 4' ~38h — ben_j_todd · 2026-11-16
- GTAgentClinic: one-line install, plug in Claude Code or Codex for a 3D hospital shift — Michael_D_Moor · 2026-11-16
- 28-model tracking table: half publish no training cutoff, Gemini 4 Carbon 8 months stale at launch — Joozio · 2026-11-16
- FutureSim Q3: LLMs now genuinely improve from experience, Opus 5.5 tops at 50% — rickasaurus · 2026-11-16