L2-Bench: An Evaluation Benchmark for Measuring LLM Capabilities in Second Language Education

This collection of arXiv preprints highlights multiple security- and safety-relevant risks across AI systems, data-driven governance, and privacy-preserving methods. Key points: (1) LLMs exhibit geopolitically aligned evaluation biases—model policy judgments change depending on which nation is presented as the endorser, creating risks for automated policy analysis and manipulation of decision-support outputs. (2) Privacy-preserving techniques: LDPKiT presents a local-differential-privacy (LDP) superimposition method for non-adversarial distillation to limit input leakage when distilling from a