Newest papers

Rethinking Post-hoc Debiasing of Large Pretrained Models

Divyaansh Mertia, Chirag Bhutra, Himanshu Shekhar, Shubham Bagwari, Pratik Mazumder · 2026 · arXiv

DOI: 10.48550/arXiv.2610.29688 · open-access PDF

Spanish

Grandes modelos pre-entrenados como CLIP se utilizan ampliamente como extractores de características congeladas. Sin embargo, sus embeddings heredan correlaciones falsas de los datos de entrenamiento, introduciendo bias en las características. Un cuerpo creciente de investigación ha buscado abordar este problema a través de la debiasing post-hoc, dejando intacta la columna vertebral. Descubrimos que a pesar de los diferentes mecanismos, estos métodos requieren que la característica de bias sea identificada y expresada en texto. Esta dependencia no es inofensiva. En este trabajo, demostramos experimentalmente que este requisito incurre en una sanción imprevista en ciertos escenarios. Para el examen

Machine translation (Meta M2M100); the English original follows.

Abstract

Large pretrained models such as CLIP are widely used as frozen feature extractors. However, their embeddings inherit spurious correlations from training data, introducing bias into the features. A growing body of research has sought to address this issue through post hoc debiasing, leaving the backbone untouched. We discover that despite differing mechanisms, these methods require the bias feature to be identified and expressed in text. This dependency is not innocuous. In this work, we experimentally demonstrate that this requirement incurs an unforeseen penalty in certain scenarios. For exam

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange