IncrLearn — Incremental and Continual Learning
in the Age of Large-Scale AI

IncrLearn addresses the full spectrum of methods for learning from time-varying, streaming, and large-scale dynamic data: incremental classification and clustering, concept drift management, novelty detection, active and continual learning. The rapid rise of large language models (LLMs) and foundation models has brought these challenges to the center of AI research: catastrophic forgetting, knowledge editing (ROME, MEMIT), continual pre-training, parameter-efficient adaptation (LoRA, adapters), streaming RAG with dynamic knowledge bases, and concept drift in LLM-based pipelines are now among the most discussed open problems in the field — and they are structurally identical to the problems this workshop has always addressed.

IncrLearn 2026 places dedicated emphasis on this convergence, welcoming both the classical incremental learning community and researchers working on continual learning for LLMs and foundation models.

Core Constraints of Incremental Learning

All incremental learning algorithms — whether applied to data streams or to large-scale models — share four fundamental constraints:

1. Distribution shifts must be detected and handled;
2. The system must operate without access to all future data;
3. New knowledge must be integrated without intensive reprocessing of past data (the root of catastrophic forgetting);
4. Updated behaviors must be available promptly without degrading prior knowledge.

Topics of Interest