continual-adaptation

SOLAR proposes a self-optimizing autonomous agent that leverages parameter-level meta-learning and multi-level reinforcement learning to enable lifelong adaptation of LLMs to non-stationary data streams, outperforming baselines on reasoning tasks.