AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

How much can language models memorize?

John Xavier Morris, Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G. Edward Suh, Alexander M Rush, Kamalika Chaudhuri, Saeed Mahloujifar

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train GPT-style transformers on uniformly random bitstrings to saturationGPT-2 architecture models, 100K-20M params · synthetic bitstringscapacity (bits per parameter): 3.51 bpp (bf16)Improved
increase training precision from bfloat16 to float32 (ablation)GPT-2 architecture models · synthetic bitstringscapacity (bits per parameter): 3.51 -> 3.83No clear effect
increase dataset size beyond model capacitylanguage models on text · FineWeb 64-token sequencesunintended memorization and generalizationMixed
dataset size exceeding model capacity in bitslanguage models on text · FineWebtest loss (double descent)Improved
increase training set sizededuplicated text training · FineWeb extraction ratetraining extraction rate: converges to test extraction rateHurt
increase dataset size relative to capacityloss-based membership inference · text and syntheticmembership inference F1: toward 0.5Hurt
fit sigmoid scaling law with capacity/dataset-size ratiomembership inference F1 prediction · GPT2-XL and GPT2-MediumF1 prediction error: predictions within 1.5 pointsImproved
membership inference versus extraction on the same modelstext-trained models · FineWebattack success: MI score 0.97 with extraction rate 0Improved
increase vocabulary size V (adds parameters) (ablation)capacity prediction · synthetic datamemorized bits: 1.49e6 -> 3.13e6 bits (V 128 to 4096)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy