FairGRPO: Fair Reinforcement Learning for Equitable Clinical Reasoning
Shiqi Dai, Wei Dai, Jiaee Cheong, Paul Pu Liang
TL;DR
FairGRPO is a hierarchical reinforcement learning method designed to promote equitable clinical diagnosis across diverse populations, reducing bias and improving fairness in medical AI systems.
Contribution
The paper introduces FairGRPO, a novel reinforcement learning approach that incorporates adaptive importance weighting and unsupervised clustering to enhance fairness without demographic labels.
Findings
Reduces predictive parity gap by 27.2% compared to baselines.
Improves F1 score by 12.49% across datasets.
Fairness improves progressively during training.
Abstract
Medical artificial intelligence systems have achieved remarkable diagnostic capabilities, yet they consistently exhibit performance disparities across demographic groups, causing real-world harm to underrepresented populations. While recent multimodal reasoning foundation models have advanced clinical diagnosis through integrated analysis of diverse medical data, reasoning trainings via reinforcement learning inherit and often amplify biases present in training datasets dominated by majority populations. We introduce Fairness-aware Group Relative Policy Optimization (FairGRPO), a hierarchical reinforcement learning approach that promotes equitable learning across heterogeneous clinical populations. FairGRPO employs adaptive importance weighting of advantages based on representation, task difficulty, and data source. To address the common issue of missing demographic labels in the…
Peer Reviews
Decision·Submitted to ICLR 2026
• The paper addresses fairness in reinforcement learning for multimodal foundation models, a highly relevant problem in clinical reasoning where demographic disparities can have critical implications. • The presentation is clear and easy to follow, with well-explained formulations and experimental design. • FairGRPO extends Group Relative Policy Optimization with an adaptive importance-weighting mechanism that normalizes rewards via inverse-temperature scaling based on group size and performan
1. The clustering-based grouping used when demographic labels are unavailable is intuitive but not further analyzed. It remains unclear what the clusters capture in practice, or under which conditions they would align with meaningful demographic or clinical subgroups. 2. The proposed FairGRPO objective is reasonable and empirically effective, but a theoretical analysis or discussion of the training objective and convergence behavior would provide a deeper understanding of its effect, rather th
1. This paper addresses fairness in reinforcement learning for multimodal clinical models, a critical but underexplored area. By focusing on fairness in critic-free RL (e.g., GRPO-style optimization), the work bridges the gap of fair ML for healthcare communities. 2. The proposed method of scaling advantages inversely by group representation and task difficulty is simple to compute and adds negligible runtime overhead. 3. Experiments span seven clinical datasets covering diverse imaging types, w
1. The theoretical grounding for the proposed fairness optimization is limited. It is unclear whether the scaling guarantees convergence or prevents overcompensation. 2. The paper employs K-means clustering on reward vectors to infer latent demographic groups, but does not analyze the robustness of this clustering step. The number of clusters, initialization, and metric choice could influence group assignments, potentially introducing instability or even amplifying unintended biases in underrepr
One of the first works to embed group fairness considerations into the RL for clinical reasoning models
Not clear how the group fairness influences individual performance, which is what is most important.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsArtificial Intelligence in Healthcare and Education · Explainable Artificial Intelligence (XAI) · Adversarial Robustness in Machine Learning