Revisiting Replay and Gradient Alignment for Continual Pre-Training of Large Language Models

• URL: http://arxiv.org/abs/2508.01908v1
• Date: Sun, 03 Aug 2025 20:07:15 GMT
• Title: Revisiting Replay and Gradient Alignment for Continual Pre-Training of Large Language Models
• Authors: Istabrak Abbes, Gopeshh Subbaraj, Matthew Riemer, Nizar Islah, Benjamin Therien, Tsuguchika Tabaru, Hiroaki Kingetsu, Sarath Chandar, Irina Rish,
• Abstract summary: Training large language models typically involves pre-training on massive corpora.<n>New data often causes distribution shifts, leading to performance degradation on previously learned tasks.<n>We take a deeper look at two popular proposals for addressing this distribution shift: experience replay and gradient alignment.
• Score: 19.136589266017694
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Training large language models (LLMs) typically involves pre-training on massive corpora, only to restart the process entirely when new data becomes available. A more efficient and resource-conserving approach would be continual pre-training, where models are updated with new data rather than retraining from scratch. However, the introduction of new data often causes distribution shifts, leading to performance degradation on previously learned tasks. In this paper, we take a deeper look at two popular proposals for addressing this distribution shift within the continual learning literature: experience replay and gradient alignment. We consider continual pre-training of models within the Llama family of architectures at a large scale across languages with 100 billion tokens of training data in each language, finding that both replay and gradient alignment lead to more stable learning without forgetting. This conclusion holds both as we vary the model scale and as we vary the number and diversity of tasks. Moreover, we are the first to demonstrate the effectiveness of gradient alignment techniques in the context of LLM pre-training and propose an efficient implementation of meta-experience replay (MER) that imbues experience replay with the benefits of gradient alignment despite negligible compute and memory overhead. Our scaling analysis across model sizes and replay rates indicates that small rates of replaying old examples are definitely a more valuable use of compute than investing in model size, but that it is more compute efficient to scale the size of the model than invest in high rates of replaying old examples.

Related papers

• Exploring Training and Inference Scaling Laws in Generative Retrieval [50.82554729023865]
  Generative retrieval reformulates retrieval as an autoregressive generation task, where large language models generate target documents directly from a query.<n>We systematically investigate training and inference scaling laws in generative retrieval, exploring how model size, training data scale, and inference-time compute jointly influence performance.
  arXiv  Detail & Related papers  (2025-03-24T17:59:03Z)
• Prioritized Generative Replay [121.83947140497655]
  We propose a prioritized, parametric version of an agent's memory, using generative models to capture online experience.<n>This paradigm enables densification of past experience, with new generations that benefit from the generative model's generalization capacity.<n>We show this recipe can be instantiated using conditional diffusion models and simple relevance functions.
  arXiv  Detail & Related papers  (2024-10-23T17:59:52Z)
• Accelerating Large Language Model Pretraining via LFR Pedagogy: Learn, Focus, and Review [50.78587571704713]
  Learn-Focus-Review (LFR) is a dynamic training approach that adapts to the model's learning progress.<n>LFR tracks the model's learning performance across data blocks (sequences of tokens) and prioritizes revisiting challenging regions of the dataset.<n>Compared to baseline models trained on the full datasets, LFR consistently achieved lower perplexity and higher accuracy.
  arXiv  Detail & Related papers  (2024-09-10T00:59:18Z)
• Adaptive Memory Replay for Continual Learning [29.333341368722653]
  Updating Foundation Models as new data becomes available can lead to catastrophic forgetting' We introduce a framework of adaptive memory replay for continual learning, where sampling of past data is phrased as a multi-armed bandit problem. We demonstrate the effectiveness of our approach, which maintains high performance while reducing forgetting by up to 10% at no training efficiency cost.
  arXiv  Detail & Related papers  (2024-04-18T22:01:56Z)
• Enhancing Consistency and Mitigating Bias: A Data Replay Approach for Incremental Learning [93.90047628101155]
  Deep learning systems are prone to catastrophic forgetting when learning from a sequence of tasks.<n>To address this, some methods propose replaying data from previous tasks during new task learning.<n>However, it is not expected in practice due to memory constraints and data privacy issues.
  arXiv  Detail & Related papers  (2024-01-12T12:51:12Z)
• Learn to Unlearn for Deep Neural Networks: Minimizing Unlearning Interference with Gradient Projection [56.292071534857946]
  Recent data-privacy laws have sparked interest in machine unlearning. Challenge is to discard information about the forget'' data without altering knowledge about remaining dataset. We adopt a projected-gradient based learning method, named as Projected-Gradient Unlearning (PGU) We provide empirically evidence to demonstrate that our unlearning method can produce models that behave similar to models retrained from scratch across various metrics even when the training dataset is no longer accessible.
  arXiv  Detail & Related papers  (2023-12-07T07:17:24Z)
• On the Costs and Benefits of Adopting Lifelong Learning for Software Analytics -- Empirical Study on Brown Build and Risk Prediction [17.502553991799832]
  This paper evaluates the use of lifelong learning (LL) for industrial use cases at Ubisoft. LL is used to continuously build and maintain ML-based software analytics tools using an incremental learner that progressively updates the old model using new data.
  arXiv  Detail & Related papers  (2023-05-16T21:57:16Z)
• Instance-Conditional Timescales of Decay for Non-Stationary Learning [11.90763787610444]
  Slow concept drift is a ubiquitous, yet under-studied problem in machine learning systems. We propose an optimization-driven approach towards balancing instance importance over large training windows. Experiments on a large real-world dataset of 39M photos over a 9 year period show upto 15% relative gains in accuracy.
  arXiv  Detail & Related papers  (2022-12-12T14:16:26Z)
• One Backward from Ten Forward, Subsampling for Large-Scale Deep Learning [35.0157090322113]
  Large-scale machine learning systems are often continuously trained with enormous data from production environments. The sheer volume of streaming data poses a significant challenge to real-time training subsystems and ad-hoc sampling is the standard practice. We propose to record a constant amount of information per instance from these forward passes. The extra information measurably improves the selection of which data instances should participate in forward and backward passes.
  arXiv  Detail & Related papers  (2021-04-27T11:29:02Z)
• Extrapolation for Large-batch Training in Deep Learning [72.61259487233214]
  We show that a host of variations can be covered in a unified framework that we propose. We prove the convergence of this novel scheme and rigorously evaluate its empirical performance on ResNet, LSTM, and Transformer.
  arXiv  Detail & Related papers  (2020-06-10T08:22:41Z)
• Efficient Learning of Model Weights via Changing Features During Training [0.0]
  We propose a machine learning model, which dynamically changes the features during training. Our main motivation is to update the model in a small content during the training process with replacing less descriptive features to new ones from a large pool.
  arXiv  Detail & Related papers  (2020-02-21T12:38:14Z)

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.