The authors treat continual learning for time series forecasting as prompt evolution.
Prompt evolution lets frozen LLMs learn new forecasting regimes without retraining
Outcome-driven prompt evolution makes a frozen LLM beat time-series foundation models on held-out data without any training.
Big Tech
Laurent Mombaerts · Jonathan Taws · Jacopo Gargano · Dinesh Gaddam · Kevin George · Srijan Tiwari · +1 more
Amazon Web Services · University of California
Research Digest··3 min read
The authors replace retraining with a continual learning loop that writes a short text guidance block into a frozen LLM's prompt.
§