Researchers build AI to detect citation cartels in scholarly networks

A trust-aware pipeline combines citation graph structure with semantic citation intent to rank suspicious paper-level communities for audit. Researchers used an LLM teacher to label 205,897 citation pairs, trained a SciBERT student, and scaled citation-intent typing to 2.04 million unique graph edges. The Composite Cartel Index (CCI), which integrates multiple metrics, identified a top community of 1,079 papers with 254.3 times more internal citations than expected—64.2% of them superficial. Comparisons against several baselines showed CCI cannot be reduced to a single heuristic, and edge excision validation confirmed its selected communities behave distinctly from random removals.