Value Action · Interactive research artifact

Value–Action Feature Atlas

Explore which learned activation patterns inside a language model are associated with choosing actions that support or conflict with a human value. Compare association screens with causal tests, then inspect plain-language explanations and examples.

How to read this atlas 56 human values Association + causal evidence

Original dataset and paper

Built on ValueActionLens by Hua Shen, Nicholas Clark, and Tanu Mitra

Mind the Value-Action Gap: Do LLMs Act in Alignment with Their Values? (EMNLP 2025 Outstanding Paper) introduced a framework and dataset of 14.8K value-informed actions across 12 cultures and 11 social topics. This atlas focuses on its action-choice setting across the same 56 values from Schwartz’s Theory of Basic Values, then adds sparse-feature interpretation and causal intervention tests.

Read the original paper
Value family -
Decision examples -
Action balance -
Causal coverage -

Ranked Features

0 rows
Rank Feature Evidence score Activation rate Source audit Model location Internal review Feature type Warning Category Interpretation