RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation

Short summary

RouteRec evaluates whether request-level hard selection or item-level learned aggregation works better for choosing among heterogeneous recommender agents under cost constraints. On MovieLens-1M, hard selection underperforms BM25, but learned aggregation with gated all-agent access reaches HR@10=0.295 with 70.2% LLM calls. The key finding: item-level aggregation is more promising than coarse request-level agent selection.

• •Request-level hard selection of recommender agents underperforms BM25 baseline (0.223 vs 0.254)
• •Item-level learned aggregation with gated LLM escalation reaches HR@10=0.295 at 70.2% LLM call rate
• •Coarse agent selection is too granular for sparse fixed-candidate recommendation settings

Generated with AI, which can make mistakes.