AI-Wiki

UC Berkeley

2 min read

UC Berkeley

Confidence 0.75 · 2 sources · last confirmed 2026-09-04

Public research university; in this wiki it appears as the institutional home of the Berkeley Function Calling Leaderboard — the benchmark that has become the de-facto standard for measuring an LLM’s tool-calling ability, and therefore the instrument behind most “small model matches frontier model on tool use” claims in the corpus.

BFCL comes out of the Gorilla project in Berkeley’s Sky Computing Lab, led by Shishir G. Patil with Ion Stoica and Joseph E. Gonzalez as senior authors — the same lab lineage behind Spark, Ray and vLLM. The leaderboard is hosted at gorilla.cs.berkeley.edu/leaderboard.html and the evaluation code is open on GitHub.

What makes Berkeley’s role structurally important to this wiki is neutrality: BFCL is maintained by an academic group rather than a model vendor, which is why vendor papers (Salesforce) and vendor position papers (NVIDIA) alike cite it as the arbiter, and why a practitioner picking a locally-runnable model (Sokolenko) can use it as a shopping list.

The lab’s two contributions to this corpus reach the same diagnosis by opposite methods, which is why it is worth tracking Berkeley as an institution rather than only as a leaderboard host: MAST hand-annotates multi-agent traces and locates failure in coordination, state and long-horizon reasoning; BFCL measures single models and finds single-turn calling saturated while memory sits at 12/100. Same lab, same fault line, one qualitative and one quantitative. See multi-agent-failure-modes.

Appears in this wiki via

Referenced (not authored) by

Relationships

Part of

Graph View

Backlinks