Jing Huang

LLM-powered similarity analysis can significantly enhance the quality and coherence of item selection in large-scale assessments, outperforming traditional metrics.

Language models struggle to consistently encode the current year, with associative and declarative representations diverging in their update responses.

Tool-using LLMs face a near-universal robustness gap, but combining Bayesian Tool Memory with reinforcement learning can boost recovery performance by over 40% in failure scenarios.