NeurIPS 2026 Workshop on
Continual Learning in the Era of
Foundation Models and Embodied Agents

News

About

This workshop focuses on continual learning as a shared challenge for foundation models and embodied agents in dynamic, open-ended, and interactive environments. Continual learning is an inherently broad field — spanning its theoretical foundations, algorithms, and systems all the way to large-scale applications — and its intersection with foundation models and agents dramatically expands this scope. We intentionally bring these areas together because they are becoming increasingly intertwined in modern AI: foundation models are becoming the backbone of agentic and embodied systems, while open-ended, non-stationary environments push both to adapt, self-improve, and acquire new skills over their lifetime.

This intersection gives rise to common challenges, including catastrophic forgetting, memory and knowledge consolidation, online adaptation, long-term skill acquisition, and safety during updates. It also connects to a fast-moving frontier of self-improving agents — agent recursive self-improvement, self-evolution, agent reinforcement learning, and the loop and harness engineering that lets agentic systems keep getting better through repeated interaction, feedback, and self-refinement. The workshop centers on continual learning problems that arise within and across all of these settings.

The workshop will bring together researchers from continual learning, foundation-model adaptation, LLM-agent self-evolution and reinforcement learning, embodied intelligence, and robot learning to address shared challenges and identify new opportunities for cross-pollination — from theory to real-world deployment.

Call for Papers

We invite submissions in two tracks: regular papers (up to 8 pages excluding references) and short papers (up to 4 pages excluding references). Contributions may take the form of research papers, position or perspective pieces, benchmarks and datasets, systems and applications papers, negative or reproducibility results, or other interdisciplinary work spanning continual, lifelong, and online learning across foundation models, LLM agents, and embodied systems. We interpret the scope broadly — from the theory and algorithms of continual learning to large-scale empirical studies, agentic self-improvement systems, and real-world deployments. All submissions are reviewed double-blind, are non-archival, and will be managed through OpenReview.

Topics of interest include but are not limited to:

• Continual and lifelong pre-training, post-training, and alignment of foundation models (LLMs, VLMs, and multimodal models)
• Continual learning, self-evolution, and long-term skill acquisition of LLM and foundation-model agents
• Agent recursive self-improvement and open-ended self-evolution of agentic systems
• Agent reinforcement learning for continual and lifelong adaptation
• Loop engineering and harness engineering for continually improving agents (scaffolding, tool use, feedback and self-refinement loops, and evaluation harnesses)
• Continual and lifelong learning for embodied agents, robotics, control, and world models
• Memory architectures, knowledge consolidation, model editing, and retrieval-augmented adaptation
• Online, test-time, and streaming adaptation under distribution shift and non-stationarity
• Catastrophic forgetting, stability–plasticity trade-offs, and forward/backward transfer
• Theory and foundations of continual learning: generalization, optimization, and scaling laws
• Safety, robustness, privacy, and reliability during continual updates and self-improvement
• Benchmarks, evaluation protocols, and metrics for lifelong and continually learning systems
• Applications and deployment: scientific discovery, healthcare, autonomous systems, and personalization

Each submission will receive at least three reviews. Submissions must not re-present finalized work previously published at ML venues.

All accepted papers will be presented in the poster sessions, and selected papers will additionally be featured as contributed oral talks. We also plan to offer a Best Paper Award.

Submission link: Submit on OpenReview

Important Dates

Schedule

Tentative one-day schedule. Morning session focuses on continual learning for foundation models; afternoon session focuses on embodied agents and robotics.

Invited Speakers

Irina Rish

Université de Montréal / Mila

Chelsea Finn

Stanford University

Abhishek Gupta

University of Washington

Gido M. van de Ven

University of Groningen

Jeff Clune

University of British Columbia / Vector Institute

Kianté Brantley

Harvard University

Organizers

Jonghyun Choi

Seoul National University

Xiao-Ming Wu

The Hong Kong Polytechnic University

Rahaf Aljundi

Toyota Motor Europe

Jorge Mendez-Mendez

Stony Brook University

Liyuan Wang

Tsinghua University

Yujie Feng

The Hong Kong Polytechnic University

Yuankai Luo

Nanjing University