Huanzhi (Hans) Mao
Ph.D. student, Sky Computing Lab, UC Berkeley
Advised by Joseph E. Gonzalez
and Ion Stoica
I work on LLM agents, with a focus on tool use: how to measure what agents can actually do, and what it takes to make them reliable enough to hand real tools.
I lead the Berkeley Function Calling Leaderboard (BFCL), the de facto benchmark for evaluating function calling in language models. It began as an AST-based evaluation of single-turn calls and now spans multi-turn stateful execution and agentic settings — web search, memory management, and format sensitivity. BFCL is developed in the open as part of the Gorilla project (13k+ stars) and ships as the bfcl-eval package.
I spent the summer of 2025 at Salesforce AI Research, where I built BFCL Audio, an evaluation for audio-native function calling. Before the Ph.D. I studied Computer Science (Honors) and Data Science at Berkeley, and served seven semesters as Head TA for CS 61A, Structure and Interpretation of Computer Programs.
News
- Jul 2026 BFCL Audio, our evaluation for audio-native function calling, is at ICML 2026.
- Jul 2026 Characterizing Agents in Production is an oral at ICML 2026.
- Dec 2025 Gave a spotlight talk on MFCL Vision at the NeurIPS 2025 NORA workshop.
- Aug 2025 Released BFCL Audio with Salesforce AI Research.
- Jul 2025 BFCL V4 is out, extending the leaderboard to agentic evaluation: web search, memory management, and format sensitivity.
- Jul 2025 The BFCL paper appeared at ICML 2025.
- Jun 2025 Started a summer internship at Salesforce AI Research.
- Sep 2024 BFCL V3 adds multi-turn, multi-step, state-based evaluation.
- Aug 2024 BFCL V2 Live replaces synthetic prompts with user-contributed real-world data.
Publications Google Scholar →
* Equal contribution unless otherwise noted.
-
EasyPPO: Stabilizing the Critic Is Key
Stabilizes PPO for LLM reinforcement learning through actor-only overlong filtering, noise-normalized critic regression, and smaller critic mini-batches.
-
BFCL Audio: An Audio Function Calling Evaluation for Large Language Models
Extends BFCL to spoken input with 6.2K expert-verified audio tool-use tasks, covering both pipelined ASR-to-LLM and end-to-end audio-in settings under controlled acoustic perturbations.
-
Assertion-Conditioned Compliance: A Provenance-Aware Vulnerability in Multi-Turn Tool-Calling Agents
Measures how multi-turn tool-calling agents respond to misleading assertions from users and from tool outputs. I advised this work and co-led the project.
-
Characterizing Agents in Production
-
FrontierCS: Evolving Challenges for Evolving Intelligence
-
AutoCode: LLMs as Problem Setters for Competitive Programming
-
SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing
-
DeployBench: Benchmarking LLM Agents for Research Artifact Deployment
-
Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live
-
CFLBench: Benchmarking Novel Control Flow Language Learning
A benchmark of 19 small deterministic programming languages that tests whether tool-augmented language models can learn unfamiliar control flow. I advised this work.
-
Concurrency without Model Changes: Future-based Asynchronous Function Calling for LLMs
An execution-layer framework that overlaps model decoding with tool execution by reasoning over symbolic futures, with no fine-tuning and no change to the function-calling protocol.
-
OpenDeepThink: Parallel Reasoning via Bradley-Terry Aggregation
-
FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale
-
The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models
The benchmark I lead. Evaluates LLM function calling via scalable abstract-syntax-tree matching across single-turn, crowdsourced live, multi-turn, and stateful agentic settings.
-
MFCL Vision: Benchmarking Tool Use in Multimodal Large Language Models for Visual Reasoning Tasks
A benchmark of 250 expert-verified tasks across five image domains that measures how reliably multimodal LLMs turn visual input into correct tool calls.
-
The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks
-
Gorilla OpenFunctions v2
An open-source function-calling model covering Python, Java, JavaScript and REST APIs, with parallel and multiple calls plus function-relevance detection.
Projects
-
Berkeley Function Calling Leaderboard (BFCL)
Project leadThe de facto benchmark for evaluating function calling in language models. BFCL has grown from AST-based evaluation of single-turn calls (v1) through user-contributed live data (v2) and multi-turn stateful execution (v3) to agentic evaluation of web search, memory management, and format sensitivity (v4). I run the releases, dataset, evaluation harness, and model onboarding.
-
Gorilla
Top contributorThe Sky Computing Lab's open-source project for teaching and evaluating language models on API and tool calls — the original Gorilla model and APIBench, OpenFunctions, GoEx, RAFT, the API Zoo, and BFCL. 13k+ stars.
Beyond Research
I sail, hike, and play more board games than I can justify. In 2017 I was the national champion in the Chinese Yachting Association Youth Sailing League.
If any of this overlaps with what you are working on, I would be glad to hear from you — email is the fastest way to reach me.