Trait, Not State: The Durability of Reading Identity in Social Highlighting

This collection of arXiv announcements highlights advances and empirical audits in retrieval, recommender systems, and LLM-agent tooling with several security and safety implications. Key items: (1) ToolSense — an open-source diagnostic framework exposing a knowledge-retrieval dissociation in parametric tool-retrieval LLMs (ToolBench), showing much lower realistic retrieval and near-random factual knowledge on probes; (2) CQC-RAG and CoDeR — methods to reduce hallucinations and constraint-violating evidence exposure by using cross-query consistency and separating topical relevance from local-‘