New study finds language models memorize about 3.6 bits per parameter
Researchers unveil a method to measure LLM memorization capacity, finding GPT-style models store roughly 3.6 bits of information per parameter, with implications for grokking and scaling.
A team of researchers has introduced a new method to quantify exactly how much a language model knows about a given data point, cleanly separating this into two components: unintended memorization (dataset-specific information retained by the model) and generalization (knowledge of the true underlying data-generation process). By isolating and removing the generalization component, they can compute total memorization, which serves as a direct estimate of a model's raw capacity.
Training hundreds of transformer models ranging from 500K to 1.5B parameters, the researchers found that GPT-style architectures store roughly 3.6 bits of information per parameter. As training datasets grow, models memorize until this capacity is saturated, after which 'grokking' kicks in and models shift toward generalization, causing unintended memorization to decline.
The work also yields new scaling laws linking model capacity and dataset size to membership inference success rates. For engineers, this offers a concrete, measurable framework for reasoning about model sizing, data privacy risks, and overfitting behavior in large language models.