Enabling Large Language Models to Generate Text with Citations
Tianyu Gao, Howard Yen, Jiatong Yu
“citation quality remains far from satisfactory”
001-chain-length · Support vs. chain length
Support rate decreases with hop count.
support_rate_hop1 0.91 · support_rate_hop4 0.55
002-hop-verification · Per-hop re-verification
Verifying at each hop prevents drift.
support_rate_hop4 0.88
Citation support degrades monotonically with each summarization hop.
Re-verifying spans at each hop restores over 90% of lost support.