Recent work suggests that large language models (LLMs) can perform multi-hop reasoning implicitly -- producing correct answers without explicitly verbalizing intermediate steps -- but the underlying mechanisms remain poorly understood. In this paper, we study how such implicit reasoning emerges by training transformers from scratch in a controlled symbolic environment. Our analysis reveals a three-stage developmental trajectory: early memorization, followed by in-distribution generalization, and eventually cross-distribution generalization. We find that training with atomic triples is not necessary but accelerates learning, and that second-hop generalization relies on query-level exposure to specific compositional structures. To interpret these behaviors, we introduce two diagnostic tools: cross-query semantic patching, which identifies semantically reusable intermediate representations, and a cosine-based representational lens, which reveals that successful reasoning correlates with the cosine-base clustering in hidden space. This clustering phenomenon in turn provides a coherent explanation for the behavioral dynamics observed across training, linking representational structure to reasoning capability. These findings provide new insights into the interpretability of implicit multi-hop reasoning in LLMs, helping to clarify how complex reasoning processes unfold internally and offering pathways to enhance the transparency of such models.
Neural networks transform high-dimensional data into compact, structured representations, often modeled as elements of a lower dimensional latent space. In this paper, we present an alternative interpretation of neural models as dynamical systems acting on the latent manifold. Specifically, we show that autoencoder models implicitly define a latent vector field on the manifold, derived by iteratively applying the encoding-decoding map, without any additional training. We observe that standard training procedures introduce inductive biases that lead to the emergence of attractor points within this vector field. Drawing on this insight, we propose to leverage the vector field as a representation for the network, providing a novel tool to analyze the properties of the model and the data. This representation enables to: (i) analyze the generalization and memorization regimes of neural models, even throughout training; (ii) extract prior knowledge encoded in the network's parameters from the attractors, without requiring any input data; (iii) identify out-of-distribution samples from their trajectories in the vector field. We further validate our approach on vision foundation models, showcasing the applicability and effectiveness of our method in real-world scenarios.
Diffusion models now set the benchmark in high-fidelity generative sampling, yet they can, in principle, be prone to memorization. In this case, their learned score overfits the finite dataset so that the reverse-time SDE samples are mostly training points. In this paper, we interpret the empirical score as a noisy version of the true score and show that its covariance matrix is asymptotically a re-weighted data PCA. In large dimension, the small time limit makes the noise variance blow up while simultaneously reducing spatial correlation. To reduce this variance, we introduce a kernel-smoothed empirical score and analyze its bias-variance trade-off. We derive asymptotic bounds on the Kullback-Leibler divergence between the true distribution and the one generated by the modified reverse SDE. Regularization on the score has the same effect as increasing the size of the training dataset, and thus helps prevent memorization. A spectral decomposition of the forward diffusion suggests better variance control under some regularity conditions of the true data distribution. Reverse diffusion with kernel-smoothed empirical score can be reformulated as a gradient descent drifted toward a Log-Exponential Double-Kernel Density Estimator (LED-KDE). This perspective highlights two regularization mechanisms taking place in denoising diffusions: an initial Gaussian kernel first diffuses mass isotropically in the ambient space, while a second kernel applied in score space concentrates and spreads that mass along the data manifold. Hence, even a straightforward regularization-without any learning-already mitigates memorization and enhances generalization. Numerically, we illustrate our results with several experiments on synthetic and MNIST datasets.
We introduce a new Bayesian perspective on the concept of data reconstruction, and leverage this viewpoint to propose a new security definition that, in certain settings, provably prevents reconstruction attacks. We use our paradigm to shed new light on one of the most notorious attacks in the privacy and memorization literature - fingerprinting code attacks (FPC). We argue that these attacks are really a form of membership inference attacks, rather than reconstruction attacks. Furthermore, we show that if the goal is solely to prevent reconstruction (but not membership inference), then in some cases the impossibility results derived from FPC no longer apply.
Hopfield networks are associative memory (AM) systems, designed for storing and retrieving patterns as local minima of an energy landscape. In the classical Hopfield model, an interesting phenomenon occurs when the amount of training data reaches its critical memory load $- spurious\,\,states$, or unintended stable points, emerge at the end of the retrieval dynamics, leading to incorrect recall. In this work, we examine diffusion models, commonly used in generative modeling, from the perspective of AMs. The training phase of diffusion model is conceptualized as memory encoding (training data is stored in the memory). The generation phase is viewed as an attempt of memory retrieval. In the small data regime the diffusion model exhibits a strong memorization phase, where the network creates distinct basins of attraction around each sample in the training set, akin to the Hopfield model below the critical memory load. In the large data regime, a different phase appears where an increase in the size of the training set fosters the creation of new attractor states that correspond to manifolds of the generated samples. Spurious states appear at the boundary of this transition and correspond to emergent attractor states, which are absent in the training set, but, at the same time, have distinct basins of attraction around them. Our findings provide: a novel perspective on the memorization-generalization phenomenon in diffusion models via the lens of AMs, theoretical prediction of existence of spurious states, empirical validation of this prediction in commonly-used diffusion models.
Neural networks are vulnerable to privacy attacks aimed at stealing sensitive data. The risks can be amplified in a real-world scenario, particularly when models are trained on limited and biased data. In this work, we investigate the impact of spurious correlation bias on privacy vulnerability. We introduce \emph{spurious privacy leakage}, a phenomenon where spurious groups are significantly more vulnerable to privacy attacks than non-spurious groups. We further show that group privacy disparity increases in tasks with simpler objectives (e.g. fewer classes) due to the persistence of spurious features. Surprisingly, we find that reducing spurious correlation using spurious robust methods does not mitigate spurious privacy leakage. This leads us to introduce a perspective on privacy disparity based on memorization, where mitigating spurious correlation does not mitigate the memorization of spurious data, and therefore, neither the privacy level. Lastly, we compare the privacy of different model architectures trained with spurious data, demonstrating that, contrary to prior works, architectural choice can affect privacy outcomes.
Post-hoc interpretability methods typically attribute a model's behavior to its components, data, or training trajectory in isolation. This leads to explanations that lack a unified view and may miss key interactions. While combining existing methods or applying them at different training stages offers broader insights, these approaches usually lack theoretical support. In this work, we present ExPLAIND, a unified framework that integrates all three perspectives. First, we generalize recent work on gradient path kernels, which reformulate models trained by gradient descent as a kernel machine, to more realistic training settings. Empirically, we find that both a CNN and a Transformer model are replicated accurately by this reformulation. Second, we derive novel parameter- and step-wise influence scores from the kernel feature maps. We show their effectiveness in parameter pruning that is comparable to existing methods, reinforcing their value for model component attribution. Finally, jointly interpreting model components and data over the training process, we leverage ExPLAIND to analyze a Transformer that exhibits Grokking. Among other things, our findings support previously proposed stages of Grokking, while refining the final phase as one of alignment of input embeddings and final layers around a representation pipeline learned after the memorization phase. Overall, ExPLAIND provides a theoretically grounded, unified framework to interpret model behavior and training dynamics.
Over-parameterized models have raised concerns about their potential to memorize training data, even when achieving strong generalization. The privacy implications of such memorization are generally unclear, particularly in scenarios where only model outputs are accessible. We study this question in the context of kernel methods, and demonstrate both empirically and theoretically that querying kernel models at various points suffices to reconstruct their training data, even without access to model parameters. Our results hold for a range of kernel methods, including kernel regression, support vector machines, and kernel density estimation. Our hope is that this work can illuminate potential privacy concerns for such models.
Refusal discovery is the task of identifying the full set of topics that a language model refuses to discuss. We introduce this new problem setting and develop a refusal discovery method, LLM-crawler, that uses token prefilling to find forbidden topics. We benchmark the LLM-crawler on Tulu-3-8B, an open-source model with public safety tuning data. Our crawler manages to retrieve 31 out of 36 topics within a budget of 1000 prompts. Next, we scale the crawl to a frontier model using the prefilling option of Claude-Haiku. Finally, we crawl three widely used open-weight models: Llama-3.3-70B and two of its variants finetuned for reasoning: DeepSeek-R1-70B and Perplexity-R1-1776-70B. DeepSeek-R1-70B reveals patterns consistent with censorship tuning: The model exhibits "thought suppression" behavior that indicates memorization of CCP-aligned responses. Although Perplexity-R1-1776-70B is robust to censorship, LLM-crawler elicits CCP-aligned refusals answers in the quantized model. Our findings highlight the critical need for refusal discovery methods to detect biases, boundaries, and alignment failures of AI systems.
Diffusion models have emerged as a powerful class of generative models, capable of producing high-quality samples that generalize beyond the training data. However, evaluating this generalization remains challenging: theoretical metrics are often impractical for high-dimensional data, while no practical metrics rigorously measure generalization. In this work, we bridge this gap by introducing probability flow distance ($\texttt{PFD}$), a theoretically grounded and computationally efficient metric to measure distributional generalization. Specifically, $\texttt{PFD}$ quantifies the distance between distributions by comparing their noise-to-data mappings induced by the probability flow ODE. Moreover, by using $\texttt{PFD}$ under a teacher-student evaluation protocol, we empirically uncover several key generalization behaviors in diffusion models, including: (1) scaling behavior from memorization to generalization, (2) early learning and double descent training dynamics, and (3) bias-variance decomposition. Beyond these insights, our work lays a foundation for future empirical and theoretical studies on generalization in diffusion models.
디퓨전 모델은 최근 다양한 생성 작업에서 뛰어난 성능을 보여주고 있지만, 훈련 데이터를 메모리화하지 않고 일반화할 수 있는 메커니즘에 대한 이해는 여전히 중요한 연구 과제로 남아 있다. 본 연구에서는 이러한 일반화와 메모리화 간의 전환에서 훈련 동역학의 역할을 깊이 있게 조사하였다. 실험과 이론적 분석을 통해 두 가지 주요 시간 척도인 초기 시간 $τ_\mathrm{gen}$과 메모리화가 발생하는 시간 $τ_\mathrm{mem}$을 식별하였다. 특히, $τ_\mathrm{mem}$은 훈련 세트 크기 $n$에 따라 선형적으로 증가하는 반면, $τ_\mathrm{gen}$은 일정하게 유지되어, 훈련 시간이 증가함에 따라 모델이 효과적으로 일반화할 수 있는 창이 형성됨을 발견하였다.
이 연구의 핵심 기여는 훈련 동역학에서의 암묵적 동적 정규화(implicit dynamical regularization) 개념을 도입하여, 고도로 과매개변수화된 설정에서도 메모리화를 피할 수 있는 방법을 제시한 것이다. 또한, 훈련 세트 크기가 모델 의존적인 임계값을 초과할 때만 과적합이 사라진다는 점을 강조하였다. 이러한 발견은 디퓨전 모델의 훈련 과정에서 메모리화가 발생하는 시점을 이해하는 데 중요한 통찰을 제공하며, 모델의 일반화 능력을 유지하는 데 필수적인 요소로 작용한다.
본 연구는 현실적 및 합성 데이터셋에 대한 표준 U-Net 아키텍처를 활용한 수치 실험과 고차원 한계에서의 처리 가능한 랜덤 피처 모델을 통한 이론적 분석을 통해 뒷받침되었다. 이러한 접근 방식은 디퓨전 모델의 훈련 동역학을 이해하는 데 기여하며, 향후 연구에서 일반화와 메모리화 간의 복잡한 관계를 탐구하는 데 중요한 기반이 될 것으로 기대된다.
Large Language Models (LLMs) are powerful but prone to hallucinations due to static knowledge. Retrieval-Augmented Generation (RAG) helps by injecting external information, but current methods often are costly, generalize poorly, or ignore the internal knowledge of the model. In this paper, we introduce R1-Searcher++, a novel framework designed to train LLMs to adaptively leverage both internal and external knowledge sources. R1-Searcher++ employs a two-stage training strategy: an initial SFT Cold-start phase for preliminary format learning, followed by RL for Dynamic Knowledge Acquisition. The RL stage uses outcome-supervision to encourage exploration, incorporates a reward mechanism for internal knowledge utilization, and integrates a memorization mechanism to continuously assimilate retrieved information, thereby enriching the model's internal knowledge. By leveraging internal knowledge and external search engine, the model continuously improves its capabilities, enabling efficient retrieval-augmented reasoning. Our experiments demonstrate that R1-Searcher++ outperforms previous RAG and reasoning methods and achieves efficient retrieval. The code is available at https://github.com/RUCAIBox/R1-Searcher-plus.
Idioms have long posed a challenge due to their unique linguistic properties, which set them apart from other common expressions. While recent studies have leveraged large language models (LLMs) to handle idioms across various tasks, e.g., idiom-containing sentence generation and idiomatic machine translation, little is known about the underlying mechanisms of idiom processing in LLMs, particularly in multilingual settings. To this end, we introduce MIDAS, a new large-scale dataset of idioms in six languages, each paired with its corresponding meaning. Leveraging this resource, we conduct a comprehensive evaluation of LLMs' idiom processing ability, identifying key factors that influence their performance. Our findings suggest that LLMs rely not only on memorization, but also adopt a hybrid approach that integrates contextual cues and reasoning, especially when processing compositional idioms. This implies that idiom understanding in LLMs emerges from an interplay between internal knowledge retrieval and reasoning-based inference.
Fact recall, the ability of language models (LMs) to retrieve specific factual knowledge, remains a challenging task despite their impressive general capabilities. Common training strategies often struggle to promote robust recall behavior with two-stage training, which first trains a model with fact-storing examples (e.g., factual statements) and then with fact-recalling examples (question-answer pairs), tending to encourage rote memorization rather than generalizable fact retrieval. In contrast, mixed training, which jointly uses both types of examples, has been empirically shown to improve the ability to recall facts, but the underlying mechanisms are still poorly understood. In this work, we investigate how these training strategies affect how model parameters are shaped during training and how these differences relate to their ability to recall facts. We introduce cross-task gradient trace to identify shared parameters, those strongly influenced by both fact-storing and fact-recalling examples. Our analysis on synthetic fact recall datasets with the Llama-3.2B and Pythia-2.8B models reveals that mixed training encouraging a larger and more centralized set of shared parameters. These findings suggest that the emergence of parameters may play a key role in enabling LMs to generalize factual knowledge across task formulations.
Existing reasoning benchmarks for large language models (LLMs) frequently fail to capture authentic creativity, often rewarding memorization of previously observed patterns. We address this shortcoming with Sudoku-Bench, a curated benchmark of challenging and unconventional Sudoku variants specifically selected to evaluate creative, multi-step logical reasoning. Sudoku variants form an unusually effective domain for reasoning research: each puzzle introduces unique or subtly interacting constraints, making memorization infeasible and requiring solvers to identify novel logical breakthroughs (``break-ins''). Despite their diversity, Sudoku variants maintain a common and compact structure, enabling clear and consistent evaluation. Sudoku-Bench includes a carefully chosen puzzle set, a standardized text-based puzzle representation, and flexible tools compatible with thousands of publicly available puzzles -- making it easy to extend into a general research environment. Baseline experiments show that state-of-the-art LLMs solve fewer than 15\% of puzzles unaided, highlighting significant opportunities to advance long-horizon, strategic reasoning capabilities.
Neural language models are black-boxes -- both linguistic patterns and factual knowledge are distributed across billions of opaque parameters. This entangled encoding makes it difficult to reliably inspect, verify, or update specific facts. We propose a new class of language models, Large Memory Language Models (LMLM) with a pre-training recipe that stores factual knowledge in both internal weights and an external database. Our approach strategically masks externally retrieved factual values from the training loss, thereby teaching the model to perform targeted lookups rather than relying on memorization in model weights. Our experiments demonstrate that LMLMs achieve competitive performance compared to significantly larger, knowledge-dense LLMs on standard benchmarks, while offering the advantages of explicit, editable, and verifiable knowledge bases. This work represents a fundamental shift in how language models interact with and manage factual knowledge.
We present the first comprehensive study of Memorization in Multilingual Large Language Models (MLLMs), analyzing 95 languages using models across diverse model scales, architectures, and memorization definitions. As MLLMs are increasingly deployed, understanding their memorization behavior has become critical. Yet prior work has focused primarily on monolingual models, leaving multilingual memorization underexplored, despite the inherently long-tailed nature of training corpora. We find that the prevailing assumption, that memorization is highly correlated with training data availability, fails to fully explain memorization patterns in MLLMs. We hypothesize that treating languages in isolation - ignoring their similarities - obscures the true patterns of memorization. To address this, we propose a novel graph-based correlation metric that incorporates language similarity to analyze cross-lingual memorization. Our analysis reveals that among similar languages, those with fewer training tokens tend to exhibit higher memorization, a trend that only emerges when cross-lingual relationships are explicitly modeled. These findings underscore the importance of a language-aware perspective in evaluating and mitigating memorization vulnerabilities in MLLMs. This also constitutes empirical evidence that language similarity both explains Memorization in MLLMs and underpins Cross-lingual Transferability, with broad implications for multilingual NLP.
We introduce the concept of protoknowledge to formalize and measure how sequences of tokens encoding Knowledge Graphs are internalized during pretraining and utilized at inference time by Large Language Models (LLMs). Indeed, LLMs have demonstrated the ability to memorize vast amounts of token sequences during pretraining, and a central open question is how they leverage this memorization as reusable knowledge through generalization. We then categorize protoknowledge into lexical, hierarchical, and topological forms, varying on the type of knowledge that needs to be activated. We measure protoknowledge through Knowledge Activation Tasks (KATs), analyzing its general properties such as semantic bias. We then investigate the impact of protoknowledge on Text-to-SPARQL performance by varying prompting strategies depending on input conditions. To this end, we adopt a novel analysis framework that assesses whether model predictions align with the successful activation of the relevant protoknowledge for each query. This methodology provides a practical tool to explore Semantic-Level Data Contamination and serves as an effective strategy for Closed-Pretraining models.
Aiming at resisting backdoor attacks in convolution neural networks and vision Transformer-based large model, this paper proposes a generalized and model-agnostic trigger-purification approach resorting to the classic Ising model. To date, existing trigger detection/removal studies usually require to know the detailed knowledge of target model in advance, access to a large number of clean samples or even model-retraining authorization, which brings the huge inconvenience for practical applications, especially inaccessible to target model. An ideal countermeasure ought to eliminate the implanted trigger without regarding whatever the target models are. To this end, a lightweight and black-box defense approach SifterNet is proposed through leveraging the memorization-association functionality of Hopfield network, by which the triggers of input samples can be effectively purified in a proper manner. The main novelty of our proposed approach lies in the introduction of ideology of Ising model. Extensive experiments also validate the effectiveness of our approach in terms of proper trigger purification and high accuracy achievement, and compared to the state-of-the-art baselines under several commonly-used datasets, our SiferNet has a significant superior performance.
Causal world models are systems that can answer counterfactual questions about an environment of interest, i.e. predict how it would have evolved if an arbitrary subset of events had been realized differently. It requires understanding the underlying causes behind chains of events and conducting causal inference for arbitrary unseen distributions. So far, this task eludes foundation models, notably large language models (LLMs), which do not have demonstrated causal reasoning capabilities beyond the memorization of existing causal relationships. Furthermore, evaluating counterfactuals in real-world applications is challenging since only the factual world is observed, limiting evaluation to synthetic datasets. We address these problems by explicitly extracting and modeling causal relationships and propose the Causal Cartographer framework. First, we introduce a graph retrieval-augmented generation agent tasked to retrieve causal relationships from data. This approach allows us to construct a large network of real-world causal relationships that can serve as a repository of causal knowledge and build real-world counterfactuals. In addition, we create a counterfactual reasoning agent constrained by causal relationships to perform reliable step-by-step causal inference. We show that our approach can extract causal knowledge and improve the robustness of LLMs for causal reasoning tasks while reducing inference costs and spurious correlations.