Bird‑flocking method reduces AI summary errors ^{CEFR B1}

A team at New York University, led by Anasse Bari with coauthor Binxu Huang, published a framework in the journal Frontiers in Artificial Intelligence. The framework is a preprocessing step that prepares long documents before a large language model generates a final summary.

In the first phase the method cleans sentences by keeping nouns, verbs and adjectives and by merging multi‑word terms so single concepts stay intact. Each sentence is converted into a numerical vector that fuses lexical, semantic and topical features. Sentences are scored for document‑wide centrality, section‑level importance and alignment with the abstract, and key sections like the Introduction, Results and Conclusion receive extra weight.

In the second phase the framework applies bird‑flocking principles — cohesion, alignment and separation — so similar sentences cluster and leaders emerge. From each cluster the highest‑scoring sentences are selected, reordered and passed to an LLM that synthesizes a fluent summary grounded in the source. The researchers tested the approach on over 9,000 documents and found it produced summaries with greater factual accuracy than LLMs alone. Bari says the framework is meant to help LLMs stay closer to source material, and the authors note it reduces but does not eliminate hallucination risk.