AlphaAgent addresses the critical problem of alpha decay in quantitative finance by introducing a novel LLM-driven framework with systematic regularization. The framework employs three key mechanisms: originality enforcement via abstract syntax tree (AST) similarity against existing alpha libraries, hypothesis-factor alignment using LLM-evaluated semantic consistency, and complexity control through AST-based structural constraints. These mechanisms guide the generation of alpha factors that balance financial rationale, novelty, and adaptability to evolving market conditions, thereby mitigating the risk of alpha decay.

The system operates through a closed-loop multi-agent architecture comprising an idea agent (generates market hypotheses), a factor agent (constructs AST-based factor expressions using an operator library), and an eval agent (backtests and provides feedback). This iterative process continuously refines factors to capture emerging market inefficiencies rather than overcrowded signals. Extensive experiments on Chinese CSI 500 and U.S. S&P 500 markets from 2021 to 2024 demonstrate that AlphaAgent significantly outperforms traditional genetic programming, reinforcement learning, and existing LLM-based methods in both performance and resistance to alpha decay, achieving annual excess returns of 11.0% and 8.74% respectively, with an 81% higher hit ratio and 30% fewer tokens consumed.