Xue

Mandate Salience Decay can lead to a 4.4x behavioral gap in financial agents over time, revealing critical vulnerabilities in their long-term deployment.

Neighboring inputs can trigger malicious behaviors in Vision-Language Agentic Systems with alarming precision, revealing a hidden vulnerability in their design.

LLMs can achieve more consistent and reliable cross-jurisdictional financial reporting by acting as constrained verifiers within a structured, agentic workflow, rather than as free-form generators.

Finally, a rigorous benchmark, VERO, to understand how coding agents can iteratively improve *other* agents, revealing which optimization strategies actually work.