Papers by Hugo Aerts
Language Models are Surprisingly Fragile to Drug Names in Biomedical Benchmarks (2024.findings-emnlp)
Copied to clipboard
Jack Gallifant, Shan Chen, Pedro Moreira, Nikolaj Munch, Mingye Gao, Jackson Pond, Leo Anthony Celi, Hugo Aerts, Thomas Hartvigsen, Danielle Bitterman
| Challenge: | Medical knowledge is context-dependent and requires consistent reasoning across various natural language expressions of semantically equivalent phrases. |
| Approach: | They create a robustness dataset to evaluate performance differences on medical benchmarks . they swap brand and generic drug names using physician expert annotations based on medical terminology . |
| Outcome: | The proposed model shows a consistent performance drop of 1-10% on medical benchmarks. |
WorldMedQA-V: a multilingual, multimodal medical examination dataset for multimodal language models evaluation (2025.findings-naacl)
Copied to clipboard
João Matos, Shan Chen, Siena Kathleen V. Placino, Yingya Li, Juan Carlos Climent Pardo, Daphna Idan, Takeshi Tohyama, David Restrepo, Luis Filipe Nakayama, José María Millet Pascual-Leone, Guergana K Savova, Hugo Aerts, Leo Anthony Celi, An-Kwok Ian Wong, Danielle Bitterman, Jack Gallifant
| Challenge: | Existing multiple-choice question and answer (QA) datasets are text-only and available in a limited subset of languages and countries. |
| Approach: | They propose a multilingual, multimodal benchmarking dataset to evaluate multimodal/vision language models in healthcare. |
| Outcome: | The WorldMedQA-V includes 568 labeled multiple-choice QAs paired with 568 medical images from four countries. |
Sparse Autoencoder Features for Classifications and Transferability (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) provide potential for uncovering structured, human-interpretable representations in Large Language Models (LLMs). |
| Approach: | They analyze SAEs for interpretable feature extraction from Large Language Models in safety-critical classification tasks. |
| Outcome: | The proposed framework outperforms hidden-state and BoW models while demonstrating cross-lingual toxicity detection and visual classification tasks. |