Large language model trading agents can blend market data, news, and executable analysis, but their behavior is typically governed by static, hand-written tool-use policies fixed before deployment. According to a new paper posted on Hugging Face, that rigidity limits how well such agents adapt their evidence gathering, tool invocation, signal verification, and risk management as market regimes shift.
The paper introduces EvolveTrade, a self-evolving framework that treats the system prompt of a tool-using trading agent as a text-parameterized policy. After each update interval, a Policy Agent revises that policy using accumulated decision traces and realized portfolio feedback, while the backbone LLM itself remains frozen. The revised policy then guides the next batch of trading decisions, allowing the agent to progressively refine its information-acquisition and portfolio-construction procedures.
Experiments spanning multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return relative to fixed-policy LLM baselines, achieving better results in most evaluated settings. Behavioral analysis indicates that self-evolved policies increase code-mediated analysis and activate computations relevant to the prevailing regime, while case-level policy-to-return attributions trace how policy-induced allocation changes translate into realized return differences.
A case study highlights the framework's risk-aware behavior: during NVIDIA's 17% single-day drop, EvolveTrade kept exposure low under its refined sizing policy, holding just 2.9% in NVDA compared with the static baseline's 10.7%, thereby limiting losses.
The authors conclude that adapting the reusable procedure governing tool use — rather than only the underlying model — is a key direction for building more robust LLM trading agents.