Full list on Google Scholar

2026

From Prescribed to Emergent: Representing and Exploiting Structure in LLM Agents Collaboration

H. Yang, Y. Mao, Z. Zhang, Y. Yao, J. Chen, T. Lan, and C. Joe-Wong

Preprint Paper Webpage

Introduces Task-Environment-Agents (TEA), a structural representation that couples an Agent Interaction Graph with a Task Activity Graph and the shared environment. It makes evolving collaboration explicit without prescribing agent reasoning, behavior, or communication order, and Infuse uses TEA records to recover, assess, and improve collaboration policies.

TEA overview: a collaboration policy is executed, recorded as task, environment, and agent views, recovered into a topology and rules, and infused into an updated policy.

Auditing Emergent LLM-Agent Collaboration through Cooperation-Obligation Coupling

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

Preprint Paper

Represents emergent collaboration through coupled cooperation-obligation graphs: a cooperation graph G of actions and artifacts, an obligation graph Q of work states, and an audit map Π between them, with a verifier deciding which artifacts count as evidence. The coupling formalizes work soundness and assignment stability; their joint satisfaction quantifies state quality and yields a conditional performance bound.

iCORE: the cooperation graph G records actions, artifacts and evidence; the obligation graph Q records work moving from open to assigned to accepted; the audit map Pi links artifacts and evidence in G to the obligations they declare and justify.

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

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

Preprint Paper Webpage

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.

Dynamic Interaction Graph representation of agent activations and events.

SHAPE: Scaling Collaborative Planning in LLM-Powered Heterogeneous Multi-Robot Systems

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

CoRL26 LEAP Workshop Paper Webpage

Characterizes what a task offers additional robots through parallel and heterogeneous participation, and studies how collaborative planning scales for teams of 3, 9, and 15 robots under Individual, Centralized, Chain, and SHAPE on eight embodied tasks in BEHAVIOR. SHAPE plans asynchronously over a shared, evolving, agent-built task graph with selective notifications, reaching higher average completion with lower coordination overhead across scales.

SHAPE overview: team composition, task configuration, collaboration patterns, and evaluation by task performance and collaboration overhead.

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

NeurIPS26 E&D Track Paper Webpage

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.

COOP squared connects symbolic agent activity with grounded environment execution.

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

TMLR26 Paper

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.

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

2025

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

NeurIPS25 LAW Workshop Paper Webpage

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.

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

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

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

NeurIPS25 SEA Workshop Paper Webpage

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.

Physical constraints in the CUBE block-pushing environment.

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

AAAI25 MARW Workshop Paper Webpage

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.

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

2024

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

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

IEEE FMSys24 Workshop · Co-located with CPS-IoT Week Paper

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.

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.

2023 and before

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.

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

JAMIA23 Paper

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.

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

KDD23 Paper

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

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

Artificial Organs 2023 Paper

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