Guardado en:
Detalles Bibliográficos
Autores principales: Korkmaz, Buse Sibel, Nair, Rahul, Daly, Elizabeth M., Chanona, Antonio del Rio
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2505.19327
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866910967618600960
author Korkmaz, Buse Sibel
Nair, Rahul
Daly, Elizabeth M.
Chanona, Antonio del Rio
author_facet Korkmaz, Buse Sibel
Nair, Rahul
Daly, Elizabeth M.
Chanona, Antonio del Rio
contents Current debiasing approaches often result a degradation in model capabilities such as factual accuracy and knowledge retention. Through systematic evaluation across multiple benchmarks, we demonstrate that existing debiasing methods face fundamental trade-offs, particularly in smaller models, leading to reduced truthfulness, knowledge loss, or unintelligible outputs. To address these limitations, we propose a contrastive learning framework that learns through carefully constructed positive and negative examples. Our approach introduces contrast computation and dynamic loss scaling to balance bias mitigation with faithfulness preservation. Experimental results across multiple model scales demonstrate that our method achieves substantial improvements in both toxicity reduction and faithfulness preservation. Most importantly, we show that our framework is the first to consistently improve both metrics simultaneously, avoiding the capability degradation characteristic of existing approaches. These results suggest that explicit modeling of both positive and negative examples through contrastive learning could be a promising direction for reducing the alignment tax in language model debiasing.
format Preprint
id arxiv_https___arxiv_org_abs_2505_19327
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Paying Alignment Tax with Contrastive Learning
Korkmaz, Buse Sibel
Nair, Rahul
Daly, Elizabeth M.
Chanona, Antonio del Rio
Machine Learning
Current debiasing approaches often result a degradation in model capabilities such as factual accuracy and knowledge retention. Through systematic evaluation across multiple benchmarks, we demonstrate that existing debiasing methods face fundamental trade-offs, particularly in smaller models, leading to reduced truthfulness, knowledge loss, or unintelligible outputs. To address these limitations, we propose a contrastive learning framework that learns through carefully constructed positive and negative examples. Our approach introduces contrast computation and dynamic loss scaling to balance bias mitigation with faithfulness preservation. Experimental results across multiple model scales demonstrate that our method achieves substantial improvements in both toxicity reduction and faithfulness preservation. Most importantly, we show that our framework is the first to consistently improve both metrics simultaneously, avoiding the capability degradation characteristic of existing approaches. These results suggest that explicit modeling of both positive and negative examples through contrastive learning could be a promising direction for reducing the alignment tax in language model debiasing.
title Paying Alignment Tax with Contrastive Learning
topic Machine Learning
url https://arxiv.org/abs/2505.19327