In-Context RL Under Non-Stationarity: New Survey
arXiv survey maps in-context reinforcement learning (ICRL) — where pretrained transformers infer task rules and adapt from interaction context alone, no weight updates at test time.
Covers decision-pretrained models, algorithm distillation, long-context meta-RL, and retrieval-augmented agents; organizes around how trial-and-error evidence, rewards, demonstrations, and retrieved experience enable learning-like computation inside the context window.