TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

research

Doubling Data Brought Grokking Four Times Sooner

Across 384 modular-arithmetic runs, data complexity dominated width in the fitted transition law.

Published Updated Story ID: mp-2026-09-12-015
Read the complete editionStory JSON

Summary

Across 384 modular-arithmetic runs, data complexity dominated width in the fitted transition law.

The reported scaling fit assigns data a minus-2.04 exponent, versus minus-0.27 for hidden width: doubling data moved generalization onset about fourfold, while doubling width shifted it roughly 1.2-fold. A sharp weight-decay boundary separated grokking from non-grokking in this small-model setting.

Why it matters

Across 384 modular-arithmetic runs, data complexity dominated width in the fitted transition law.

Limits and context

No additional limitation was separately recorded.

Key claims

  1. Across 384 modular-arithmetic runs, data complexity dominated width in the fitted transition law.

    Evidence: source-2026-09-12-017

Sources

  1. arXiv preprint 2609.10657arXiv · primary research

Corrections

No corrections have been recorded for this story.