Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems

Feed of recent arXiv papers (Jun 10 2026) on recommender systems, LLM alignment, retrieval, diagnostics and benchmarks. Key contributions: Mult-DPO — a multinomial DPO objective that tractably aligns LLMs to set-wise (multi-positive) preferences and bounds a Plackett–Luce marginalized loss; MetaPlate — a counterfactual-guided RAG+LLM system that uses CGM and wearable signals to produce personalized meal recommendations to reduce postprandial hyperglycemia (expert-in-the-loop evaluation; privacy/health-data implications; code-backed RAG retrieval over USDA); τ-Rec — a verifiable benchmark and R