Continual Learning: Navigating the Future of Adaptive AI

Latest 50 papers on continual learning: Nov. 30, 2025

The dream of AI that learns continuously, adapting to new information without forgetting the old, has long been a holy grail in machine learning. However, the notorious ‘catastrophic forgetting’ dilemma has remained a formidable barrier. Recent breakthroughs, as showcased in a collection of cutting-edge research, are paving the way for truly adaptive AI systems capable of lifelong learning. These innovations span diverse domains, from medical imaging and robotics to cybersecurity and communication systems, marking a pivotal moment in the quest for intelligent agents that evolve with their environments.

The Big Idea(s) & Core Innovations

At the heart of these advancements is the persistent effort to mitigate catastrophic forgetting and enhance model plasticity. Several papers tackle this challenge by introducing novel architectural designs and training paradigms. For instance, researchers from JPMorgan Chase, in their paper “Continual Learning of Domain Knowledge from Human Feedback in Text-to-SQL”, leverage human feedback to distill tacit domain knowledge into structured memory, enabling Text-to-SQL agents to continuously refine their performance. Similarly, Z. Gao and P. Morel introduce Prompt-Aware Adaptive Elastic Weight Consolidation for Continual Learning in Medical Vision-Language Models, significantly reducing forgetting in medical AI by selectively protecting parameters based on task-specific linguistic patterns.

The idea of dynamic adaptation is further explored in “Dynamic Nested Hierarchies: Pioneering Self-Evolution in Machine Learning Architectures for Lifelong Intelligence” by Akbar Anbar Jafari et al. from University of Tartu, which proposes self-evolving architectures that autonomously adjust optimization levels and frequencies. This neuroplasticity-inspired approach enables models to adapt to non-stationary environments. Another intriguing angle comes from Hyung-Jun Moon and Sung-Bae Cho at Yonsei University in their work, “Expandable and Differentiable Dual Memories with Orthogonal Regularization for Exemplar-free Continual Learning”, which introduces dual memory architectures to explicitly store shared and task-specific knowledge, achieving state-of-the-art results without needing exemplar buffers.

A groundbreaking shift is seen in “AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection” by Saleh Momeni et al. from University of Illinois Chicago. They propose an analytic, gradient-free method for class-incremental learning that avoids catastrophic forgetting entirely, achieving performance comparable to joint training, an impressive feat that challenges conventional gradient-based approaches. For the first time, in “Intrinsic preservation of plasticity in continual quantum learning”, Yi Q Chen and Shi Xin Zhang reveal that quantum neural networks inherently preserve plasticity, offering a structural advantage over classical models in continual learning.

Under the Hood: Models, Datasets, & Benchmarks

These advancements are often powered by innovative architectures, specialized datasets, and rigorous benchmarking, pushing the boundaries of continual learning:

• OpenCML Framework (OpenCML: End-to-End Framework of Open-world Machine Learning to Learn Unknown Classes Incrementally by Jitendra Parmar et al.): An end-to-end framework for open-world machine learning that incrementally learns unknown classes using custom loss functions and BIRCH clustering.
• MSVQA Dataset and UNIFIER Framework (Multimodal Continual Learning with MLLMs from Multi-scenario Perspectives by Kai Jiang et al. from Northwestern Polytechnical University): Introduces a new dataset and the UNIFIER framework to study and mitigate catastrophic forgetting in Multimodal Large Language Models (MLLMs) across diverse visual scenarios.
• Stellar VLA Framework (Continually Evolving Skill Knowledge in Vision Language Action Model by Yuxuan Wu et al. from Shanghai Jiao Tong University): A knowledge-driven continual learning framework for Vision-Language-Action (VLA) models, featuring task-centric and hierarchical task-skill variants. Code is expected on GitHub.
• MedPEFT-CL Framework (MedPEFT-CL: Dual-Phase Parameter-Efficient Continual Learning with Medical Semantic Adapter and Bidirectional Memory Consolidation by Ziyuan Gao from University College London): A parameter-efficient framework for medical vision-language segmentation tasks, using semantic adapters and bidirectional memory consolidation. Code available at https://github.com/ziyuan-gao/MedPEFT-CL.
• CLTS (Continual Learning via Text-Image Synergy) (Caption, Create, Continue: Continual Learning with Pre-trained Generative Vision-Language Models by Indu Solomon et al. from IIITB): A modular architecture leveraging BLIP and Stable Diffusion to use text captions for replay, reducing memory by 63x. Code is likely at https://github.com/iiitb-nlpir/CLTS.
• FSC-Net (Fast-Slow Consolidation Networks) (FSC-Net: Fast-Slow Consolidation Networks for Continual Learning by Mohamed El Gorrim from United Arab Emirates University): A dual-network architecture inspired by neuroscience, separating fast task adaptation from slow knowledge consolidation, primarily through pure replay. Code at https://github.com/MedGm/FSC-Net.
• LwP (Learning with Preserving) Framework (Learning with Preserving for Continual Multitask Learning by Hanchen David Wang et al. from Vanderbilt University): Addresses Continual Multitask Learning (CMTL) by preserving the geometric structure of shared representation spaces without a replay buffer. Code available at https://github.com/AICPS-Lab/lwp.
• PIECE (Parameter Importance Estimation-based Continual Enhancement) (Parameter Importance-Driven Continual Learning for Foundation Models by Lingxiang Wang et al. from Beihang University): Selectively updates a minimal subset of parameters (0.1%) in foundation models using Fisher Information or second-order normalization. Code at https://github.com/wanglingxiang0717/PIECE.
• Hydra Mitigation Method (Data Heterogeneity and Forgotten Labels in Split Federated Learning by Joana Tirana et al. from University College Dublin): Addresses catastrophic forgetting in Split Federated Learning by training multiple copies of the last layers in part-2 of the model. Code at https://github.com/jtirana98/Hydra-CF-in-SFL.
• WebCoach Framework (WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance by Genglin Liu et al. from UCLA and Amazon): Enhances web agents with persistent cross-session memory and retrieval-based coaching, improving robustness and efficiency on the WebVoyager benchmark. Code at https://github.com/genglinliu/WebCoach.
• KAN-LoRA (Catastrophic Forgetting in Kolmogorov-Arnold Networks by Mohammad Marufur Rahman et al. from Wake Forest University): A Kolmogorov-Arnold Network-based adapter for continual fine-tuning of Language Models, investigated for its forgetting properties and performance in knowledge editing. Code at https://github.com/marufur-cs/AAAI26.
• CoSO (Continuous Subspace Optimization) (Continuous Subspace Optimization for Continual Learning by Quan Cheng et al. from Nanjing University): A framework that fine-tunes pre-trained models within multiple gradient-derived subspaces to mitigate catastrophic forgetting. Code at https://github.com/lamda-nju/CoSO.

Impact & The Road Ahead

These advancements herald a future where AI systems are not only powerful but also perpetually adaptive and resilient. The ability to learn continually without forgetting past knowledge is crucial for real-world deployment in dynamic environments—from self-driving cars needing to adapt to new road conditions (as highlighted in “Continual Reinforcement Learning for Cyber-Physical Systems: Lessons Learned and Open Challenges” by Kim N. Nolle et al.) to medical AI continually integrating new diagnostic protocols. The development of exemplar-free methods like PANDA (Patch And Distribution-Aware Augmentation for Long-Tailed Exemplar-Free Continual Learning by Siddeshwar Raghavan et al.) and dual-memory architectures further addresses privacy concerns and computational constraints inherent in lifelong learning. Furthermore, the theoretical insights into quantum neural networks and analytic methods offer radically new paradigms for developing robust continual learners. The road ahead involves further pushing the boundaries of efficiency, scalability, and robustness, ensuring that AI can not only learn but truly evolve in an ever-changing world.