The rapid progress of Large Language Models (LLMs) has produced systems with remarkable capabilities in language understanding, reasoning, programming, and tool use, while advances in post-training, reinforcement learning, retrieval-augmented generation, and agentic scaffolding now let agents tackle complex tasks in coding, web navigation, scientific discovery, and embodied control. Yet today's frontier models remain largely static: after costly pre- and post-training, their knowledge, skills, and reasoning are mostly fixed, and once deployed they adapt through prompting, retrieval, memory, or external tools rather than genuine internal learning and consolidation.
This contrasts sharply with human intelligence, where people continuously acquire knowledge, refine representations, and reorganize beliefs through interaction. Current LLMs and embodied agents instead cannot continually learn at test time: they struggle to internalize new information after deployment, fail to improve from repeated mistakes, and can forget prior skills when updated naively (catastrophic forgetting).
The NeurIPS 2026 Workshop on Towards Test-Time Continual Learning Agents (TTCL) brings together researchers across continual learning, large language models, reinforcement learning, embodied AI, memory systems, cognitive science, robotics, and multimodal learning. We define Test-Time Continual Learning Agents as AI systems that continuously acquire, consolidate, and refine knowledge and capabilities during deployment, without catastrophic forgetting or repeated large-scale retraining. This goes beyond updating facts: it asks how agents improve perception, reasoning, planning, exploration, skill acquisition, and long-term decision-making through ongoing experience in virtual and physical worlds—capabilities especially important for robotics, scientific discovery, personalized assistants, education, healthcare, and human–AI collaboration.
Core Questions
- What does it mean for an agent to truly learn at test time, beyond retrieval and prompting?
- What memory, consolidation, exploration, and self-reflection mechanisms enable scalable test-time continual learning without catastrophic forgetting?
- How do we evaluate such learning and ensure its safety and robustness across long horizons and open-ended, multimodal settings?
Goals
- Convene experts across ML, continual learning, NLP, RL, robotics, cognitive science, multimodal learning, and AI safety.
- Solicit work spanning lifelong and online learning, test-time adaptation, meta-learning, intrinsic motivation, cognitive architectures, and self-improving agents.
- Promote rigorous benchmarks and evaluation protocols for long-horizon agent learning, including a hands-on community challenge.
- Bridge the siloed LLM-agent, embodied-agent, continual-learning, and memory communities around shared terminology and open problems.
