QuantaAlpha is an evolutionary framework for LLM-driven alpha mining that addresses key challenges in financial factor discovery: noisy backtesting feedback, non-stationary markets, and limited exploration. The framework treats each end-to-end mining run as a trajectory and applies two evolutionary operators—mutation and crossover—to iteratively improve factors. Mutation localizes and revises suboptimal steps via self-reflection, while crossover recombines validated segments from high-reward trajectories, enabling structured exploration and reliable reuse of effective patterns. To ensure robustness, QuantaAlpha enforces semantic consistency across hypotheses, symbolic expressions, and executable code, and applies complexity and redundancy controls to prevent overfitting and factor crowding.

Extensive experiments on the CSI 300 index demonstrate that QuantaAlpha significantly outperforms traditional machine learning models, deep learning models, factor libraries, and prior LLM-based agentic systems (RD-Agent and AlphaAgent). Using GPT-5.2, it achieves an Information Coefficient of 0.1501, an annualized return of 27.75%, and a maximum drawdown of only 7.98%. Ablation studies confirm that each component—diversified planning, mutation, crossover, and generation constraints—contributes non-trivially to overall performance. Moreover, factors mined on CSI 300 transfer effectively to CSI 500 and S&P 500 without any adaptation, yielding cumulative excess returns of 160% and 137% over four years, respectively. This demonstrates strong generalization under market distribution shifts and highlights the framework's potential for real-world quantitative investment.