Plurality of Human Preferences for AI Systems

LLM groupthink reduces the plurality of outputs and preferences, motivating research into diversity-enhancing methods.

Framework argues evaluation and alignment must preserve minority perspectives instead of collapsing labels by plurality vote.

Both findings challenge the assumption that RLHF and fine-tuning reliably produce aligned models that reflect diverse human preferences.

LLM output convergence problems directly relate to the challenge of representing plurality of human preferences in AI systems.

IFLLM dataset reveals diverse gazing and mouse behaviors reflecting plurality of user preferences for LLM responses.

The value-theory study critiques preference stand-ins that flatten culturally situated human values.

Dialect-tax findings show pipeline inequalities persist even when semantic equivalence is recognized.

Broad five-value user demand versus narrow drifting model constitutions leaves large user constituencies uncovered.

Demand for constitutional value rankings is broad while frontier supply is narrow, formalizing plurality shortfalls in defaults.

Scaling laws for social simulation fidelity reveal that opinion and behavioral modeling improve with compute, informing how AI systems represent human preference diversity.