Apple study finds plan-then-retrieve tool discovery wins only on multi-tool LLM agent queries
Overview
Apple researchers Mingfei Zhang, Amit Surana, Ziauddin Majid, Chenxi Wang and Michael Xu compare two ways LLM agents find APIs in large registries.
In plan-then-retrieve (PtR), the model first generates Abstract Tool Stubs, hypothesized function signatures for needed capabilities, and uses them as retrieval queries. In retrieve-then-plan (RtP), tools are retrieved directly from the user query.
On ToolBench's 46,984 endpoints, the authors report that PtR outperforms RtP on multi-tool queries, while RtP outperforms PtR on single-tool queries. The extra discovery call roughly doubles inference cost, so the authors suggest routing queries by complexity.
Written by AI from the articles below · updated Oct 15, 3:55 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Apple Machine Learning ResearchOfficialPlan-then-retrieve tool discovery helps multi-tool LLM agent queries but hurts single-tool ones
AIApple researchers compare two ways LLM agents find APIs in large registries: retrieving tools from the user query versus first generating Abstract Tool Stubs for retrieval. On ToolBench's 46,984 endpoints, the plan-then-retrieve approach beats retrieve-then-plan on multi-tool queries but loses on single-tool ones. The extra discovery call roughly doubles inference cost, which is why the authors suggest routing queries by complexity.
Heat trend
Not enough continuous observations to show a trend yet.