Science ❯ Computer Science ❯ Artificial Intelligence ❯ Machine Learning

Model Evaluation

Poetic Prompts Defeat AI Safety Across Models, Averaging 62% Success

Single-turn verse reformulations boosted jailbreak rates far beyond prose baselines, revealing a systemic weakness in current safety evaluations.

RAG Advances Spotlight Structured Retrieval Gains and Safety Risks in Agentic RL Search

LLM-Agent Research Coalesces Around Generalizability, RLVR, and Grounded Retrieval

LLM Vulnerabilities Drive Emergence of Self-Improving AI Frameworks