LibContinual: A Comprehensive Library towards Realistic Continual Learning

TL;DR

Continual Learning faces catastrophic forgetting under non-stationary data streams. LibContinual delivers a unified, modular library that integrates 19 CL algorithms across five families, supports realistic online settings, and introduces a unified memory-budget protocol plus a category-randomized semantic setting to stress-test robustness. The study shows that training-from-scratch methods collapse offline assumptions in online CL, while PTM-based, parameter-efficient approaches achieve strong performance with modest memory. Results also reveal that clean memory scaling does not guarantee better accuracy, and many methods rely on semantic structure, underscoring the need for resource-aware and semantically robust strategies. Overall, LibContinual provides a reproducible, extensible platform to benchmark and develop more practical continual learning systems.

Abstract

A fundamental challenge in Continual Learning (CL) is catastrophic forgetting, where adapting to new tasks degrades the performance on previous ones. While the field has evolved with diverse methods, this rapid surge in diverse methodologies has culminated in a fragmented research landscape. The lack of a unified framework, including inconsistent implementations, conflicting dependencies, and varying evaluation protocols, makes fair comparison and reproducible research increasingly difficult. To address this challenge, we propose LibContinual, a comprehensive and reproducible library designed to serve as a foundational platform for realistic CL. Built upon a high-cohesion, low-coupling modular architecture, LibContinual integrates 19 representative algorithms across five major methodological categories, providing a standardized execution environment. Meanwhile, leveraging this unified framework, we systematically identify and investigate three implicit assumptions prevalent in mainstream evaluation: (1) offline data accessibility, (2) unregulated memory resources, and (3) intra-task semantic homogeneity. We argue that these assumptions often overestimate the real-world applicability of CL methods. Through our comprehensive analysis using strict online CL settings, a novel unified memory budget protocol, and a proposed category-randomized setting, we reveal significant performance drops in many representative CL methods when subjected to these real-world constraints. Our study underscores the necessity of resource-aware and semantically robust CL strategies, and offers LibContinual as a foundational toolkit for future research in realistic continual learning. The source code is available from \href{https://github.com/RL-VIG/LibContinual}{https://github.com/RL-VIG/LibContinual}.

Figures (8)

• Figure 1: Conceptual illustration of the three implicit assumptions in continual learning evaluation and our proposed experimental dimensions to investigate them. Standard evaluation paradigms (gray) often rely on idealized conditions. (1) Data Stream (the Orange Axis): They assume offline access to task data for multi-epoch training, whereas we test models in a strict single-pass online CL setting. (2) Storage Constraint (the Teal Axis): They permit inconsistent memory cost accounting, making fair comparison difficult; we enforce a unified memory budget to normalize evaluation. (3) Semantic Structure (the Yellow Axis): They typically use tasks with high intra-task semantic homogeneity. We introduce a more challenging category-randomized setting, preventing models from relying on task-level shortcuts and thus testing for more robust representations.
• Figure 2: Architecture of the proposed Libcontinual.
• Figure 3: An illustration of the three continual learning settings defined by inter- and intra-task semantic structure. In the traditional setting, tasks are subsets of a single dataset, making them semantically homogeneous. In the cross-domain setting, each task remains homogeneous but originates from a different domain, introducing a domain shift between tasks. In our proposed category-randomized setting, the assumption of intra-task homogeneity is broken. Each task is a heterogeneous mix of classes from different domains, forcing the model to learn disparate concepts simultaneously.
• Figure 4: The taxonomy of continual learning methods categorizes them into five major algorithmic strategies: regularization-based, replay-based, optimization-based, representation-based, and architecture-based methods.
• Figure 5: The taxonomy of continual learning methods from a storage-centric perspective. The five categories, image-based, feature-based, model-based, parameter-based, and prompt-based, are illustrated with representative examples of the type of content stored in memory.