Learning, Fast and Slow: Towards LLMs That Adapt Continually
arXiv·high signal
Rishabh Tiwari et al. present a dual-system framework for continual LLM adaptation, combining parameter updates (slow learning via RL) with in-context learning (fast adaptation). The approach addresses a core limitation of current LLMs — they're trained once and deployed static — by enabling models to incorporate new knowledge without catastrophic forgetting. Directly relevant for agent builders who need models that improve from deployment experience.