Det-MASO: Diversity-Regularized Multi-Agent Sequence Optimization for Factor Generation

To address the limitations of existing single-agent factor generation methods in exploring the factor space and the tendency of multi-agent methods to produce homogeneous factors, this study proposes the Det-MASO multi-agent sequence optimization framework. The framework introduces a differentiable log-determinant penalty on the inter-agent correlation matrix in the joint objective, with computational cost growing linearly with the number of assets, which suppresses agent homogenization during training. Backtesting shows that it improves the information coefficient by 1.1%, annualized return by 6.4%, and Sharpe ratio by 0.33.