Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test

• URL: http://arxiv.org/abs/2506.21551v3
• Date: Fri, 03 Oct 2025 21:00:58 GMT
• Title: Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test
• Authors: Ziyue Li, Chenrui Fan, Tianyi Zhou,
• Abstract summary: We investigate when an LLM memorizes the training data, when its generalization on downstream tasks starts to improve, and what happens if there is a lag between the two.<n>Our study, for the first time, shows that grokking still emerges in pretraining mixture-of-experts (MoE)<n>Our primary discovery is that the pathways evolve from random, non-smooth across layers, instance-specific to more structured and transferable across samples, despite the converged pretraining loss.
• Score: 22.499052329934603
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: This paper presents the first study of grokking in practical LLM pretraining. Specifically, we investigate when an LLM memorizes the training data, when its generalization on downstream tasks starts to improve, and what happens if there is a lag between the two. Unlike existing works studying when a small model generalizes to limited and specified tasks during thousands epochs' training on algorithmic data, we focus on a practical setting for LLMs, i.e., one-epoch pretraining of next-token prediction on a cross-domain, large-scale corpus, and generalization on diverse benchmark tasks covering math/commonsense reasoning, code generation, and domain-specific retrieval. Our study, for the first time, verifies that grokking still emerges in pretraining mixture-of-experts (MoE) LLMs, though different local data groups may enter their grokking stages asynchronously due to the heterogeneity of their distributions and attributions to others. To find a mechanistic interpretation of this local grokking, we investigate the dynamics of training data's pathways (i.e., expert choices across layers in MoE). Our primary discovery is that the pathways evolve from random, non-smooth across layers, instance-specific to more structured and transferable across samples, despite the converged pretraining loss. This depicts a transition from memorization to generalization. Two novel metrics are developed to quantify these patterns: one computes the pathway similarity between samples, while the other measures the consistency of aggregated experts between subsequent layers for each sample. These training data based metrics induce zero cost but can faithfully track and monitor the generalization of LLMs on downstream tasks, which, in conventional settings, requires costly instruction tuning and benchmark evaluation.

Related papers

• Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs [21.541258368039955]
  We find that layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample.<n>In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample.
  arXiv  Detail & Related papers  (2025-07-10T17:59:53Z)
• Low-Perplexity LLM-Generated Sequences and Where To Find Them [0.0]
  We introduce a systematic approach centered on analyzing low-perplexity sequences - high-probability text spans generated by the model.<n>Our pipeline reliably extracts such long sequences across diverse topics while avoiding degeneration, then traces them back to their sources in the training data.<n>For those that do match, we quantify the distribution of occurrences across source documents, highlighting the scope and nature of verbatim recall.
  arXiv  Detail & Related papers  (2025-07-02T15:58:51Z)
• Grokking Explained: A Statistical Phenomenon [4.113597666007784]
  Grokking, or delayed generalization, is an intriguing learning phenomenon where test set loss decreases sharply only after a model's training set loss has converged.<n>This paper formalizes and investigates grokking, highlighting that a key factor in its emergence is a distribution shift between training and test data.
  arXiv  Detail & Related papers  (2025-02-03T19:28:11Z)
• Exact, Tractable Gauss-Newton Optimization in Deep Reversible Architectures Reveal Poor Generalization [52.16435732772263]
  Second-order optimization has been shown to accelerate the training of deep neural networks in many applications. However, generalization properties of second-order methods are still being debated. We show for the first time that exact Gauss-Newton (GN) updates take on a tractable form in a class of deep architectures.
  arXiv  Detail & Related papers  (2024-11-12T17:58:40Z)
• What Do Learning Dynamics Reveal About Generalization in LLM Reasoning? [83.83230167222852]
  We find that a model's generalization behavior can be effectively characterized by a training metric we call pre-memorization train accuracy. By connecting a model's learning behavior to its generalization, pre-memorization train accuracy can guide targeted improvements to training strategies.
  arXiv  Detail & Related papers  (2024-11-12T09:52:40Z)
• A Closer Look at Benchmarking Self-Supervised Pre-training with Image Classification [51.35500308126506]
  Self-supervised learning (SSL) is a machine learning approach where the data itself provides supervision, eliminating the need for external labels. We study how classification-based evaluation protocols for SSL correlate and how well they predict downstream performance on different dataset types.
  arXiv  Detail & Related papers  (2024-07-16T23:17:36Z)
• SoK: Membership Inference Attacks on LLMs are Rushing Nowhere (and How to Fix It) [16.673210422615348]
  More than 10 new methods have been proposed to perform Membership Inference Attacks (MIAs) against LLMs.<n>Contrary to traditional MIAs which rely on fixed-but randomized-records or models, these methods are mostly trained and tested on datasets collected post-hoc.<n>This lack of randomization raises concerns of a distribution shift between members and non-members.
  arXiv  Detail & Related papers  (2024-06-25T23:12:07Z)
• The Right Time Matters: Data Arrangement Affects Zero-Shot Generalization in Instruction Tuning [86.19804569376333]
  We show that zero-shot generalization happens very early during instruction tuning.<n>We propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement.
  arXiv  Detail & Related papers  (2024-06-17T16:40:21Z)
• Alpaca against Vicuna: Using LLMs to Uncover Memorization of LLMs [61.04246774006429]
  We introduce a black-box prompt optimization method that uses an attacker LLM agent to uncover higher levels of memorization in a victim agent.<n>We observe that our instruction-based prompts generate outputs with 23.7% higher overlap with training data compared to the baseline prefix-suffix measurements.<n>Our findings show that instruction-tuned models can expose pre-training data as much as their base-models, if not more so, and using instructions proposed by other LLMs can open a new avenue of automated attacks.
  arXiv  Detail & Related papers  (2024-03-05T19:32:01Z)
• Take the Bull by the Horns: Hard Sample-Reweighted Continual Training Improves LLM Generalization [165.98557106089777]
  A key challenge is to enhance the capabilities of large language models (LLMs) amid a looming shortage of high-quality training data. Our study starts from an empirical strategy for the light continual training of LLMs using their original pre-training data sets. We then formalize this strategy into a principled framework of Instance-Reweighted Distributionally Robust Optimization.
  arXiv  Detail & Related papers  (2024-02-22T04:10:57Z)
• Task-customized Masked AutoEncoder via Mixture of Cluster-conditional Experts [104.9871176044644]
  Masked Autoencoder(MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. We propose a novel MAE-based pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE) MoCE trains each expert only with semantically relevant images by using cluster-conditional gates.
  arXiv  Detail & Related papers  (2024-02-08T03:46:32Z)
• On Inter-dataset Code Duplication and Data Leakage in Large Language Models [4.148857672591562]
  This paper explores the phenomenon of inter-dataset code duplication and its impact on evaluating large language models (LLMs) Our findings reveal a potential threat to the evaluation of LLMs across multiple SE tasks, stemming from the inter-dataset code duplication phenomenon. We provide evidence that open-source models could be affected by inter-dataset duplication.
  arXiv  Detail & Related papers  (2024-01-15T19:46:40Z)
• Understanding Emergent In-Context Learning from a Kernel Regression Perspective [55.95455089638838]
  Large language models (LLMs) have initiated a paradigm shift in transfer learning.<n>This paper proposes a kernel-regression perspective of understanding LLMs' ICL bahaviors when faced with in-context examples.<n>We find that during ICL, the attention and hidden features in LLMs match the behaviors of a kernel regression.
  arXiv  Detail & Related papers  (2023-05-22T06:45:02Z)
• Theoretical Characterization of the Generalization Performance of Overfitted Meta-Learning [70.52689048213398]
  This paper studies the performance of overfitted meta-learning under a linear regression model with Gaussian features. We find new and interesting properties that do not exist in single-task linear regression. Our analysis suggests that benign overfitting is more significant and easier to observe when the noise and the diversity/fluctuation of the ground truth of each training task are large.
  arXiv  Detail & Related papers  (2023-04-09T20:36:13Z)
• Learning Curves for Sequential Training of Neural Networks: Self-Knowledge Transfer and Forgetting [9.734033555407406]
  We consider neural networks in the neural tangent kernel regime that continually learn target functions from task to task. We investigate a variant of continual learning where the model learns the same target function in multiple tasks. Even for the same target, the trained model shows some transfer and forgetting depending on the sample size of each task.
  arXiv  Detail & Related papers  (2021-12-03T00:25:01Z)
• What training reveals about neural network complexity [80.87515604428346]
  This work explores the hypothesis that the complexity of the function a deep neural network (NN) is learning can be deduced by how fast its weights change during training. Our results support the hypothesis that good training behavior can be a useful bias towards good generalization.
  arXiv  Detail & Related papers  (2021-06-08T08:58:00Z)
• Double Descent and Other Interpolation Phenomena in GANs [2.7007335372861974]
  We study the generalization error as a function of latent space dimension in generative adversarial networks (GANs) We develop a novel pseudo-supervised learning approach for GANs where the training utilizes pairs of fabricated (noise) inputs in conjunction with real output samples. While our analysis focuses mostly on linear models, we also apply important insights for improving generalization of nonlinear, multilayer GANs.
  arXiv  Detail & Related papers  (2021-06-07T23:07:57Z)
• Exploring Memorization in Adversarial Training [58.38336773082818]
  We investigate the memorization effect in adversarial training (AT) for promoting a deeper understanding of capacity, convergence, generalization, and especially robust overfitting. We propose a new mitigation algorithm motivated by detailed memorization analyses.
  arXiv  Detail & Related papers  (2021-06-03T05:39:57Z)
• Adversarial Encoder-Multi-Task-Decoder for Multi-Stage Processes [5.933303832684138]
  In multi-stage processes, decisions occur in an ordered sequence of stages. We introduce a framework that combines adversarial autoencoders (AAE), multi-task learning (MTL), and multi-label semi-supervised learning (MLSSL) Using real-world data from different domains, we show that our approach outperforms other state-of-the-art methods.
  arXiv  Detail & Related papers  (2020-03-15T19:30:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.