A Kitchener-Waterloo EdTech company has been recognized for its research that proposes a solution to a common AI “hallucination” that happens when models confidently present incorrect results when extracting information from documents.
ApplyBoard announced today that its research team won the Best Paper Award at the Institute of Electrical and Electronics Engineers (IEEE)’s 15th International Conference on Pattern Recognition Systems (ICPRS 2025) in London, UK.
“In education technology, accuracy is essential… A single data error can affect a student’s admission decision and their future.”
That paper, Embedding Confidence to Enhance Trust in AI Document Entity Extraction, focuses on tackling the “silent failure” problem—when AI systems hallucinate, or generate incorrect information, without indicating uncertainty. The paper suggests a new way to verify the accuracy and reliability of large language models (LLMs) when processing student transcripts, resumes, and other unstructured documents: a “confidence score” that helps gauge the trustworthiness of the LLM’s output.
While LLMs can already help organizations extract information from documents much faster and more efficiently than humans, they are also prone to misreading students’ GPAs or incorrectly transcribing their course grades and failing to flag potential errors. ApplyBoard says its proposed system has proven so effective that the startup plans to implement it into its online platform, which helps international students apply to colleges and universities abroad.
ApplyBoard VP of product development and paper co-author Sina Meraji, along with fellow ApplyBoard researchers Matthew MacDonald, Sina Khosravi, and Arash Ramin, surveyed existing techniques and developed the “confidence score.”
ApplyBoard describes the system as “a quality assurance layer” that analyzes mathematical embeddings of AI outputs to give each piece of extracted data a confidence score. The company’s research team claimed that when tested, its system was able to distinguish between accurate extractions and errors at a 98 percent accuracy rate.
RELATED: ApplyBoard lays off employees due to “policy changes across major study destinations”
“This recognition from the IEEE validates our commitment to making AI not just powerful, but trustworthy,” Meraji said in a statement. “In education technology, accuracy is essential… A single data error can affect a student’s admission decision and their future.”
ApplyBoard plans to use this tech to implement a “traffic light” approach within its own platform that prioritizes applications for human review based on their confidence scores and surfaces ones with lower certainty for greater scrutiny.
The company expects this system to process hundreds of thousands of student applications annually and reduce turnaround times from application to submission.
Feature image courtesy ApplyBoard. Photo via LinkedIn.