Tag
该论文提出一种将LLM代理的自进化限制在运行时harness（指令文本、工具调用逻辑与原语组合）的框架，模型权重保持固定，使每次适应都成为可审查的diff，并通过哈希链记录的准入门进行管控。模拟实验显示受控双循环仅放行144/7449个候选变更且未恶化误差，而无约束系统放行了309个有害变更。
This paper presents a design-science audit protocol for detecting supervision drift in proxy-labeled credit-risk prediction models, with multi-signal diagnostics tested on LendingClub data.
This paper builds a multi-scale stacking ensemble for credit risk scoring and audits LLM-generated explanations, finding that ranking gains are real but small while the narrative explanations are often unfaithful, with SHAP and LIME agreeing on important features but not their order or sign.
The Bank for International Settlements warns that an AI-driven market selloff could quickly spill into credit, citing past stress episodes and an opaque private credit system.