Pith. sign in

REVIEW 23 cited by

How much do language models memorize?

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2505.24832 v3 pith:WXZVI4V2 submitted 2025-05-30 cs.CL

How much do language models memorize?

classification cs.CL
keywords modelsmemorizationmodelcapacitylanguagegeneralizationcontainsestimate
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

We propose a new method for estimating how much a model knows about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: unintended memorization, the information a model contains about a specific dataset, and generalization, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point "grokking" begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 23 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Can a Language Model Learn Facts Continually in Its Weights?

    cs.CL 2026-07 conditional novelty 7.0

    Facts written into LLM weights are stored but question-keyed; broad data creates usable knowledge while later writes destroy access rather than storage, leaving context more reliable than weights for composition and r...

  2. Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning

    cs.AI 2026-07 conditional novelty 7.0

    Memorized facts in fine-tuned LLMs often sit off the mid-layer reasoning path; relocating those representations recovers most multi-hop generalization failures.

  3. LLM-Based Examination of Eligibility Criteria from Securities Prospectuses at the German Central Bank

    cs.CL 2026-06 unverdicted novelty 7.0

    LLMs are applied in a generative pipeline for extracting, normalizing, and interpreting eligibility criteria from securities prospectuses, achieving up to 91% precision in document-level decisions with a conservative bias.

  4. Detecting Functional Memorization in Code Language Models

    cs.LG 2026-06 unverdicted novelty 7.0

    Authors demonstrate functional memorization in code LLMs via counterfactual midtraining comparison on functional equivalence metrics beyond textual overlap.

  5. Erase Persona, Forget Lore: Benchmarking Multimodal Copyright Unlearning in Large Vision Language Models

    cs.CV 2026-05 unverdicted novelty 7.0

    CoVUBench is the first benchmark framework for evaluating multimodal copyright unlearning in LVLMs via synthetic data, systematic variations, and a dual protocol for forgetting efficacy and utility preservation.

  6. Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity

    cs.LG 2026-04 unverdicted novelty 7.0

    Incompressible Knowledge Probes enable log-linear estimation of LLM parameter counts from factual accuracy on obscure questions, showing continued scaling of knowledge capacity across open and closed models.

  7. Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity

    cs.LG 2026-04 conditional novelty 7.0

    Incompressible Knowledge Probes map black-box factual accuracy to order-of-magnitude parameter count via a log-linear fit on 93 open models (R²=0.910), and reject the Densing Law for factual capacity.

  8. Can a Language Model Learn Facts Continually in Its Weights?

    cs.CL 2026-07 unverdicted novelty 6.0

    Weight writes create question-keyed knowledge that later sequential writes redirect, so context—not weights—is the reliable channel for facts that must survive or be composed.

  9. Fast Unlearning at Scale via Margin Self-Correction

    cs.LG 2026-06 unverdicted novelty 6.0

    MASC achieves competitive forget-retain trade-offs in language model unlearning at lower computational cost via margin self-correction and an online stopping criterion on TOFU, MUSE News, and MUSE Books.

  10. Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning

    cs.LG 2026-05 unverdicted novelty 6.0

    In a structured-output NW matrix task, Transformers generalize fastest at intermediate dataset sizes while larger sets can accelerate memorization in partial-competence regimes.

  11. Before Forgetting, Learn to Remember: Revisiting Foundational Learning Failures in LVLM Unlearning Benchmarks

    cs.CV 2026-05 unverdicted novelty 6.0

    LVLM unlearning benchmarks fail due to initial memorization failures on fictitious data; ReMem benchmark with multi-hop and multi-image scaling plus Exposure metric enables reliable learning and unlearning diagnosis.

  12. Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts

    cs.CL 2026-04 conditional novelty 6.0

    Loss-based pruning of training data to limit facts and flatten their frequency distribution enables a 110M-parameter GPT-2 model to memorize 1.3 times more entity facts than standard training, matching a 1.3B-paramete...

  13. Deep sequence models tend to memorize geometrically; it is unclear why

    cs.LG 2025-10 unverdicted novelty 6.0

    Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.

  14. AtlasKV: Augmenting LLMs with Billion-Scale Knowledge Graphs in 20GB VRAM

    cs.CL 2025-10 unverdicted novelty 6.0

    AtlasKV integrates billion-scale KGs into LLMs parametrically with sub-linear complexity and low memory by converting triples into key-value representations handled by the model's attention.

  15. Access Paths for Efficient Ordering with Large Language Models

    cs.DB 2025-08 unverdicted novelty 6.0

    Introduces the LLM ORDER BY semantic operator with algorithmic improvements, a semantic-aware external merge sort, and a budget-aware optimizer that selects near-optimal access paths for LLM-based ordering.

  16. Enabling Global, Human-Centered Explanations for LLMs:From Tokens to Interpretable Code and Test Generation

    cs.SE 2025-03 unverdicted novelty 6.0

    CodeQ aggregates token rationales into code categories to enable global interpretability of LLMs, claiming over 50% entropy reduction and revealing model preference for syntactic cues plus human misalignment in a 37-p...

  17. Train Smarter, Not Longer: Memorization-Guided Data Reuse for Efficient LLM Training

    cs.LG 2026-07 conditional novelty 5.0

    Properly spaced multi-epoch reuse of high-quality data, guided by a measured memorization window, continues to improve LLM performance far beyond the common four-epoch heuristic.

  18. The Future of Facts: Tracing the Factual Generation-Verification Gap

    cs.CL 2026-05 unverdicted novelty 5.0

    Empirical tracing across model families shows verification precedes and outlasts generation for facts, with updates producing simultaneous verification of old and new answers.

  19. Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay

    cs.LG 2026-05 unverdicted novelty 5.0

    Self-generated replay from language models nearly eliminates catastrophic forgetting during finetuning except when models are pretrained close to saturation.

  20. Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds

    cs.LG 2026-05 unverdicted novelty 5.0

    Grokking emerges near the model size where memorization timescale T_mem(P) intersects generalization timescale T_gen(P) on modular arithmetic.

  21. LaCy: What Small Language Models Can and Should Learn is Not Just a Question of Loss

    cs.CL 2026-02 unverdicted novelty 5.0

    LaCy augments next-token loss with spaCy parser signals to train SLMs on learn-vs-delegate decisions, producing higher FactScores in cascades than loss-only or LLM-judge baselines.

  22. PromptMN: Pseudo Prompting Language

    cs.CL 2026-06 unverdicted novelty 4.0

    PromptMN is a pseudo-prompting DSL that adds compact typed directives to natural language to improve clarity, reusability, and reverse engineering of AI instructions.

  23. Causal Connections: Leveraging Multilingual Fine-Tuning for Financial QA@FinCausal 2026

    cs.CL 2026-06 unverdicted novelty 2.0

    Fine-tuned multilingual LLMs achieve top shared-task scores on financial causality extraction in English and Spanish.