A team from Northwestern University, led by Ph.D. student Zhenyu Pan, has made significant advances in redefining the balance between model performance and fairness in AI systems, moving beyond traditional federated learning approaches. The team’s work, presented in two complementary studies, leverages reinforcement learning (RL) to enable models to autonomously explore and optimize across multiple objectives—such as accuracy, fairness, and security—without relying on rigid, one-size-fits-all training strategies. In the first study, called FairReason, the team introduced a novel single-model framework where RL acts as a dynamic exploration engine. Instead of jointly training for both reasoning performance and bias mitigation in a single objective, the model uses a policy layer to experiment with different data mixtures and training strategies. This allows the system to explore the trade-off between performance and fairness in a flexible, adaptive way—without being forced to exactly match a teacher model or ground-truth labels. The result is a more robust and balanced output distribution, even when the model is not explicitly trained to reduce bias. In the second study, Evo-MARL, the team applied multi-agent reinforcement learning in a red-team-blue-team environment. Here, RL directly optimizes a joint objective: task correctness and system security. To ensure stability, the framework incorporates KL regularization, while a co-evolving attack pool continuously generates novel adversarial inputs, creating a dynamic and challenging training environment. This enables agents to learn not only how to perform tasks effectively but also how to defend against evolving threats—adapting to distribution shifts and maintaining resilience over time. Together, these two approaches highlight RL’s role as a versatile “explorer” in complex AI optimization landscapes. In FairReason, it explores the optimal balance between performance and fairness across varying data compositions. In Evo-MARL, it navigates the evolving tension between functionality and safety in adversarial settings. The research was motivated by growing concerns in the AI community: while RL-based post-training has proven effective in boosting model reasoning, it may also inadvertently amplify biases or weaken safety. A preliminary literature review revealed a lack of systematic analysis on how different post-training methods—such as supervised fine-tuning (SFT), knowledge distillation (KD), and RL—impact the trade-off between intelligence and fairness. There was also no clear, quantitative guidance for developers aiming to build models that are both capable and responsible under resource constraints. To address this gap, the team structured their work around two complementary paths—FairReason and Evo-MARL—aligned with the upcoming Trustworthy Foundation Models Workshop at ICCV 2025. With support from NVIDIA’s GPU cloud computing platform, they conducted large-scale experiments and validated their approaches in dynamic, adversarial multi-agent environments. The project was executed under intense time pressure. Pan led the effort alongside two interns: Zhang Yutong, a high school sophomore from Fudan High School, and Zhang Yiting, a senior at South China University of Technology. Despite their differing academic backgrounds and experience levels, both demonstrated exceptional technical ability and dedication. The team completed the full cycle—from experimental design and code implementation to result analysis and paper writing—in under two weeks, working through late nights and constant iteration. The process was fast-paced and demanding, but the team’s strong collaboration and shared focus allowed them to deliver high-quality results. Looking ahead, the team plans to expand FairReason by testing a wider range of model sizes and data types, aiming to uncover generalizable patterns—potentially leading to a “scaling law” for fair and effective model training. In Evo-MARL, they aim to evolve the current system into a more flexible, heterogeneous multi-agent framework, where agents with different roles, capabilities, and objectives can coexist and interact, better simulating real-world complexity. The work, available on arXiv (https://arxiv.org/abs/2508.03864v1 and https://arxiv.org/abs/2507.23067), represents a step forward in building AI systems that are not only smart but also fair, safe, and adaptable. Pan’s personal website can be found at https://pzyseere.github.io/.