FGR-ColBERT: Identifying Fine-Grained Relevance Tokens During Retrieval

Batch of recent IR/ML arXiv papers (Apr 2026) covering advances in retrieval, LLM integration, multimodal representation, recommendation scaling, dataset construction, and agentic RAG robustness. Key contributions: FGR‑ColBERT — distills LLM token‑level relevance into an efficient ColBERT variant that matches retrieval recall and significantly improves token‑level F1 versus a much larger LLM; Evidence Units — parser‑independent grouping of visual assets and context to form semantically complete chunks that boost retrieval metrics; UniMixer — a unified architecture and scaling framework for Re‑