Value-Weighted Routing Optimizes Recommender LLM Calls.

When designing recommender systems that combine inexpensive heuristics with costly large language models (LLMs), routing decisions are typically based on item difficulty—sending "hard" cases to the LLM. This paper argues that such a framing is incomplete, as it fails to account for the distinct business value associated with different items. A difficult, low-value item does not carry the same cost of error as a difficult, high-value item. The "Value Router" study, conducted through a fully synthetic retail merchandising simulation, demonstrates the benefits of value-weighted routing. By considering both estimated difficulty and estimated value, the Value Router achieved comparable recall of high-value items (60%) to a difficulty-only baseline but with substantially higher precision (98.3% vs. 94.3%). The research also highlighted how aggregate metrics can obscure underlying issues, revealing that overall results were often driven by category differences rather than per-item discrimination. Furthermore, in a simulated Black Friday demand surge, value-weighted routing proved more robust than static or seasonally tuned approaches, underscoring its importance for dynamic retail environments.