Research map / Large language models / Language, safety and interpretability
Emergent and Predictable Memorization in Large Language Models
Stella Biderman, USVSN Sai Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Gregory Anthony, Shivanshu Purohit, Edward Raff
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 |
|---|---|---|---|
| use fully-trained smaller Pythia models to predict 12B memorization | Pythia-1.4B predictor · 12B memorized sequences | recall: 0.554 | Hurt |
| use fully-trained 70M model as predictor of 12B memorization | Pythia-70M predictor · 12B memorized sequences | precision vs recall: 95.6% precision, 19.7% recall | Mixed |
| increase checkpoint training progress from 23M to 126M sequences | Pythia-6.9B checkpoint predictor · final model memorization | recall: 0.513 -> 0.918 | Improved |
| increase checkpoint training progress from 23M to 126M sequences | Pythia-12B checkpoint predictor · final model memorization | recall: 0.500 -> 0.916 | Improved |
| increase checkpoint training progress from 23M to 126M sequences | Pythia-6.9B checkpoint predictor · final model memorization | precision: 0.919 -> 0.945 | No clear effect |
| increase checkpoint progress from 23M to 126M sequences (ablation) | Pythia-12B-deduped checkpoint predictor · final model memorization | recall: 0.523 -> 0.920 | Improved |
| increase model size at fixed low compute budget | equi-compute Pythia predictors · 12B memorization recall | recall | Improved |
| use smallest model at >=10% of final compute budget | equi-compute Pythia predictors · 12B memorization recall | recall | Improved |
| double the number of matching tokens required for memorized (ablation) | precision/recall predictors · 12B memorization | precision and recall | No clear effect |
| train on deduplicated Pile (Pythia-deduped) (ablation) | Pythia-12B · fraction memorized | memorization rate: slightly lower than 1.62% | Improved |
| replace Pythia with Pythia-deduped in predictor analyses (ablation) | checkpoint and size predictors · 12B-deduped | precision and recall | No clear effect |
Closest papers by meaning
- Quantifying and Analyzing Entity-Level Memorization in Large Language Models (AAAI 2024 · similarity 0.91)
- Mitigating Memorization in Language Models (ICLR 2025 · similarity 0.91)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.90)
- Rethinking LLM Memorization through the Lens of Adversarial Compression (NeurIPS 2024 · similarity 0.90)
- Characterizing Memorization in Diffusion Language Models: Generalized Extraction and Sampling Effects (NeurIPS 2026 · similarity 0.90)
- Learning the Signature of Memorization in Autoregressive Language Models (NeurIPS 2026 · similarity 0.90)
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.
