Hanqing Yang

Hanqing Yang

PhD Student in ECE at CMU

Open Source

Talk is cheap...

I am constantly inspired by open-source projects and grateful for the openness behind them. Many of my projects began with ideas and tools shared by others. I hope something here can become a starting point for someone else.

More is coming.

Python · MIT License

DIG Dynamic Interaction Graph

DIG records emergent multi-agent execution as a time-evolving graph of agent activations and events. It is agnostic to the language model, environment, and number of agents.

Observe
Wrap agents from an existing framework, or let DIG deliver events itself.
Inspect
Trace events, activations, reactions, routing, and lineage as a run unfolds.
Heal
Detect structural failures and record information injection or rerouting interventions.

RepositoryPaperProject page

DIG represents multi-agent execution using agent activations and the events that flow between them.
Agent activations and events form a dynamic, inspectable record of collaboration.

Python · Multi-agent simulation

Multi-Agent Crafter Decentralized generative agents

A Minecraft-inspired environment for studying LLM agents that plan, remember, share resources, and work toward a collective objective.

Simulate
Run configurable teams in a resource-gathering and crafting environment.
Remember
Use persistent memory and knowledge graphs for long-horizon planning.
Observe
Follow individual agents or the full team through a multi-view interface.

RepositoryPaperProject page

Multiple LLM agents collaborate in the Multi-Agent Crafter environment.
Agents coordinate resource gathering and crafting through shared and local memory.

Python · Jupyter notebooks

LLM-Based Digital Twin Human-in-the-loop control

Code for simulating heterogeneous human feedback and training HVAC controllers around collective comfort and energy use.

Simulate
Model a population with heterogeneous thermal preferences.
Control
Run centralized and decentralized controller experiments.
Evaluate
Reproduce analysis, visualization, and confidence-interval workflows.

RepositoryPaper

An LLM-based digital twin simulates population feedback used to train an HVAC controller.
Simulated population dynamics connect human feedback to controller training.