Plan or Retrieve First? A Scale Study of Tool Discovery Architectures

Apple Research Blog
Read the full article
Researchers compared two methods for large language models to discover relevant APIs from large tool registries: Retrieve-then-Plan (RtP) and Plan-then-Retrieve (PtR). RtP works better for single-tool queries, while PtR excels on complex multi-tool queries, with trade-offs in inference cost and retrieval accuracy.

More in Machine Learning

"Mommy and daddy" play nice: Apple and Meta team up to bring open-source Muse to Macs

Covered by 6 sources

Meta and Apple, Longtime Foes, Team Up to Put Open-Weight AI on Macs

Covered by 4 sources
Machine Learning3 min read

Nvidia’s new Nemotron 3.5 Super VL is another step forward for US open-weight AI

Covered by 4 sources