A new research paper from KAIST introduces EvolveTrade, a framework that lets large language model trading agents evolve their own behavior over time — not by fine-tuning the model, but by revising the text instructions that govern how it uses tools and makes decisions.
The key insight is treating the system prompt as a text-parameterized policy. After each trading interval, a dedicated Policy Agent reviews accumulated decision traces and realized portfolio performance, then rewrites the policy for the next batch of decisions. The backbone LLM stays fixed throughout.
This approach addresses a fundamental limitation of current LLM trading agents: their tool-use policies are typically hand-written and frozen before deployment, leaving them unable to adapt when market conditions shift. EvolveTrade enables continuous adaptation without the cost or risk of model fine-tuning.
Experiments across multiple market regimes and two different LLM backbones show the self-evolved policies consistently improve Sharpe Ratio and Cumulative Return compared to fixed-policy baselines. Behavioral analysis reveals the evolved policies increase code-mediated analysis and activate regime-relevant computations — the agent learns to write more analytical code when markets become volatile.
The researchers also provide case-level attribution, tracing how specific policy changes led to measurable differences in portfolio returns. This interpretability is notable in a field where LLM agents are often treated as black boxes.