Researchers have proposed a new way to think about routing user prompts to the most suitable large language model (LLM) expert, framing the task as a contextual bandit problem with limited feedback. The work, posted on Hugging Face, aims to maximize response quality in an online setting where feedback is scarce.
In the formulation, the system faces K actions representing different experts and d features that encode incoming prompts, operating over a horizon of T rounds. The proposed algorithms strategically choose which rewards to observe in order to keep regret low.
The paper reports theoretical guarantees: in the full-information setting, regret is bounded by O(d T / m), while in the bandit setting the bound is O(d T K / m), where m is a budget on feedback. The authors state that experiments show their approach efficiently learns high-quality routing strategies across diverse LLMs from limited feedback.
The work sits alongside a growing body of routing research, including recent papers on workload-budget-constrained routing, cost-aware multi-objective bandits, and top-k Pareto bandits.