Research analyzing LLM citation granularity across four model scales (8B-120B) reveals that enforcing fine-grained sentence-level citations degrades attribution quality by up to 276% compared to optimal granularity. Paragraph-level citations consistently perform best, with larger models suffering disproportionately from fine-grained constraints that disrupt multi-sentence reasoning needed for accurate attribution.
Research
Are Finer Citations Always Better? Rethinking Granularity for Attributed Generation
Fine-grained sentence-level citations degrade LLM attribution quality by up to 276% versus paragraph-level granularity, with larger models suffering most from the constraint.
Monday, April 6, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline
Tags
research