Skip to content
MAGICS Lab

Highlights

Key papers.

A small curated set of the lab's papers, each with a one-line hook and a per-paper page summarizing the idea and linking out.

A reinforcement-learning manager observes aggregated outcomes from a network of LLM agents and, each step, selects an information-disclosure policy that modulates what each agent sees in its prompt. The agents' interaction graph is fixed; only the information layer is reshaped.

ASME Journal of Mechanical Design, 148(4): 041708 · 2026

Adaptive Information Modulation: Designing Governance Mechanisms for Multi-Agent Artificial Intelligence Systems

Govern a multi-agent LLM system by modulating what each agent sees — not who it talks to.

Qiliang Chen, Sepehr Ilami, Nunzio Lorè, Babak Heydari

Three-layer model diagram: inventor types layer (heterogeneity in search distance and adoption approach), organization-design layer (core-periphery networks for information sharing), and performance-landscape layer (a rugged NK-style landscape), with arrows describing how types are allocated to network positions and how agents search and feed back performance scores.

Strategic Management Journal · 2026

Core or Periphery: Examining Where to Allocate Heterogeneous Inventors and the Impact on Firms' Innovation

Where heterogeneous inventors sit in a firm's network changes what the firm can discover.

Babak Heydari, Shinjinee Chattopadhyay, Soumyakant Padhee, Samina Karim

Three-panel summary: (1) study goal — what shapes LLM strategic behavior more, game structure or contextual framing; (2) experimental setup — four games × five framings × GPT-3.5, GPT-4, LLaMa-2 producing cooperate-or-defect choices; (3) key findings — GPT-3.5 is context-driven, GPT-4 is structure-centric and bimodal, LLaMa-2 balances both.

Nature Scientific Reports · 2024

Strategic Behavior of Large Language Models and the Role of Game Structure versus Contextual Framing

Do LLMs reason strategically, or just react to framing? Both — and the mix depends on the model.

Nunzio Lorè, Babak Heydari

Variational autoencoder applied to network topology — original graph is encoded into a Gaussian latent space, then decoded into a reconstructed graph.

ASME Journal of Mechanical Design · 2025

Resource Governance in Networked Systems via Integrated Variational Autoencoders and Reinforcement Learning

Govern a multi-agent network by acting in a learned latent space, not on the raw graph.

Qiliang Chen, Babak Heydari

Hierarchical Graph Reinforcement Learning system diagram: HGRL framework observes and intervenes in an evolving network of autonomous agents that update their types by imitation.

ASME Journal of Computing and Information Science in Engineering · 2025

Adaptive Network Intervention for Complex Systems: A Hierarchical Graph Reinforcement Learning Approach

Hierarchical RL that steers an evolving multi-agent network toward cooperation.

Qiliang Chen, Babak Heydari

Nature Computational Science · 2022

Micro-level Social Structures and the Success of COVID-19 National Policies

Two numbers per country — household size and daily contacts — explain a lot of the pandemic's first wave.

Qingtao Cao, Babak Heydari

Reliability Engineering & System Safety · 2024

Building Back Better: Modeling Decentralized Recovery in Sociotechnical Systems Using Strategic Network Dynamics

After a disruption, agents at the bottom re-form the network themselves. Modeling that matters.

Negin Maddah, Babak Heydari