MARVEL: Modular Abstention for Reliable and Versatile Expert LLMs
Jul 1, 2025·
,,,,,,·
1 min read
Bingbing Wen
Faeze Brahman
Zhan Su
Shangbin Feng
Yulia Tsvetkov
Lucy Lu Wang
Bill Howe
Abstract
We present MARVEL, a modular abstention framework for building reliable and versatile expert large language models. Our approach enables models to selectively abstain from answering questions they are uncertain about, improving reliability and trustworthiness.
Type
Publication
ICML Reliable Foundation Model Workshop & NeurIPS Submission
We present MARVEL, a modular abstention framework for building reliable and versatile expert large language models. Our approach enables models to selectively abstain from answering questions they are uncertain about, improving reliability and trustworthiness in AI systems.
