An LLM-powered Agentic Recommendation System for Connected TV Content Discovery

Collection of recent IR/recsys papers (July 2026) covering LLM-powered recommendation and retrieval engineering, efficient RAG and embedding techniques, model compression for edge devices, and an adversarial-attack study. Key contributions include: an agentic LLM+ML hybrid for Connected TV recommendations addressing latency and heterogenous context integration; soft-token fusion to inject numerical/embedding features into LLM inputs for retrieval; empirical demonstration that multilingual prompt-injection attacks can reliably inflate LLM relevance judgments and evade current defenses; scalable