Large Language Models (LLMs) have achieved remarkable performance across a wide range of tasks, yet they suffer from a fundamental limitation: once trained, their knowledge remains largely static. Updating or correcting this knowledge typically requires expensive and impractical retraining procedures, making current models ill-suited for dynamic, real-world environments.

This special session focuses on how to enable LLMs to evolve after deployment, allowing them to adapt to new information, revise outdated or incorrect facts, selectively forget learned content, and remain reliable over time. The session brings together researchers working on continual learning, knowledge editing, machine unlearning, and parameter-efficient adaptation techniques for LLMs.

As real-world information changes, static language models increasingly produce outdated or incorrect outputs. While Retrieval-Augmented Generation (RAG) methods can partially mitigate this issue by injecting external knowledge at inference time, they do not fully address outdated internal representations or scenarios where retrieval is impractical.

The inability to efficiently update, adapt, or forget knowledge raises significant practical, ethical, and safety concerns, especially as LLMs are deployed in high-impact and safety-critical applications. This special session aims to explore methods and evaluation frameworks that allow LLMs to evolve continuously, without full retraining and without catastrophic forgetting.

Recent advances in continual and incremental pre-training, targeted knowledge editing, and machine unlearning demonstrate promising directions toward adaptive language models. By bringing these communities together, the session aims to foster cross-fertilization and accelerate progress toward reliable, adaptive, and accountable language technologies.

We invite original research contributions on (but not limited to) the following topics:

• Online and continual learning methods for LLMs

• Knowledge editing and targeted model updates

• Machine unlearning and selective forgetting

• Parameter-efficient adaptation strategies, including test-time and inference-time methods

• Evaluation protocols and benchmarks for adaptability, retention, and forgetting

• Automated knowledge updating in Retrieval-Augmented Generation (RAG) systems, including ingestion and incremental updates

• Paper Submission Deadline: March 15, 2026

• Notification of Acceptance: May 15, 2026

• Camera-Ready Submission: June 15, 2026

• Author Registration Deadline: June 30, 2026

• Conference Dates: September 21–23, 2026

• Papers must be submitted according to the IEEE EAIS 2026 submission guidelines.

• For formatting instructions and submission details, please refer to the main conference website

• All submissions will be peer-reviewed and evaluated based on originality, technical quality, relevance to the special session, and clarity of presentation.