One of the primary concerns surrounding artificial intelligence is its tendency to “hallucinate” and invent false information when summarising long documents. These errors are highly problematic because they spread falsehoods and force users to waste time manually checking the AI’s output for mistakes.
To combat this, computer scientists at New York University (NYU) have developed a new algorithmic framework inspired by the natural phenomenon of bird flocking.
Published in the journal Frontiers in Artificial Intelligence, the framework serves as a preprocessing step for large language models (LLMs) to produce more reliable, factual summaries of long texts.
Anasse Bari, a computer science professor at NYU’s Courant Institute School of Mathematics, Computing, and Data Science, explained the root of the hallucination problem: “One contributing factor is that when input text is excessively long, noisy, or repetitive, model performance degrades, causing AI agents and LLMs to lose track of key facts, dilute critical information among irrelevant content, or drift away from the source material entirely.”
Sentences as virtual birds
To counteract this data drift, Bari and co-author Binxu Huang designed an algorithm that treats every sentence in a long document — such as a legal analysis or scientific study — as a virtual bird.
First, the algorithm cleans and scores every sentence based on its document-wide centrality, section-level importance, and alignment with the abstract.
Next, the framework groups the sentences into clusters that mirror how real birds self-organise into flocks using three simple rules: cohesion (staying close to nearby birds), alignment (moving in the same direction), and separation (avoiding crowding).
This ensures that sentences with similar meanings naturally cluster together without overcrowding. By selecting only the highest-scoring sentences from each “flock,” the algorithm extracts a diverse range of key points, covering background, methods, results, and conclusions, without repeating a single theme.
This concisely curated and structured summary is then passed to the LLM to generate the final fluent output.
Grounding AI in reality
The researchers evaluated their new algorithm on more than 9,000 documents. They found that combining the bird-flocking framework with an LLM generated summaries with significantly greater factual accuracy than using an LLM alone.
“The intention was to ground AI models more closely to the source material while reducing repetition and noise before generating a final summary,” Bari said.
However, the researchers stress that this preprocessing step is not a magical cure-all for AI errors.
“The goal is to help the AI generate summaries that stay closer to the source material,” Bari concluded. “While this approach has the potential to partially address the issue of hallucination, we do not want to claim we have solved it — we have not.”
