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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| train GPT-style transformers on uniformly random bitstrings to saturation | GPT-2 architecture models, 100K-20M params · synthetic bitstrings | capacity (bits per parameter): 3.51 bpp (bf16) | Improved |
| increase training precision from bfloat16 to float32 (ablation) | GPT-2 architecture models · synthetic bitstrings | capacity (bits per parameter): 3.51 -> 3.83 | No clear effect |
| increase dataset size beyond model capacity | language models on text · FineWeb 64-token sequences | unintended memorization and generalization | Mixed |
| dataset size exceeding model capacity in bits | language models on text · FineWeb | test loss (double descent) | Improved |
| increase training set size | deduplicated text training · FineWeb extraction rate | training extraction rate: converges to test extraction rate | Hurt |
| increase dataset size relative to capacity | loss-based membership inference · text and synthetic | membership inference F1: toward 0.5 | Hurt |
| fit sigmoid scaling law with capacity/dataset-size ratio | membership inference F1 prediction · GPT2-XL and GPT2-Medium | F1 prediction error: predictions within 1.5 points | Improved |
| membership inference versus extraction on the same models | text-trained models · FineWeb | attack success: MI score 0.97 with extraction rate 0 | Improved |
| increase vocabulary size V (adds parameters) (ablation) | capacity prediction · synthetic data | memorized bits: 1.49e6 -> 3.13e6 bits (V 128 to 4096) | Improved |
Closest papers by meaning
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.90)
- Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws (ICLR 2025 · similarity 0.89)
- Mitigating Memorization in Language Models (ICLR 2025 · similarity 0.89)
- Quantifying Generalization Complexity for Large Language Models (ICLR 2025 · similarity 0.88)
- Learning the Signature of Memorization in Autoregressive Language Models (NeurIPS 2026 · similarity 0.88)
- Quantifying and Analyzing Entity-Level Memorization in Large Language Models (AAAI 2024 · similarity 0.88)
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.
