HuggingGPT: Orchestrate expert models with an LLM controller¶

Four-stage planning, capability matching, dependency execution and aggregation.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Zhejiang University / Microsoft Research Asia.
First public release	2023-03-30
Author code	microsoft/JARVIS.
Local adapter or method key	`hugginggpt`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

ChatGPT decomposes requests into a dependency graph, selects experts by their Hugging Face descriptions, executes in topological order, and synthesizes outputs.

flowchart LR
  A[Plan dependent subtasks] --> B[Match expert descriptions]
  B[Match expert descriptions] --> C[Execute topologically]
  C[Execute topologically] --> D[Synthesize outputs]

Original paper figure¶

Figure shows task planning and model orchestration. From the original paper; copyright remains with the authors.

Core equation¶

\[ G=(V,E)=\operatorname{Plan}(x),\qquad m_v=\arg\max_m \operatorname{match}(d_v,c_m),\qquad o_v=f_{m_v}(o_{\mathrm{pa}(v)}). \]

Reported results¶

The paper gives language, vision, speech and cross-modal system cases and qualitative evaluation; no production A/B.

Local reproduction¶

PlanBench mini, 120 episodes, seed 42. Tool descriptions proxy model capabilities and each subtask keeps its dependencies. Baseline vs HuggingGPT: joint success 1.0000/1.0000, mean cost 64.5000/2.3500, model matches/edges 0/0 vs 360/240. Metrics. No ChatGPT or hundreds of actual experts; the perfect score measures orchestration only.

auto-research agent-eval --method hugginggpt --benchmark planbench-mini \
  --episodes 120 --memory-size 24 --seed 42