Continual Learning: Navigating Non-Stationary Worlds with Dynamic Architectures and Adaptive Forgetting

Latest 30 papers on continual learning: May. 9, 2026

The world isn’t static, and neither should our AI models be. In the rapidly evolving landscape of AI/ML, the ability for models to learn new tasks or adapt to changing data distributions without forgetting previously acquired knowledge – a challenge known as continual learning (CL) – is paramount. This digest explores a fascinating collection of recent research, revealing breakthroughs across diverse domains, from medical imaging and robotics to traffic prediction and even physical systems, all striving to overcome catastrophic forgetting and achieve true lifelong learning.

The Big Idea(s) & Core Innovations

At the heart of continual learning lies the stability-plasticity dilemma: how to remain stable on old tasks while adapting plastically to new ones. Recent work tackles this from multiple angles, often leveraging dynamic architectures, adaptive parameter management, and novel data strategies.

One significant trend is the use of Mixture-of-Experts (MoE) and parameter-efficient fine-tuning (PEFT) techniques like LoRA. For instance, in “Scene-Adaptive Continual Learning for CSI-based Human Activity Recognition with Mixture of Experts”, researchers from The Hong Kong Polytechnic University propose SAMoE-C, which uses an attention-based semantic router to activate scene-specific expert networks, ensuring constant inference cost while adapting to new environments. Similarly, “SPHERE: Mitigating the Loss of Spectral Plasticity in Mixture-of-Experts for Deep Reinforcement Learning” by Peking University and BIGAI formalizes plasticity loss in MoE policies as a decline in spectral plasticity, introducing SPHERE to maintain spectral diversity and prevent rank collapse, crucial for continual reinforcement learning.

For Large Language Models (LLMs), adaptive intervention in representation space is emerging. “CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning” from Iowa State University proposes a framework that learns low-rank interventions on hidden representations using KL divergence as a unified signal for routing, regularization, and merging, significantly reducing trainable parameters compared to LoRA. Complementing this, “Skill Neologisms: Towards Skill-based Continual Learning” by the University of Cambridge introduces soft tokens, called skill neologisms, integrated into the vocabulary to learn composable skills without modifying model weights, enabling zero-shot composition of independently learned skills.

Replay-based strategies continue to evolve, becoming more intelligent and resource-efficient. “Replay-Based Continual Learning for Physics-Informed Neural Operators” from Tsinghua University and Bauhaus-Universität Weimar applies replay with distillation and LoRA for physics-informed neural operators, selectively replaying only poorly performing samples. For EEG signals, Beihang University’s “Adaptive Data Compression and Reconstruction for Memory-Bounded EEG Continual Learning” (ADaCoRe) uses saliency-driven keyframe protection and polyphase compression to preserve critical information under strict memory constraints, achieving superior performance with significantly smaller buffers.

Interestingly, the very nature of forgetting itself is being re-examined. In “Learning to Forget: Continual Learning with Adaptive Weight Decay” by The Swiss AI Lab and University of Alberta, FADE dynamically adapts per-parameter weight decay rates online via meta-gradient descent, viewing decay as a controlled forgetting mechanism. Even physical systems, as explored in “Sequential Learning and Catastrophic Forgetting in Differentiable Resistor Networks” by the University of Limerick, exhibit catastrophic forgetting, controlled by task conflict and degree of adaptation, providing a physically interpretable testbed for CL.

Under the Hood: Models, Datasets, & Benchmarks

These advancements are underpinned by robust new models, specialized datasets, and comprehensive benchmarks:

• Dynamic and Adaptive Architectures:
  • SAMoE-C (The Hong Kong Polytechnic University): Mixture-of-Experts framework with attention-based semantic router for CSI-based Human Activity Recognition.
  • CRAFT (Iowa State University): Utilizes low-rank interventions on hidden representations of LLMs (e.g., Llama-3.2-1B-Instruct, Gemma-2B-it).
  • NORACL (University of Zurich & ETH Zurich): Employs on-demand neuronal growth triggered by Effective Dimension and Fisher Information, starting from compact networks.
  • ADR-SNN (Chinese Academy of Sciences): Brain-inspired Spiking Neural Networks with adaptive dynamic routing to create sparse neural pathways for energy-efficient CL, tested on CIFAR100 and ImageNet.
  • CoMemNet (Shanghai Jiao Tong University): A dual-branch contrastive learning framework for traffic prediction using an embedding-based backbone and Node-Adaptive Temporal Memory Buffer (TMRB-N).
  • TSN-Affinity (AGH University of Krakow): A Decision Transformer-based approach for Continual Offline Reinforcement Learning using sparse task-specific subnetworks and Affinity Routing.
  • LoDA (Xidian University): Decomposes LoRA’s update space into general and isolated subspaces using projection energy objectives, validated with ViT-B/16.
  • Sentinel-VLA (University of Sydney): A metacognitive Vision-Language-Action model with a dedicated ‘sentinel’ module for active status monitoring, trained on EC-Gen data and used with VLM/Action experts.
• Specialized Datasets & Benchmarks:
  • MEP-BENCH: A 31-task multi-track benchmark for multi-component neuroplastic CL systems (MPCS).
  • WildFC, AIGenImages2026: Evolving datasets for AI-generated image detection, constructed via an automated fact-check retrieval pipeline.
  • MM-Fi dataset: Multi-modal CSI dataset with 27 activities across 4 environments for HAR.
  • Continual LEGO: An expansion of the LEGO compositional reasoning task for Transformer CL capabilities.
  • MMWHS, NCI-ISBI13, I2CVB, PROMISE12, LAScarQS, LiTS, FeTS: Diverse medical imaging datasets for continual medical image segmentation benchmark.
  • CoIN-6, CoIN-Long-10: Benchmarks for federated multimodal continual learning used with LLaVA and Qwen2.5-VL backbones.
  • ISRUC, FACED, PhysioNet-MI: Benchmarks for EEG continual learning (sleep staging, emotion decoding, motor imagery).
  • PEMSD3(S), PEMSD4(L), PEMSD8(M): Large-scale traffic datasets for traffic prediction.
  • TRACE benchmark: Used for evaluating LLM adaptation in CRAFT.
  • AndroZoo, CICMalDroid 2020, CIC-AndMal2017: Datasets for Android malware detection.
• Public Code & Resources:
  • CoMemNet: For continual traffic prediction.
  • TSN-Affinity: For continual offline reinforcement learning.
  • Physical-learning-resistor-networks: For resistor network sequential learning.
  • Fade: For adaptive weight decay in continual learning.
  • WildFC: Project page for automated in-the-wild data collection.
  • Sphere-RL: Project page for SPHERE regularizer in MoE-RL.

Impact & The Road Ahead

The implications of this research are profound. Continual learning is moving beyond simply preventing forgetting, embracing adaptive and dynamic intelligence that can learn, evolve, and even repair itself. From medical AI that adapts to new diseases without re-training on sensitive patient data (“CoRE: Concept-Reasoning Expansion for Continual Brain Lesion Segmentation” by Southeast University), to self-correcting robots that learn on-demand and recover from errors (“Sentinel-VLA: A Metacognitive VLA Model with Active Status Monitoring for Dynamic Reasoning and Error Recovery” by University of Sydney), the future promises truly autonomous and robust AI.

New theoretical insights are also redefining our understanding, such as how representational dimensionality controls modularity’s effectiveness in “When Does Structure Matter in Continual Learning? Dimensionality Controls When Modularity Shapes Representational Geometry” (IT University of Copenhagen), or how external memory in LLM agents merely relocates the CL challenge from parameters to retrieval dynamics (“When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents” by Nanyang Technological University).

The overarching trend points toward AI systems that are not only efficient and robust, as highlighted by the HERCULES framework for NAS (Politecnico di Milano’s “HERCULES: Hardware-Efficient, Robust, Continual Learning Neural Architecture Search”), but also deeply interpretable and adaptive, moving away from static models to dynamic, lifelong learners. The progress is clear: continual learning is critical for building the next generation of intelligent systems capable of thriving in our unpredictable world.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.