Research Summary: Verifiable, Articulable, and Tacit Components of Preference

Recent research highlights that human preferences in various domains, including creative fields, often contain components that are difficult to articulate or verify, termed 'tacit.' While modern AI models are typically improved through articulated rules and verifiers, these tacit components are frequently overlooked. A new large dataset, CreativePreferences, has been introduced to study these aspects, revealing 'articulability gaps' and 'verifiability gaps' in preference modeling.

Why it matters

This research underscores a fundamental limitation in current AI development regarding the nuanced understanding of human preferences, particularly in subjective or creative contexts. Addressing these 'tacit' components is crucial for developing AI systems that can genuinely align with complex human values and decision-making, impacting product design, user experience, and automated content generation.

Key insights

• Human preferences across diverse domains include 'tacit' components that resist clear articulation or verification.
• Current AI model development primarily relies on explicitly articulated rules, rubrics, and verifiers (e.g., in RLAIF and RLVR).
• The 'tacit' elements of preferences are generally understudied in AI model improvement.
• A new large-scale dataset, CreativePreferences (2.8M texts, 317M human judgments across 7 creative domains), has been developed to investigate these preferences.
• The study models preference labels using executable programs (Verifiable), rubric banks (Articulable), and densely trained models (Tacit), leading to the VAT framework.
• The research identifies significant 'articulability gaps' (VAT-VA) and 'verifiability gaps' (VA) when trying to capture human preferences purely through articulated or verifiable means.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.03025