Breaking

ICML Paper Pins Down an LLM's Memory: About 3.6 Bits Per Parameter

July 17, 2026 at 11:45 EDT

Research & Papers
Foundation Models
Security

Researchers from Meta, Google DeepMind, Cornell and NVIDIA have put a concrete number on how much a language model can memorize: roughly 3.6 bits per parameter, implying a 7-billion-parameter model can hold on the order of 3GB of training data verbatim. The finding comes from the paper "How much do language models memorize?" (arXiv:2505.24832), which was accepted to ICML 2026 and named an Outstanding Paper Honorable Mention (ICML listing).

ICML 2026 · Meta · Google DeepMind · Cornell · NVIDIA

How Much Can a Language Model Memorize? About 3.6 Bits Per Parameter

Researchers put a hard number on raw storage capacity — implying a 7-billion-parameter model can hold roughly 3GB of training data verbatim, and offering a cleaner way to separate memorization from real learning.

3.6 bits

Storage capacity per parameter (bfloat16)

~3GB

Verbatim data a 7B-parameter model can hold

500K–1.5B

Parameter range across several hundred models tested

Precision barely changes capacity

Moving from bfloat16 to fp32 lifts storage only marginally — so higher precision offers little relief against data leakage.

3.60

bfloat16

3.83

fp32

bits stored per parameter — nearly doubling numerical precision adds just ~0.23 bits

Fixed capacity drives a clean transition

1 · MEMORIZE

Small datasets fit inside capacity, so the model stores examples verbatim.

→

2 · SATURATE

Data outgrows capacity; unintended memorization drops sharply.

→

3 · GENERALIZE

"Grokking" begins — the model learns the distribution instead of storing it.

What it clarifies

A firmer empirical basis for privacy: for large models on large datasets, inferring whether any single example was in training becomes hard — each point contributes little to total capacity.

The caveats

Rare or unique strings remain prone to verbatim retention, and the tiny fp32 gain means precision tuning offers little protection against data leakage.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.