AlgoEvolve applies LLM-driven program evolution to algorithmic trading

A new cs.AI paper, “AlgoEvolve,” explores whether large language models can act as semantic mutation operators for evolving algorithmic trading programs.

The domain is harder than static coding benchmarks because trading strategies must deal with noisy, changing environments and evaluation that can be sensitive to overfitting. That makes it a useful stress test for LLM-guided program search.

The paper adds to a growing research thread where models are not just writing code once, but repeatedly generating, mutating, and selecting programs inside an automated discovery loop.