What changed in BFCL's V4?

V4 rebalanced BFCL’s scoring weight, not just its task list. V3 split evenly across Live single-turn calls, Non-Live single-turn calls, and Multi-Turn, 33% each, with no agentic category at all. V4 cuts Live and Non-Live to 10% each, carves out a 10% Irrelevance slice, keeps Multi-Turn at 30%, and hands the other 40% to a new Agentic category. The Gorilla team’s own reason: “As single-turn tasks approach saturation, weighting now favors complex, multi-step agentic tasks.”

Agentic is three graded skills none of V3 had: Web Search, where the model gets a search tool and a fetch tool for questions past its training cutoff; Memory Management, testing key-value, vector, and recursive-summary memory across a session; and Format Sensitivity, a diagnostic that reruns the same calls across 26 prompt and schema variations and finds models return higher accuracy with Python and JSON response formats than with either XML variant. Why a benchmark’s score drops on a new version covers the general pattern; V4 is a case of it done on purpose, with the harder categories named and weighted rather than found by accident.