Unverified50% confidenceFactExact time
Grokking现象最早由Power等人在2022年的论文中报告,小型Transformer在模运算任务上训练准确率达到100%后,验证准确率在成千上万额外训练步骤后突然跃升至完美
1
Sources
50%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
Cite This Claim
Stable URI
https://kongchang.com/claim/694288API
curl https://kongchang.com/api/v1/knowledge/claims/694288MCP
get_claim(id=694288)