TraderBench：AI代理在对抗性资本市场中的稳健性如何？

TraderBench is a newly proposed benchmark designed to address two major challenges in evaluating AI agents in finance. Traditional static benchmarks require costly expert annotation and fail to capture the dynamic decision-making process central to real-world trading. At the same time, LLM-based judges introduce uncontrolled variance in domain-specific tasks. TraderBench combines expert-verified static tasks (knowledge retrieval, analytical reasoning) with adversarial trading simulations scored purely on realized performance metrics such as Sharpe ratio, returns, and drawdown, thereby eliminating judge variance entirely. The framework features two novel tracks: crypto trading with four progressive market-manipulation transformations, and options derivatives scoring based on P&L accuracy, Greeks, and risk management.