ATLAS: Adaptive Trading with LLM AgentS Through Dynamic Prompt Optimization and Multi-Agent Coordination

Researchers introduce ATLAS, a multi-agent framework that uses large language models for autonomous trading by combining dynamic prompt optimization with real-time market feedback. The system addresses key challenges in deploying LLMs for finance: adapting to delayed, noisy market signals and converting model outputs into executable orders.

ATLAS represents a meaningful step forward in applying large language models to quantitative finance, tackling a fundamental problem: LLMs excel at reasoning but struggle with the feedback loops required for financial decision-making. Traditional machine learning models benefit from immediate, clean signals; markets deliver delayed, noisy feedback that obscures causality. The framework's innovation lies in Adaptive-OPRO, a prompt-optimization technique that iteratively refines trading instructions based on stochastic market outcomes rather than relying on fixed instructions or simple reflection mechanisms.

The approach emerges from growing recognition that LLMs possess useful financial reasoning capabilities but lack structural mechanisms to operationalize that reasoning. By implementing an order-aware action space—constraining model outputs to executable market orders—ATLAS bridges the critical gap between language generation and market execution. This design choice eliminates the interpretive risk of converting abstract signals into trades.

Across multiple equity regimes and LLM families, Adaptive-OPRO consistently outperformed fixed-prompt baselines, while reflection-based feedback produced inconsistent results. This finding challenges the assumption that simply asking models to reflect on their mistakes drives improvement in financial contexts. The framework integrates heterogeneous data streams—market prices, news sentiment, corporate fundamentals—allowing the agent to synthesize multifaceted information into coherent trading signals.

The research suggests that LLM-based trading systems require domain-specific optimization techniques rather than generic fine-tuning approaches. However, the paper leaves practical questions unaddressed: real-world transaction costs, robustness across market conditions, and whether the framework scales to portfolio-level decision-making with execution constraints.

• →ATLAS uses Adaptive-OPRO, a novel prompt-optimization technique, to dynamically refine trading instructions based on real-time market feedback rather than fixed prompts.
• →The framework constrains LLM outputs to an order-aware action space, ensuring model decisions map directly to executable market orders.
• →Adaptive-OPRO consistently outperforms fixed prompts across different equity regimes and LLM families, while reflection-based feedback fails to deliver systematic gains.
• →The system integrates market data, news sentiment, and corporate fundamentals through multi-agent coordination to inform trading decisions.
• →Results demonstrate that LLM-based trading requires domain-specific optimization rather than generic fine-tuning or reflection mechanisms.