Memorization in Language Models

Large language models (LLMs) can expose memorized training sequences under prefix-based extraction: given a prefix from a training example, the model may assign high probability to the original continuation. In deployed systems, however, prefixes are rarely evaluated in isolation. They often appear together with instructions, retrieved documents, or other task-specific context, as in retrieval-augmented generation (RAG). This motivates examining whether contextual conditioning mitigates memorization or merely changes the set of memorized samples that become extractable. We investigate this issue through paired item-level measurements of probabilistic suffix extraction. For each prefix-suffix pair, we score the target suffix under an empty prompt and under retrieved contexts of varying relevance, across three open-weight instruction-tuned models. We find that context does not simply erase memorization. Instead, extractable memorization consists of a context-robust core and a context-sensitive boundary. Many samples that are extractable without context remain extractable under the retrieved context, especially as the prefix length increases. At the same time, context mainly affects marginal samples near the extraction threshold: it suppresses some exposures, but also enables new ones that are missed by prefix-only evaluation. These findings qualify the view that RAG reduces memorization risk. Context can lower aggregate extraction by suppressing boundary cases, yet robustly extractable samples persist, and context-enabled extractability remains security-relevant.

Knowledge that a language model appears to forget during finetuning often remains stored and can be recovered, a phenomenon called spurious forgetting. Finetuning on new facts can even produce forgetting that undoes itself: recall of the old facts collapses, recovers as training continues on new facts alone, and only then erodes for good. We seek to understand when such forgetting is not catastrophic. A minimal associative memory reproduces these dynamics with three ingredients: keys with shared structure, concentrated new values, and normalization in the network. Finetuning moves all old representations along a common direction, hiding the old facts while preserving their relative geometry; normalization withdraws this shift once the new facts are learned, whereas fact-specific changes accumulate and cause the erosion. Moreover, subtracting the common shift eliminates the collapse in a Transformer trained on synthetic data, and removing a single direction from each weight update restores old facts in a pretrained language model. Forgetting thus combines a shared, reversible loss of access with a slow erosion of individual facts, and only the second is catastrophic. Which one dominates depends on whether the new data move old memories together or apart.

While unlearning seeks to negate undesired capabilities acquired through learning, little research has examined how the way models learn shapes their subsequent unlearning. In this paper, we investigate this connection from the perspectives of memorization and generalization, the two most representative yet competing strategies that models employ during training. We first classify memorization- and generalization-heavy models using grokking in modular addition and compare their responses to unlearning, showing that the latter suffer greater retain damage, i.e., a larger performance drop on the retain set. Furthermore, we conduct a finer-grained analysis by introducing bucketed modular addition, in which the respective contributions of the two strategies can be explicitly controlled across the memorization-generalization spectrum. In this setup, we reaffirm that the same trend persists and is nearly monotonic. We further demonstrate that this relationship also holds in LLM unlearning across verbatim and factual recall settings. Finally, we provide two practical insights for developing better unlearning methods, highlighting the importance of accounting for learning dynamics in unlearning.

Large language models (LLMs) reliably perform entity copying, in which a model copies tokens referring to an entity, termed entity tokens, from the prompt into its output to answer a question. Although entity copying is straightforward for most LLMs, existing research does not provide a systematic account of which layers specialize in this fundamental task or how other tokens in the same sequence, termed context tokens, influence the model's ability to copy the entity tokens. To address these questions, we conduct experiments on Qwen3-8B using two novel methods: genie-in-a-bottle, which controls exactly which layers can participate in an entity-copying task, and attention lobotomy, which cuts off specific tokens' attention to entity tokens without affecting the remaining attention distribution. We find that two distinct groups of layers in the second half of the model are both necessary and sufficient for entity copying. Moreover, in addition to the decoding position's attention to entity tokens, context tokens' attention to entity tokens also proves necessary for copying the exact tokens, even though context tokens do not store entity information themselves unless they satisfy particular semantic properties. Our findings establish the critical role of late layers in entity copying under the guidance of context tokens, calling for future work on how models propagate and consume entity information.

Large Language Models (LLMs) perform well on reasoning benchmarks, but it remains unclear whether this reflects genuine contextual reasoning or reliance on facts memorized in their parameters. We investigate this by distinguishing two possibilities: a broad \textit{memorization bias}, where familiar content improves reasoning performance, and the \textit{Strong Parametric Shortcut Hypothesis}, where models skip reasoning entirely and recall stored answers. To test these effects, we introduce \textbf{MemoReason}, a human-curated benchmark that pairs factual reasoning tasks with structurally identical \fictitiousterm{} versions where real entities like people, companies, or dates are systematically replaced by \fictitiousterm{} ones of the same type. This \scorerevision{preserves task structure and specified reasoning operations} while varying the familiarity of the context, allowing controlled measurement of how the parametric memory affects reasoning. \revision{Our evaluation of recent LLMs reveals consistent and statistically significant performance drops of up to 15.7% in the fictitious setting, demonstrating a clear memorization bias.} However, a targeted analysis of \revision{questions failed in the fictitious setting} shows that models rarely respond with the corresponding factual answer, indicating that direct parametric shortcuts are not the dominant failure mode. These findings suggest that parametric memory influences reasoning through mechanisms more complex than simple factual recall. \textbf{MemoReason} provides a controlled framework for studying these mechanisms and for extending paired factual-fictitious{} evaluation to broader reasoning settings.

Memorization has been proposed as a mechanism to explain how language models fit the tail of their training distributions, but its training dynamics are not understood well. In this work, we take a fine-grained look at memorization by decomposing the loss trajectory of memorized sequences over training and model parameters. Across the Pythia family, we study memorization of duplicated training sequences (recitation) and rare ones (recollection). We find that memorization in both cases is characterized by sequence-level gradient alignment, though recitation suffers from misalignment with other training influences which causes forgetting, explaining the necessity for higher duplication of these examples. We further show that the lower model layers are the most involved in memorization and forgetting. Predicting memorization, our decomposition improves over a cross-entropy baseline, especially in larger models and early in training. Intervening on a small set of highly influential parameters we are able to ablate memorization in the final model. Together, these findings advance our understanding of how memorization develops during training and offer insights for predicting and intervening on it.

We reproduce and stress-test the work of Yu et al. (2023), who characterize how language models (LMs) arbitrate between memorized knowledge and contradictory in-context statements. We replicate their world-capitals experiments on 31 models spanning Pythia, GPT-2, Qwen3, and Ministral families, including base and post-trained variants, and extend evaluations to five additional knowledge relation types from the ParaConflict dataset. We empirically confirm most of their original findings: larger models and higher-frequency entities tend to favor memorized answers, with substantial family-level variance. However, several conclusions do not generalize cleanly: entity-frequency effects disappear on Qwen3-14B and 32B; post-training shifts the memory-context trade-off inconsistently across families; question phrasing alone can change a model's reliance on memorized knowledge by up to 80 percentage points; and semantically unrelated prose can mimic coherent supporting context. Our results clarify where Yu et al.'s claims hold and to what extent they generalize to other prompts.

State Space Models (SSMs) have emerged as a compelling alternative to Transformers, enabling sequence modeling with constant memory and linear compute. Although SSMs exhibit reasonable performance and favorable computational characteristics, they continue to lag behind Transformers on tasks that require in-context learning and precise retrieval, slowing their adoption for large-scale language modeling. In this work, we demonstrate that both the success and failure of SSMs in these domains can be explained by studying the role of the gating mechanism, a prevalent component in modern recurrent networks. Specifically, we show through theory and experiments that this gating mechanism causes SSMs to first learn an in-weights "memorization" solution, while delaying, or even preventing, convergence to a correct in-context learning solution. Importantly, this happens even in cases where there are no fundamental limitations due to the architecture or its memory capacity. On the other hand, we find that gating is often beneficial for improving generalization to long sequence lengths. Our results illuminate the crucial role of the gating mechanism in shaping both the training dynamics and generalization of SSMs, and provide a basis for understanding and improving linear-time models.

Factual hallucination in closed-book question answering is often treated as a coverage problem: a model fails because the relevant fact is absent from its internal memory. This view misses a second source of error. Even when a fact has been observed, finite memory may force it to be stored only approximately. We study this effect through a simple coverage--compression model of factual recall. We consider an unstructured question-answering task with $N$ possible queries and $K$ possible answers. A learner observes $M$ training facts, compresses them into at most $B$ bits, and answers uniformly drawn test queries without retrieval. For a uniformly random ground-truth mapping, we prove $\mathcal{E} \geq \frac{M}{N}\delta^\star\!\left(\frac{B}{M}\right) + \left(1-\frac{M}{N}\right)\left(1-\frac{1}{K}\right)$, where $\delta^\star(r)$ is the inverse rate-distortion function of a uniform $K$-ary source under zero-one loss. The two terms separate compression distortion on observed facts from missing coverage on unobserved facts. The bound gives a compact way to reason about selective memory, forced compression, structure, retrieval, abstention, and long-context organization. We study the predicted signatures with theory-implied simulations and controlled fact-injection probes in modern language models that vary fact load and effective trainable memory. The result is not a complete theory of hallucination, but an information-theoretic account of a separable failure mode: lossy recall of observed facts under finite memory.

Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize surface-level text spans. To address this, we present an information-weighted cross-entropy loss that rescales token-level contributions using TF-IDF statistics, emphasizing semantically informative tokens while down-weighting ubiquitous ones. Experiments on five decoder-only LLMs ranging from 1.1B to 13B parameters show consistent reductions in memorized substring length while preserving perplexity and downstream task performance. Under LoRA fine-tuning, TF-IDF reduces average substring memorization length by 14% across all five models. Under full-weight fine-tuning on TinyLLaMA 1.1B, the reduction reaches 58%. Our approach is architecture-agnostic and can be incorporated into existing training pipelines with less than 3% computational overhead, offering a lightweight and principled way to mitigate memorization without disrupting standard training dynamics.

When a language model finds a sentence unusually cheap to predict, it is tempting to conclude that the sentence was in its training data. Almost every published test of that inference has had to guess which sentences were in the training data, the members, and which were not. This paper removes the guessing. Two model families, OLMo-2 and Pythia, publish their pretraining corpora, and a public index over those corpora returns the exact number of times any sentence appeared in each. Those counts make three questions answerable directly. The answers form a pincer, closing from two sides. At the duplication levels ordinary text actually has, five models from 1B to 13B parameters carry at most a faint trace of their own exposure. We measure that trace with a design that reads the same sentence through two models, which cancels fluency and quality by construction, and it comes to a rank correlation near -0.08, where -1 would be a perfect relation and 0 none. Where the trace does become strong, above roughly a thousand copies, the two corpora agree on which sentences those are, because they are the famous ones, so exposure can no longer be told apart from fame. Two further measurements show how apparent membership signal gets manufactured. A common way to build a non-member is to change one word of a member. The model does prefer the original, but the gap is the same, within noise, whether the original appeared once or a hundred times, so what the model is rewarding is the author's word choice, not memory. Above a thousand copies the gap grows with model size on the twelve sentences we can test there, at the same boundary where the pincer closes. And swapping the controls for sentences that differ from the members in register moves a detector from 0.83 to 0.94 AUC, on a scale where 0.5 is a coin flip and 1.0 is perfect separation. We release the sentence banks, counts, and code.

Large Language Models (LLMs) can memorize and reproduce sensitive, copyrighted, or otherwise undesirable training content, creating privacy, safety, and regulatory concerns. Machine unlearning offers a practical alternative to full retraining, but many existing methods apply broad or fixed parameter updates that can degrade utility and remain brittle under deployment changes such as post-training quantization, where forgotten knowledge may partially re-emerge. We propose Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer-level unlearning framework that selects transformer layers using a forget-to-retain significance score. This score identifies layers with high influence on the forget set and low sensitivity to the retain set, allowing FOM-UL to concentrate updates where they are most effective while leaving most of the model unchanged. This targeted update strategy improves the forgetting-utility trade-off and provides an empirical path toward quantization-resilient unlearning by reducing the chance that small, diffuse updates are erased by low-bit rounding. Across TOFU, KnowUnDo, and MUSE-style evaluations, FOM-UL reduces residual memorization compared with strong GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines while preserving retain-set utility close to the vanilla model. Under 8-bit and 4-bit post-training quantization, FOM-UL maintains stronger memorization suppression and utility preservation than competing methods, and adversarial prompt evaluations show lower recovery of forgotten content. Overall, FOM-UL provides an efficient unlearning strategy that improves targeted forgetting, utility preservation, and deployment robustness without claiming formal guarantees of erasure.

We approach a cognitive simulation perspective on episodic and semantic memory in multiple-choice question answering by incorporating text from individual text corpora (ITC) into retrieval-augmented generation and DoRA fine-tuning. We web-crawl the search histories of 515 participants who answered 36 multiple-choice knowledge items and analyze a stratified subsample of 150 participants. For each participant, one DoRA adapter consolidates their ITC into a small language model (SLM) whose baseline correctness falls below the participants' lowest quartile. The adapter measurably writes the ITC into the weights: it fits its own participant's held-out text better than other participants' texts (dz =1.27), an individuality effect that increases with ITC size in rank order. On the generalized knowledge test, however, the adapter adds knowledge rather than alignment with the individual: log-loss match improves, whereas match accuracy under a bias-corrected PMI readout does not, and retrieval adds nothing on top. Our results demonstrate that ITCs can be consolidated into the weights of SLMs, an encouraging basis for individualized tutoring agents, and we discuss how to move from there toward a realistic simulation of episodic and semantic memory at the individual level.

Diffusion language models (DLMs) have recently emerged as an alternative modeling paradigm to autoregressive LMs, offering advantages such as parallel generation and bidirectional context modeling. Despite growing interest in their generative capabilities, the privacy risks of DLMs remain underexplored. We identify a phenomenon termed token-level memorization asymmetry through theoretical analysis of diffusion training dynamics. Building on this finding, we propose Q-Skew, a quantile-weighted skewness-based indicator for membership inference on finetuned DLMs. Experiments across multiple fine-tuning datasets and models show that our method outperforms existing baselines. Moreover, we show that Q-Skew can also facilitate other privacy violations, such as PII extraction. Our findings reveal a previously underexplored privacy attack surface and highlight the need for systematic privacy evaluation of DLMs.

Preference optimization is widely used to align large language models with human preferences, but preference-data composition may also influence privacy-relevant memorization. We examine whether adding synthetic privacy-preference pairs to Direct Preference Optimization (DPO) is associated with lower canary-based memorization signals without modifying the objective or introducing a formal privacy mechanism. We propose Privacy-Pressure Preference Mixing (P3M), a data-composition protocol that varies the amount of privacy-preference data while keeping helpfulness and harmlessness preference data fixed. We evaluate a non-privacy Baseline and privacy-mixing ratios of 0.5, 1.0, and 2.0 using Gemma 3 270M-IT across five random seeds and validate the same four conditions using 4-bit-quantized Gemma 2 2B-IT across three seeds. Overall, under the tested conditions, privacy-preference mixing is associated with lower mean canary suffix log-likelihood proxy values across both model settings and lower aggregate membership-inference attack performance relative to the Baseline in the mixed-source 2B evaluation. Specifically, across the privacy-aware 2B configurations, the mean area under the receiver operating characteristic curve (AUROC) ranges from 0.596 to 0.629, and the mean area under the precision-recall curve (AUPRC) ranges from 0.541 to 0.575, compared with 0.804 and 0.790, respectively, for the Baseline. However, the reduction in membership distinguishability does not hold uniformly across data sources. Moreover, the relationship between the privacy ratio and harmlessness preference accuracy varies by model setting, whereas helpfulness preference accuracy remains broadly stable. These findings suggest that P3M should be viewed as a lightweight empirical protocol for examining privacy-utility-safety trade-offs rather than as a formal privacy guarantee or a defense against extraction attacks.

A single training example's contribution to a finished model is normally estimated rather than measured, because measuring it takes two expensive full pre-training runs that differ in one row of one batch. We ran that counterfactual 24 times at a small scale. We trained 32 GPT-2 models at 124M parameters from scratch on OpenWebText, over four conditions and eight seeds. At step 200 of 9,536, at peak learning rate, we replaced one row of a 256-row batch with a fixed context injection carrying a 194-token passage. The three injected conditions are: 1. fluent prose with a corpus-attested subject, 2. fluent prose with a fabricated subject matched to it within 0.14% on full-batch gradient delta, and 3. random keyboard characters. The fourth condition is an uninjected twin. The passage is learned from one exposure and then decays. Fifty steps after injection, the arm that saw a passage predicts it better than the arm that did not by 0.039 and 0.044 nats of cross-entropy on the passage, at eight of eight seeds with p < 0.0001. At the final step we do not detect that difference for either passage, at p = 0.25 and p = 0.71, against minimum detectable effects of 0.025 and 0.079 nats, nor between the two passages, at p = 0.54. Every geometric measure we report is taken after that decay. Our pre-registered contrast on interpolation loss barrier is +0.0068 with p = 0.509, against a minimum detectable effect of 0.032 barrier units. Held-out cross-entropy is -0.00044 with p = 0.310. Per-layer centered kernel alignment does not detectably separate any condition at any layer. Weight displacement reaches 44.1% of the seed-to-seed Euclidean distance and is 92% settled by the midpoint of training, while the barrier reaches 3.0% of the seed-to-seed barrier. Those two figures sit roughly 15 times apart, and that is a lower bound. The injection relocates the model within its basin without moving it out.

The extent to which large language models for code rely on memorization over genuine understanding remains highly debated. While current literature frequently reports widespread memorization, evaluating the underlying probing techniques across dense architectures reveals a severe breakdown in their utility at scale. Traditional encoder-style probes using perturbations such as synonym fuzzing or dead-code insertion struggle to expose memorization in scaled models, even on known-contaminated benchmarks, and decoder-style probes that rely on log probabilities show similar performance degradation. The specific mode of failure for these probes, particularly why such techniques disrupt smaller models but fail to impact larger ones, motivates us to untangle representation load from memorization rather than treating them as a single phenomenon. By applying invertible mathematical transforms to numeric problems, we isolate these two factors and reveal that scaled encoders successfully absorb substantial representation load while still converging on the correct family of solutions. In practical software engineering, this ability to adapt to varying surface forms is what truly matters for usability and generalizability in LLM and agentic applications. Whether a specific solution was seen during training becomes a much less pressing question because although memorization inflates scores on contaminated benchmarks, factoring out representation load makes it debatable how much we should truly care if a functional answer was originally memorized. Future evaluations must therefore be built around separating these phenomena rather than relying on methodologies that quietly entangle them.

Whether idiosyncratic, item-specific knowledge is learned before abstract class-level generalizations, or vice versa, is a central question in language learning, with exemplar and abstraction-based theories making opposite predictions. Recent methods have claimed to show that, at least for large language models, abstract knowledge is learned first. We show that these methods fall short: pure memorizer models with no abstract representations can appear, by the same criteria, to learn either item-specific or class-level knowledge first, depending on their sensitivity to individual observations, with the transition point governed by the distributional properties of the input. We further argue that the distinction between item-specific and abstract knowledge may be ill-defined for distributed representations, as a word's class-level properties may not be separable from its item-specific properties.

Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.

Language models are almost always quantized before they are deployed, and a growing line of work asks whether quantization also lowers their privacy risk. That work measures privacy almost entirely with membership inference. We think this is the wrong thing to measure for the risk that most people actually worry about, namely a model reproducing its training data word for word, and we measure that directly. Using the Pythia models and the public set of sequences each of them is known to have memorized, we track verbatim extraction across five precision levels, from full precision down to four bits, and across three model sizes, while measuring general capability (perplexity) at every point. We find two things. Quantization is a selective forgetter: verbatim memorization falls off faster than capability at every precision and every model size we tried, and this holds under two unrelated quantization algorithms and two evaluation corpora. But the selectivity is not enough to make quantization a privacy defense, which cuts against the optimistic reading of earlier membership-inference results. At the largest model we study, four-bit quantization still reproduces most of the memorized sequences while giving up only a few percent of capability, and the fraction of memorized data that survives quantization grows with model size. We conclude that compression should not be treated as a way to remove memorized training data, and that extraction, not membership inference, is the number practitioners should be watching. All code, sampled evaluation data, and per-configuration results are released.

In this paper, we push the boundary of LLM reasoning by testing them in a Chinese language game, xiehouyu, with novel xiehouyu created by linguists that had not existed before to avoid data contamination. We use multiple-choice questions (MCQ), free-form explanation generation, and new xiehouyu creation to evaluate LLMs' ability to understand and create xiehouyu. In MCQ, we use the delta of accuracy ($Δ_{acc}$) between existing but low-frequency xiehouyu and novel ones as an index for memorization. $Δ_{acc}$ for native speakers is very low, suggesting similar processing mechanisms. However, we found that frontier Chinese models have on average a $Δ_{acc}$ of 23.6%, while English-centric models tested have a mean $Δ_{acc}$ of 5.1%, suggesting that frontier Chinese models are likely trained with much larger Chinese data, thus memorizing more low-frequency xiehouyu. For novel xiehouyu, Gemini 3.1 Pro demonstrated remarkable ability with acc 92.6, which is 24% higher than human accuracy. In xiehouyu creation, those created by LLMs receive much worse ratings than those by humans. These results suggest that claims about the reasoning abilities of LLMs may need careful re-examination considering the data contamination issue, and that LLMs' creativity in language-related tasks may still be behind human experts, at least in Chinese xiehouyu.

LoRA adapters are often shared on their own, and the amount of information they can hold about their training data bounds what sharing them can reveal. Following Morris et al. (2026), we measure this amount in bits by training adapters on frozen language models to memorize random token sequences. LoRA adapters store 2 to 3.4 bits per trainable parameter, less than full fine-tuning of the same model. The number of parameters alone does not set this amount: at equal size, adapters on MLP layers store more than adapters on attention layers, and a randomly initialized frozen model supports as much storage as a pretrained one once the scale of its output logits can be trained. We then train the same adapter on a task with supervised fine-tuning (SFT) and with GRPO. At the same accuracy, SFT stores about three times as much information specific to its training examples, and it stores most of the bits of secrets planted in the training data, which GRPO does not store because it rarely samples them.

While large language models (LLMs) can solve advanced reasoning problems in seconds, we show that even frontier models fail to perform a much simpler operation: exactly copying an input string that lies well within their context windows. We attribute this failure to positional encodings in Transformer architectures, whose inductive bias favors copying through a shortcut based on matching local contexts rather than carefully locating the corresponding input positions. To address this issue, we introduce 2D-RoPE, which organizes text into a 2D grid rather than a 1D sequence and assigns each token a row ID and a column ID. Under this view, copying becomes simply retrieving input tokens at a fixed column offset, which makes the task easy to learn. In synthetic copy experiments, shallow Transformers with 2D-RoPE achieve perfect copying at input lengths hundreds of times longer than those seen during training, whereas standard positional encodings fall far behind. We further show that the advantage of 2D-RoPE language models on copy tasks consistently holds in large-scale pretraining on DCLM with model sizes up to 1.4B parameters. Overall, our results suggest that viewing text in 2D can benefit language modeling, and we hope this encourages future work to further explore the potential of 2D positional encodings.

Recent work on extractable memorization in LLMs suffers from two contrasting validity problems. Some studies overstate extraction, e.g., relying on sequences too short to distinguish memorization from predictability. Others imply that extraction is unreliable evidence of memorization, since models can also reproduce real-world text they weren't explicitly trained on. In different ways, both overlook what makes a valid extraction claim: the model must generate a training sequence with high enough probability to indicate memorization. To determine what's high enough, one has to perform a matched comparison: measuring the generation probabilities of both the training sequences of interest and comparable non-training sequences. Because non-training sequences cannot have been memorized, their probabilities provide a baseline for predictability; a training sequence exceeding this baseline provides evidence of memorization. We formalize matched comparisons in two ways: (1) a conformal test that calibrates a threshold to a chosen FPR when training and non-training sequences are sampled from populations, and (2) a census that calibrates against a matched non-training document when the object is a single document (e.g., a book). We show that matched comparisons enable rigorous, calibrated memorization claims, and reveal where prior setups have validity issues. For instance, on Wikipedia OLMo 2 32B reproduces non-training 10-token suffixes roughly 24% as often as training ones: that share of the training generation rate reflects false positives, not memorization. For Llama 3.1 70B on books, the thresholds we calibrate are as low as 1e-27, supporting memorization claims for sequences that no feasible sampling budget would extract. Based on these results, we refine "extractable memorization" to require a valid memorization claim and near-certain generation within a realistic budget.

Large language models (LLMs) store factual knowledge in their parameters. While recent work has shown that this knowledge resides in MLP layers, existing constructive and mechanistic interpretability models of fact-storage in LLMs fail to explain the surprising empirical phenomenon that they store facts at an information-theoretically optimal rate. In this work, we develop a theoretical account of this phenomenon. We develop the first Transformer-compatible fact-storing MLP closed-form construction that satisfies the following three properties empirically observed in LLMs: it (i) attains optimal fact storage scaling, (ii) handles arbitrary input/output geometries, and (iii) works inside Transformers. Key to our work is to analyze the decoding margin of MLPs, whereas prior work only studies MLP fact storage. Under isotropic embeddings, our construction achieves information-theoretically optimal storage capacity scaling and requires $10$-$104\times$ fewer parameters at matched fact count than prior constructions. For arbitrary key and value embeddings, we show that our construction attains the same storage capacity scaling, up to penalization factors depending on the embedding geometries. Moreover, we demonstrate that our constructed MLPs can be used within Transformer blocks for factual recall tasks at optimal capacity scaling, requiring $15$-$63\times$ fewer parameters at matched fact count than prior constructions. Finally, as a proof-of-concept, we show that fact-storing MLPs enable modular fact editing by swapping a Transformer's MLP with a new one.

Fine-tuning LLMs to inject new knowledge faces a critical challenge: LLMs can quickly memorize new facts, yet fail to use them for downstream reasoning tasks. We formalize this failure as the Knowing-Using Gap, characterized by an accuracy gap and a temporal lag between memorization and generalization. To understand this phenomenon, we fine-tune LLMs with unseen knowledge and monitor the spatial permeation dynamics of the knowledge internally with self-patching, an adaptation of activation patching that scans all layer pairs at every fine-tuning check-point. Self-patching identifies activation locations where relocating representations substantially improves failed generalization cases. These results are consistent with a knowledge-circuit misalignment hypothesis: memorized representations can exist internally but may not be routed to computation-effective layers. Experiments are done cross-domain for the robustness of this finding. Building on this diagnosis, we propose layer-wise representation self-distillation (LRSD) that aligns knowledge representation from late storage layer to middle layer. LRSD keeps improving generalization after fine-tuning saturates and nearly doubles multi-hop chaining accuracy on Qwen, while leaving memorization intact.

The training paradigm of large language models has shifted from traditional one-pass training to multi-epoch training, as reasonable reuse of limited high-quality data can improve both model performance and sample efficiency. Meanwhile, excessive repetition introduces the risk of overfitting and diminishing returns. Determining when and how to reuse data effectively thus emerges as a natural but under-explored question. Through a novel observation of model's "Memorization Window" signals derived from loss retention dynamics and downstream evaluation scores, we propose "Memorization-guided Data Reuse", a training paradigm that adaptively determines when and how data should be reused, enabling principled decisions on the number of training epochs and the scheduling of data replays. Our preliminary experiments reveal a consistent memorization-driven regime: performance continues to improve with repetition far beyond current practice (e.g., the commonly cited four-epoch limit). While a full scheduler remains future work, these insights provide a foundation for memorization-aware training schedules, helping to determine reuse budgets and move toward training LLMs smarter rather than longer with limited high-quality data.

We audit a fixed prefix-window mean-NLL memorization probe (K=20) on a Qwen2.5-VL-7B canary testbed and report three post-hoc cases where it disagrees with full-span secret NLL or greedy exact-recall. C3 (false negative, window truncation): damage lands on hex tokens outside K=20; the probe stays flat while hit@1 drops. C4 (false positive, non-secret drift): the probe moves, but approximately 99% sits on non-secret preamble; the secret span and hit@1 are unchanged. C5 (ambiguous in-window drop): the probe falls on an undertrained baseline while full-span hex is positive and hit@1=0. Recommendation: report (i) full-span secret NLL, (ii) a span-localised decomposition, (iii) behavioural exact-recall at k>=4, and (iv) decoy probes before asserting secret-specificity. Evidence is on controlled canaries in one backbone; magnitudes are testbed-specific.

CSIRTs increasingly fine tune language models on vulnerability scan records, but these records expose internal network topology and create privacy risks under regulations such as GDPR and LGPD. We present the first empirical study of how DP SGD and HMAC pseudonymization interact when fine tuning small language models with 1B to 3B parameters on structured CSIRT data. We evaluate 96 LoRA adapters across four SLMs and four training regimes, including raw fine tuning, QLoRA with large batch training, and DP SGD with epsilon equal to 2 and 8. We also audit memorization using 20 planted canaries, four extraction attacks, and a dual attack targeting HMAC pseudonymized identifiers. Our results show three main findings. First, matched update controls reproduce the observed reduction in memorization by reducing the number of optimizer updates alone, accounting for 66 percent to 132 percent of the measured effect, with a mean of 100 percent across three seeds and four models. In this setting, DP SGD provides the formal privacy guarantee but does not produce additional measurable reductions in memorization. Second, HMAC pseudonymization removes the original identifiers from the exposure surface, reducing exposure by 40 percent to 61 percent, while pseudonymized identifiers remain close to the expected random baseline and do not become a secondary memorization target. Third, F1 scores remain between 0.19 and 0.28 across all 96 adapters using four shot prompting, indicating that, under the evaluated training budget, 1B to 3B SLMs do not achieve operationally useful performance.

Language models are running out of high-quality training data, and even aggressively deduplicated corpora retain some amount of repetition. Earlier controlled studies predated Chinchilla-style scaling laws and could only measure the cost of repetition indirectly. We revisit repetition in the Chinchilla era, using a fitted no-repetition scaling law to report Compute-Equivalent Gain and Compute-Equivalent Loss. We show that under this modernized paradigm, repetition damage is systematic in three ways. First, holding compute allocated to repeated data constant, eval loss peaks at an intermediate repeat count $\Rep$; repeating a moderately sized subset a moderate number of times damages performance more than repeating a large subset a few times or a small subset many times. Second, the location of this peak is well-fit by a power law in model size; this scaling law reveals that the most damaging number of repeated data grows more quickly than compute. Finally, when repeated documents consume 10% of the FLOPs budget in a controlled exact-document repetition setting, the compute-equivalent loss can be large: on FineWeb-Edu-Dedup, the most damaging repeat count for a Qwen3-style 344M-parameter model at $\OT=1$ matches the loss of a no-repetition run using 67% of the FLOPs. We demonstrate that these phenomena are not language-model-specific, and can be analytically understood in a simple statistical model: a misspecified linear regression with verbatim duplicates reproduces the same qualitative loss peak, quantifying how such peaks can arise from a statistical tradeoff between memorization and generalization. Our findings add precision to the study of duplication in language models, allowing practitioners to quantify the wasted compute incurred by the presence and repeat structure of duplicates in pretraining corpora.