Overview

• A new review highlights persistent hallucinations in LLM outputs and calls for integrated fact-checking frameworks using advanced prompting, domain-specific fine-tuning and retrieval-augmented generation
• Privacy research shows LLM-powered recommenders can be inverted to recover about 65 percent of user interactions and infer age and gender in 87 percent of cases
• Security assessments reveal that up to 94.4 percent of LLM agents are vulnerable to direct prompt injection, 83.3 percent to RAG backdoors and all tested models to inter-agent trust exploits
• P-Aligner, a lightweight instruction pre-alignment module, improves coherence and achieves average win-rate gains of 28.35 percent on GPT-4-turbo and 8.69 percent on Gemma-2-SimPO
• TRAIL introduces joint inference with dynamic knowledge graph refinement, outperforming existing KG-augmented and RAG baselines by 3 to 13 percent while supporting continual learning without retraining