Coherence Ladders: A Parametric Variation Dataset for Testing LLM Preference Coherence
Dataset Description
This dataset contains 141 parametrically varied preference ladders designed to operationally test whether large language models exhibit genuinely coherent preferences. Each ladder varies a single normatively relevant property across 7 tiers of increasing magnitude, enabling a monotonicity test: if a model coherently values property P, its preference for outcomes with… See the full description on the dataset page: https://huggingface.co/datasets/Teddybearresearch/coherence-ladders-141.
