Year: 2025
Type: conference-paper
Abstract: Large language models are increasingly used to synthesize scientific literature, yet they remain prone to hallucination -claims that are linguistically fluent but lack support in the cited sources.We tackle hallucination detection in the SCIHAL 2025 challenge by augmenting SCIBERT with Triplet and InfoNCE contrastive objectives in addition to cross-entropy classification.The system achieves validation macro-F 1 scores of 0.626±0.004on the coarsegrained hallucination detection task (Sub-task 1) and 0.632 ± 0.012 on the fine-grained detection task (Sub-task 2), exceeding a plain SCIB-ERT baseline by more than three points.The official blind test set scores reach macro-F 1 scores of 0.51 and 0.43 for Sub-tasks 1 and 2, respectively, securing fifth place in both leaderboards.Confusion matrix analysis shows that contrastive learning markedly improves majority classes, whereas sparse categories, especially Missing Information, remain challenging despite aggressive attempts to mitigate class imbalance.
Authors: Crivoi Carla, Ana Sabina Uban
Institution: University of Bucharest
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4412944220