Steerable AI models balance conflicting user goals better
Which Objectives Need a Dial? Predicting Objective Conflict and Covering Trade-offs in Steerable Pluralistic Alignment
Artificial IntelligenceComputation and LanguageComputers and Society
Summary
People often want different things from AI, so one model cannot satisfy everyone. The authors study how AI models can be adjusted, or steered, to handle different goals at once. They find ways to predict when two goals clash and methods to cover many trade-offs without making a separate model for each. Their results guide building AI systems that better respect diverse preferences.
What this means in practice
- •For ai product developers: Build AI products that can adjust output according to multiple, sometimes conflicting user preferences without retraining separate models for each.$Commercial implications: Allows creation of customizable AI services that cater to diverse user values, enabling market differentiation and broader user appeal.
- •For ai system engineers: Optimize steering strategies to cover many trade-offs between objectives efficiently using parameter merging and model selection.
Authors
David Tsoi, Esra Dönmez
Abstract
People hold diverse, sometimes conflicting values, so no single aligned model can satisfy everyone. Pluralistic alignment therefore calls for steerable models that can balance competing objectives differently. Multi-Objective Direct Preference Optimization (MODPO) does this by using an objective weight to span a continuum of trade-offs. We study two questions: when can one model improve two objectives simultaneously, and how can many trade-offs be covered without training a separate model for each? Across seven objective pairs from HelpSteer and UltraFeedback, two pre-training measurements predict whether objectives align or conflict for human-annotated data, but not for AI-annotated data, where response length and repetition confound reward-model scores. For broader trade-off coverage, selecting the nearest trained model and merging model parameters both help, but neither consistently matches direct training. These findings yield practical guidance for building steerable models that serve diverse preferences.