Enabling Large Language Models to Generate Text with Citations
Tianyu Gao, Howard Yen, Jiatong Yu
LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QA
Jiajie Zhang, Yushi Bai, et al.
FactCG: Enhancing Fact Checkers with Graph-based Multi-hop Data
Deren Lei, Yaxi Li, et al.
“smaller retrieval units are easier for humans to verify against the generated statement”
“fine-grained sentence-level citations allow users to locate the exact supporting evidence”
“verification accuracy degrades as the evidence context grows”
001-tfidf · TF-IDF baseline
A stdlib TF-IDF retriever reaches usable recall@8 on claim→span matching.
recall_at_8 0.81 · span_support_rate 0.74
002-dense · Dense embeddings
Dense retrieval (BGE-M3) beats TF-IDF by a wide margin.
recall_at_8 0.84 · span_support_rate 0.76
003-anchor-robustness · Anchor robustness across versions
Quote-text anchors re-resolve after re-parsing with a different PDF extractor.
anchor_survival_rate 0.97
Anchoring citations at the span level makes them self-verifying: the quoted text either exists in the source or it does not.
Sentence-granularity passages are easier to verify than paragraph chunks without losing retrieval recall.
Span anchors survive re-indexing and new document versions because they never reference chunk or vector ids.
Dense retrieval is required for competitive recall on claim-to-source matching.