LLM Memorization: The 3.6 Bits Per Parameter Threshold
gabriberton · x · 2026-07-03
The paper "How much do language models memorize?" reveals that LLMs continuously memorize training data until saturation occurs. This saturation point is reached at 3.6 bits per parameter; beyond this, the model begins to compress → generalize → exhibit double descent → and experience grokking. This establishes a quantitative threshold for the transition from memorization to generalization.
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27