EvolveTrade: KAIST Study Shows AI Agents Evolve Prompts
A new KAIST paper introduces EvolveTrade, letting AI agents self-evolve system prompts to bind Python calculations to portfolio allocation, hitting a 4.00 Sharpe ratio.
Stock photo for illustration only, not from the actual event
- KAIST's EvolveTrade research fixes the gap where AI agents run tools but ignore quantitative output.
- A separate policy agent reviews execution traces and rewrites operational prompts every 5 trading days.
- Testing shows a 4.00 Sharpe ratio and 10.56% cumulative return, outperforming static baselines.
Most autonomous agent benchmarks today evaluate tool use as a superficial information-gathering exercise. Developers hand models API keys for stock prices, search tools for financial news, and a Python REPL, then check if it makes correct API calls before writing a summary. In practice, giving agents a code interpreter does little if the system prompt treats code execution merely as descriptive background chatter while allocating capital using vague qualitative heuristics.
To fix this, a new paper from KAIST, EvolveTrade (Kim et al., arXiv:2609.17632), stops hardcoding the tool-use contract at deployment. Instead, the researchers treat the operational system prompt as a text-parameterized policy that continuously adapts based on historical performance.
Stock photo for illustration only, not from the actual event
In standard setups like LiveTradeBench using GPT-5-mini, the static agent ran Python diligently, calculating that JPMorgan had a 20-day return of +3.45% with a 2.105 Sharpe ratio, and Caterpillar sat at 1.497. Yet, final allocations barely budged from an arbitrary baseline, leaving JPMorgan at 6% and Caterpillar at 4%. The quantitative metrics remained trapped inside the reasoning trace without driving actual portfolio weight adjustments.
Under EvolveTrade, a separate policy agent reviews execution traces and realized returns every five trading days, rewriting the operational prompt to dictate how the trading agent utilizes its tools. Tested across sideways, drawdown, and bull markets with GPT-5-mini and Gemini-2.5-Flash, the evolved prompt systematically shifted metric calculation out of qualitative text and directly into executable code allocation.
"In the September 10, 2025 trace highlighted by the authors, the static agent called Python once for aggregate stats. The evolved agent called Python 11 times: running metric calibration, validating risk constraints, and generating exact percentage weights."
Kim et al., EvolveTrade
The concept of self-evolving prompts tackles a major challenge in production AI systems known as prompt bloat or context degradation. By allowing the policy length to fluctuate non-monotonically (between 5,000 and 20,000 characters) based on market regimes, EvolveTrade demonstrates that automated systems can incorporate validation checks and prune redundant rules dynamically. For developers deploying autonomous financial agents, this proves that deterministic binding between tool execution and final actions is critical to avoid fragile qualitative LLM prose.
Across 50-day testing, EvolveTrade maintained a 4.00 Sharpe ratio and a 10.56% cumulative return against the static agent's 2.94 Sharpe and 8.88% return, while keeping maximum drawdown at 2.96% versus 4.40%. For software engineers building production workflows, the lesson is clear: unless your system prompt explicitly forces models to bind final actions to deterministic calculations, runtime tool outputs will remain underutilized.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment