Hanqing Yang

Hanqing Yang

PhD Student in ECE at CMU

Collaboration among intelligent entities

From exploring what constitutes collaboration, to representing emergent structure, to examining what becomes computationally possible.

I approach collaboration through how it unfolds: agents interact, tasks evolve, interdependencies emerge, and individual contributions become collectively consequential. These questions are agnostic to entity type; my current work studies how they play out among LLM agents.

Google Scholar

Research directions

Definition

Interdependence · collective capability · collaborative primitives

What constitutes collaboration among intelligent entities, and which primitives distinguish it from mere group behavior?

Representation

Expressiveness · generalizability · scalability

What representational structure can capture emergent collaboration across tasks, teams, and scales—and under what assumptions?

Computation

Formal properties · evaluation · optimization

What can be derived, evaluated, and optimized once collaboration is represented, and at what computational cost?

Definition, representation, and computation

Definition

Interdependence, collective capability, and collaborative primitives

What constitutes collaboration among intelligent entities, and which primitives distinguish it from mere group behavior?

Definition begins with candidate primitives rather than a settled checklist. The LLM-based digital twin placed heterogeneous human feedback into a shared reward objective. Decentralized generative agents organized shared and local memory for planning under dependency constraints. CUBE made collective capability a property of the task: some goals are individually infeasible. COOP2 then formalized what cooperative tasks require through four categories of constraints that guard progress. The Five Ws survey situated these questions across communication structures in MARL, emergent language, and LLM-based multi-agent systems.

Across these systems, interdependencies emerge as agents interact and share, while heterogeneous capabilities can yield collective capabilities for individually infeasible tasks or efficiency gains. These are candidate collaborative primitives, not a finished definition; which others matter remains open.

Representation

Expressiveness, generalizability, and scalability

What representational structure can capture emergent collaboration across tasks, teams, and scales—and under what assumptions?

DIG represents agent interaction as event passing between activations in a bipartite graph, with minimal assumptions about organization or execution. COOP2 represents multi-agent interaction with the environment as an interplay between symbolic reasoning and grounded transitions. DR. WELL makes a related symbolic-grounded structure explicit through roles, plans, and a shared world model that changes as agents act.

At a more abstract level, iCORE couples a graph of observed cooperation with a graph of evolving obligations and the evidence connecting them. Each step exposes a different tradeoff: expressiveness asks how much unconstrained, adaptive collaboration is captured; generalizability, what transfers versus must be redefined; and scalability, how complexity grows with system size. Their balance—and the assumptions each representation requires—remains open.

Computation

Formal properties, evaluation, and optimization

What can be derived, evaluated, and optimized once collaboration is represented, and at what computational cost?

A representation becomes useful when it supports questions that final task success cannot answer. DIG makes reachability and progress failures visible in an event-activation graph, then supports online healing through information injection and rerouting. COOP2 aligns agent dynamics with grounded task requirements to observe task progress, coordination cost, and constraint deficits; its repair case study uses predicted deficits to open targeted communication channels for replanning. iCORE formalizes work soundness and assignment stability in coupled cooperation-obligation graphs; their joint satisfaction quantifies state quality and yields a conditional performance bound.

Together, these projects sharpen three roles for computation. Formal properties characterize feasibility, optimality, guarantees, and complexity. Evaluation quantifies collaborative states and links contributions to collective outcomes. Optimization improves outcomes at both agent and collective levels. The open question is which claims and interventions remain tractable as tasks, teams, and representations grow. Failures of evaluation or guarantees also expose what a representation is missing and feed back into the definition of collaboration.

Research works

2026 · Preprint, under review

Auditing Emergent LLM-Agent Collaboration through Cooperation-Obligation Coupling

Z. Zhang*, H. Yang*, C. Joe-Wong, and T. Lan

Represents emergent collaboration through coupled cooperation-obligation graphs that formalize work soundness and assignment stability. Their joint satisfaction quantifies state quality and yields a conditional performance bound. [Paper]

G interaction · Q obligation · Π evidence-backed coupling

2026 · Preprint, under review

COOP2: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems

H. Yang, N. Nourzad, S. Chen, M. Siew, J. Chen, and C. Joe-Wong

Formalizes cooperation from two connected sides: what cooperative tasks require and how LLM agents reason, communicate, use tools, and act across cognitive and primitive layers. Aligning both makes cooperation observable, while the COOP2-Repair case study highlights how the formulation can support targeted replanning. [Paper] [Project page]

COOP squared connects symbolic agent activity with grounded environment execution.

2026 · Preprint, under review

DIG to Heal: Scaling General-Purpose Agent Collaboration via Explainable Dynamic Decision Paths

H. Yang, H. Lee, Y. Yao, Z. Liu, K. Liu, J. Chen, and C. Joe-Wong

Represents asynchronous multi-agent execution as a time-evolving graph of agent activations and events. The graph exposes reachability and progress failures and supports online healing through information injection and rerouting. [Paper] [Project page]

Dynamic Interaction Graph representation of agent activations and events.

2026 · TMLR · Survey Certification

The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why: A Survey from MARL to Emergent Language and LLMs

J. Chen, H. Yang, Z. Liu, and C. Joe-Wong

Surveys communication structures across MARL, emergent language, and LLM-based multi-agent systems through who talks to whom, when, what, and why. It reveals how organizational assumptions are often built into communication mechanisms rather than collaboration examined directly. [Paper]

Figure 11 from the Five Ws survey: an overview of LLM-agent components, multi-agent communication designs, applications, and challenges.

2025 · NeurIPS LAW Workshop

DR. WELL: Dynamic Reasoning and Learning with Symbolic World Model for Embodied LLM-Based Multi-Agent Collaboration

N. Nourzad*, H. Yang*, S. Chen, and C. Joe-Wong

Coordinates embodied agents through two phases: agents negotiate and commit to roles, then independently execute symbolic plans grounded in a shared world model. Working above raw trajectories makes collaboration more reusable and interpretable while allowing the world model to improve across episodes. [Paper] [Project page]

DR. WELL overview: agents negotiate roles, execute symbolic plans, revise a shared world model, and validate plans in the environment.

2025 · NeurIPS SEA Workshop

CUBE: Collaborative Multi-Agent Block-Pushing Environment for Collective Planning with LLM Agents

H. Yang*, N. Nourzad*, S. Chen, and C. Joe-Wong

Designs weighted blocks, force, congestion, collisions, and timing so task-induced dependencies and collective capability become explicit. Some goals are individually infeasible, allowing collaboration to be studied as a property of the task rather than only of the policy. [Paper] [Project page]

Physical constraints in the CUBE block-pushing environment.

2025 · AAAI MARW Workshop

LLM-Powered Decentralized Generative Agents with Adaptive Hierarchical Knowledge Graph for Cooperative Planning

H. Yang, J. Chen, M. Siew, T. Lorido-Botran, and C. Joe-Wong

First Place (tie), Decentralized and Multi-Agents Track · Berkeley RDI LLM Agents MOOC Hackathon

Introduces DAMCS, whose adaptive hierarchical knowledge graph organizes multimodal experience into shared and local memory. Decentralized agents can reason from past interactions while sharing relevant knowledge rather than entire histories during long-horizon planning under dependency constraints. [Paper] [Project page]

Agents in Multi-Agent Crafter cooperate through structured memory and communication.

2024 · IEEE International Workshop on Foundation Models (FMSys) · Co-located with CPS-IoT Week

An LLM-Based Digital Twin for Optimizing Human-in-the-Loop Systems

H. Yang, M. Siew, and C. Joe-Wong

Uses an LLM-based digital twin to simulate heterogeneous human feedback for adaptive HVAC control. Aggregated occupant preferences enter the reward objective, making collective preference part of what the controller learns to optimize. [Paper]

Figure 1 from the digital-twin paper: simulated population dynamics and aggregated thermal preferences train an agent-in-the-loop controller for comfort and energy savings.

Earlier Work

Machine Learning for Healthcare

Before focusing on multi-agent systems, I worked closely with clinicians on decision support and scarce-resource allocation across international, national, and local datasets. These projects moved from validating risk scores, to aligning predictions with clinical decision horizons, to estimating treatment effects when a potentially beneficial intervention cannot be given to every eligible patient.

2023 · Journal of the American Medical Informatics Association

Multi-Horizon Predictive Models for Guiding Extracorporeal Resource Allocation in Critically Ill COVID-19 Patients

B. Xue, N. Shah, H. Yang, T. Kannampallil, P. R. O. Payne, C. Lu, and A. S. Said

Developed calibrated predictions at multiple time horizons for critically ill COVID-19 patients, aligning risk estimates with when scarce ECMO allocation decisions must be made. [Paper]

2023 · ACM SIGKDD

Assisting Clinical Decisions for Scarcely Available Treatment via Disentangled Latent Representation

B. Xue, A. S. Said, Z. Xu, H. Liu, N. Shah, H. Yang, P. R. O. Payne, and C. Lu

Used disentangled latent representations to separate prognostic factors from heterogeneous treatment effects, supporting decisions when treatment capacity is limited. [Paper]

2023 · Artificial Organs

Validation of ECMO Mortality Prediction and Severity-of-Illness Scores in an International COVID-19 Cohort

N. Shah, B. Xue, Z. Xu, H. Yang, E. Marwali, H. Dalton, P. P. R. Payne, C. Lu, A. S. Said, and the ISARIC Clinical Characterisation Group

Evaluated mortality prediction and severity-of-illness scores in an international COVID-19 cohort, testing how established tools behave across a diverse clinical population. [Paper]