Pluralistic Alignment in Language Models: Objectives, Human References, and Preference Representations

Key points

• Yet existing work often mixes three distinct questions: what pluralistic behavior a model should exhibit, whose preferences define the reference, and what evidence is sufficient to demonstrate success.1
• We organize the literature around a two-layer taxonomy.1
• Layer 1 distinguishes representing a range of human-supported viewpoints (Overton Pluralism), adapting faithfully to a specified perspective or preference condition (Steerable Pluralism), and matching the response distribution observed in a target population (Distributional Pluralism).1
• We synthesize a protocol-bounded corpus of 70 works on pluralistic alignment and neighboring areas, using the objective-by-human-reference matrix to map existing approaches and identify underexplored combinations.1

Sources (1)

• [1]Pluralistic Alignment in Language Models: Objectives, Human References, and Preference Representations
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 05:32 AM

  Pluralistic alignment aims to represent meaningful variation in human preferences rather than collapse heterogeneous judgments into a single target.

  Yet existing work often mixes three distinct questions: what pluralistic behavior a model should exhibit, whose preferences define the reference, and what evidence is sufficient to demonstrate success.

Extractive summary: sentences quoted from the sources.