Center for AI Business Innovation

Federated learning (FL), a machine learning framework designed to protect data privacy, has gained significant attention from researchers in recent years. Unlike traditional centralized approaches, FL enables multiple clients to collaboratively train a global model without sharing raw data, thereby preserving privacy and compliance with data governance policies.

Each client independently updates the global model received from the server using its local data and then shares the updated model parameters with the server. The server aggregates these updates to produce a new global model, which is subsequently broadcast to the clients for the next training round. FL has been recognized as an effective method for incentivizing data sharing, which helps address the data availability challenges critical in the era of large language models (LLMs).

By enabling distributed training while respecting privacy, FL fosters collaboration across organizations, institutions and devices, where data cannot be directly pooled due to privacy or regulatory constraints. Despite its promise, FL has several limitations. Its convergence speed and final performance often fall short compared to centralized methods, even when client data is independently and identically distributed (i.i.d.). These challenges are magnified when applying FL to deeper neural networks, which demand more robust optimization techniques.

Another significant issue is data heterogeneity: in many real-world scenarios, client data is non-i.i.d., leading to imbalances that degrade the performance of the global model. Conventional FL methods struggle to handle such variability effectively, resulting in reduced model accuracy and stability. To address these challenges, our center has been actively developing innovative solutions both theoretically and experimentally. We are exploring advanced optimization algorithms to accelerate convergence, strategies to mitigate the impact of non-i.i.d. data, and frameworks tailored for large-scale, deep neural networks in federated settings. These efforts aim to push the boundaries of FL, making it more robust, efficient, and adaptable to diverse real-world applications.

The use of large-language model agents in social simulations has emerged as a trending research topic, driven by advancements in generative AI. Unlike traditional rule-based agent modeling, employing LLMs as agents offers significantly greater flexibility. This approach is also widely regarded as a means of verifying and enhancing LLMs’ ability to perform human-like, deliberate reasoning.

A key question in agent modeling is how accurately these models mirror real-world scenarios. Some studies have demonstrated LLMs' capability to replicate basic human behaviors and reasoning. However, issues such as data contamination and value misalignment may introduce unwanted biases, causing LLMs to become overly familiar with or skewed toward the problems being studied.

These biases can compromise the quality of social simulations, particularly in complex, long-term scenarios where higher-order interactions — such as cooperation, confrontation, deception and persuasion — play a central role. In contrast, our ongoing research emphasizes that in social simulations, agents should operate independently of prior assumptions. Instead, they must focus on contextual factors and actively adapt their actions based on historical interactions, ensuring greater realism and robustness in simulating social dynamics.

Participatory budgeting is a decision-making process in which public resources are allocated based on the preferences of individuals within a community. Traditionally, most studies in this field rely on an additive utility function, where each individual assigns a private utility value to every candidate project, and the total utility of a funded subset is the simple sum of these values.

While this additive model simplifies both analysis and implementation, it often fails to capture the complexities of real-world scenarios. For example, building two playgrounds in the same neighborhood may not yield twice the utility of building a single playground, due to diminishing returns or overlapping benefits. To address these limitations, we broaden the framework of participatory budgeting by introducing general utility functions.