VISPA: Pluralistic Alignment via Automatic Value Selection and Activation

TL;DR

VISPA is a training-free framework that enables pluralistic alignment of language models by controlling value expression through dynamic selection and internal activation steering, improving their ability to reflect diverse perspectives in high-stakes domains.

Contribution

VISPA introduces a novel, training-free method for pluralistic alignment that uses internal activation mechanisms for value control and representation, addressing limitations of existing approaches.

Findings

PERFORMS well across multiple models and settings

ADAPTABLE with different steering methods and values

ENHANCES pluralism in high-stakes applications

Abstract

As large language models are increasingly used in high-stakes domains, it is essential that their outputs reflect not average} human preference, rather range of varying perspectives. Achieving such pluralism, however, remains challenging. Existing approaches consider limited values or rely on prompt-level interventions, lacking value control and representation. To address this, we introduce VISPA, a training-free pluralistic alignment framework, that enables direct control over value expression by dynamic selection and internal model activation steering. Across extensive empirical studies spanning multiple models and evaluation settings, we show VISPA is performant across all pluralistic alignment modes in healthcare and beyond. Further analysis reveals VISPA is adaptable with different steering initiations, model, and/or values. These results suggest that pluralistic alignment can be achieved through internal activation mechanisms, offering a scalable path toward language models that serves all.