Validate complexity-adaptive routing classifiers

Validate empirical classifiers for routing tool-discovery queries between Retrieve-then-Plan and Plan-then-Retrieve architectures, including the proposed zero-cost signal based on Retrieve-then-Plan’s top-1 cosine similarity, to determine whether they reliably identify single-tool versus multi-tool query complexity.

Background

The paper finds a complexity-dependent crossover between Retrieve-then-Plan (RtP) and Plan-then-Retrieve (PtR): RtP performs better on single-tool queries, whereas PtR performs better on multi-tool queries. This motivates routing each query to the architecture expected to perform best.

The authors propose several pre-LLM routing signals, including query length, clause count, named-entity density, conjunctive sub-goals, and especially RtP’s top-1 cosine similarity. However, these signals are not empirically validated as classifiers in the study, leaving unresolved whether they can support reliable complexity-adaptive routing without introducing additional model calls.

References

We leave empirical validation of such classifiers to future work; complexity-adaptive routing over retrieval strategies has established precedent in Adaptive-RAG~\citep{jeong2024adaptiverag}, which routes between no-retrieval, single-step, and iterative retrieval by query complexity.

— Plan or Retrieve First? A Scale Study of Tool Discovery Architectures  (2610.19068 - Zhang et al., 13 Oct 2026) in Section 3, subsection “Discovery as the Binding Constraint” (following the discussion of complexity-adaptive routing)