Can I read Noise Contrastive Alignment for LMs on EtoBox?

Noise Contrastive Alignment for LMs by thai.nq107 is a document available to read on EtoBox.

What is Noise Contrastive Alignment for LMs about?

This paper presents a framework for aligning language models (LMs) with explicit rewards using Noise Contrastive Estimation (NCE), introducing two algorithms: NCA and InfoNCA. Unlike existing methods like Direct Preference Optimization (DPO), which rely on pairwise preference data, the proposed methods can directly utilize reward datasets with scalar evaluations, leading to improved performance in complex reasoning tasks. Experimental results demonstrate that InfoNCA and NCA outperform traditional preferenc

Author

thai.nq107

Language

EN