Paper: How Much Do LLMs Remember? ~3.6 Bits Per Param

pmttyji · reddit · 2026-07-07

A new paper proposes a method to estimate a model's memory of individual data points to measure its capacity. The authors formally decompose memory into "incidental memory" (information specific to a dataset) and "generalization" (information about the true data generation process). By completely eliminating generalization, the total memory yields the capacity estimate: GPT-style models store approximately 3.6 bits per parameter.

When training on gradually increasing datasets, models first exhaust their capacity for memorization before entering the grokking phase, where incidental memory decreases and generalization increases. The researchers trained hundreds of transformers ranging from 500,000 to 1.5 billion parameters, deriving a series of scaling laws related to capacity, data scale, and membership inference.

Related event: ICML Paper: LLMs Store 3.6 Bits of Memory Per Parameter(2 posts)→

Original post →

More from Research

Research channel →