Online Learning with LLM Experts from Limited Feedback
The paper treats LLM routing as a budgeted bandit problem, learning which expert to use while asking for feedback only sparingly.
The authors model prompts as feature vectors and LLM experts as the actions chosen over time. They give regret bounds for both full-information and bandit feedback settings under a feedback budget far smaller than the number of rounds. Experiments in the paper report that the approach learns effective routing policies across multiple LLMs with limited reward observations. HF Daily Papers' note
The authors model prompts as feature vectors and LLM experts as the actions chosen over time. They give regret bounds for both full-information and bandit feedback settings under a feedback budget far smaller than the number of rounds. Experiments in the paper report that the approach learns effective routing policies across multiple LLMs with limited reward observations. HF Daily Papers' note
score 4