Tag
论文提出 Investor's Last Exam (ILE)，一个由买方投资者主导的股票研究基准，评估AI智能体在冻结信息集下的研究质量、预测准确度与模拟投资表现，结果显示最强模型（GPT-6 Astra、Claude Opus 5.5）研究能力强但预测表现仍显著不足。
A developer has launched a public beta of FaultFence, a local-first testing kit for financial AI agents that includes 17 security scenarios such as duplicate payments, tool poisoning, and budget rule bypasses, and is seeking feedback from 10-15 developers.
Investor's Last Exam (ILE) 是一个由资深买方投资者主导的股票研究基准，使用冻结证据、专家评分标准和金融依赖图评估 AI 智能体的端到端研究能力，结果显示顶尖模型研究能力强但在预测和投资判断上仍然失败。
8004scan and Krexa are partnering to explore how ERC-8004 identity and reputation standards can support financial agents on Solana, integrating agent deployment with onchain credit infrastructure.
This paper audits self-evolution mechanisms in financial AI agents, revealing capability improvements alongside security risks such as prompt injection drift and execution-interface mismatches, emphasizing the need for holistic auditing.
Introduces FinPersona-Bench, a benchmark to measure how well autonomous financial agents maintain their assigned behavioral mandates over time, revealing Mandate Salience Decay (MSD) that worsens with temporal distance and varies by model and agent profile.
This paper proposes an orchestration framework for financial agents that maps traditional algorithmic trading components to AI agents, demonstrating improved returns in stock and BTC trading compared to market benchmarks.