Skip to content
Synced · WeChat

The Ma Jiaqi Failure Exposed an LLM Issue He Spotted in the Shower a Year Earlier

「马嘉祺」让大模型翻车,而他一年前洗澡时就发现了问题

FaceMind links low-frequency token degradation to two papers: SLoW appeared at EMNLP 2025, Adam's Law was accepted as an ACL 2026 Oral, and high-frequency rewriting raised DeepSeek-V3 math accuracy from 63.55% to 71.54%.

Why it matters: The FaceMind team links large models failing on low-frequency words to the SLoW and Adam's Law papers, laying out a mechanistic account of low-frequency token degradation.

Read the original ↗Export Markdown