How to Read the Value–Action Feature Atlas

Research question

Which internal features help explain how a language model chooses between concrete actions that support or conflict with a human value—and which features have causal influence?

Models can describe values abstractly, but their concrete choices may reveal different priorities. Across 56 human values, this atlas brings together feature rankings, interpretations, real examples, and causal tests in one inspectable interface.

1. Choose a value.Select a value such as Social Power, Equality, or Helpful.
2. Choose the evidence.Use an association screen (S1–S3) or a causal validation stage (V1–V2).
3. Inspect a feature.Read its interpretation, examples, warnings, and available intervention evidence together.

Dataset and attribution

This work starts from ValueActionLens

ValueActionLens was introduced by Hua Shen, Nicholas Clark, and Tanu Mitra in “Mind the Value-Action Gap: Do LLMs Act in Alignment with Their Values?”, an EMNLP 2025 Outstanding Paper. The framework compares what language models say they value with the actions they choose in concrete scenarios. Its dataset contains 14.8K value-informed actions across 12 cultures and 11 social topics.

This atlas focuses on the action-choice setting across the same 56 values from Schwartz’s Theory of Basic Values and extends that behavioral starting point with sparse-feature interpretation and causal intervention tests.

Read the original ValueActionLens paper

Evidence stages

One public vocabulary: S1–S3 for discovery, V1–V2 for causal validation

S1Answer-site association. Ranks features associated with the model’s choice at the answer token.

S2Content-span association. Ranks features active while the model processes the two action descriptions.

S3Pre-answer contribution. Estimates which features contribute to the choice immediately before the answer.

V1Single-feature intervention. Tests whether manipulating one feature changes support for an action.

V2Feature-group intervention. Tests the joint effect of a related group of features.

Key distinction: S1–S3 identify candidate relationships. V1–V2 test causal influence. A high association score is useful discovery evidence, but it is not causal proof by itself.

Using the atlas

What the visible controls mean

Human value: Selects one of the 56 value items and its broader value family.
Evidence stage: Moves between the three association screens and two causal validation stages.
Ranked for: Switches between features associated with Action 1 and Action 2.
Model location: Restricts where the feature was measured, such as an action span, answer token, or the state immediately before the answer.
Feature type: Separates stronger value-related interpretations from broader semantic features, formatting controls, and uncertain cases.
Warnings: Shows whether a feature may reflect a scaffold, formatting pattern, or other artifact rather than the value itself.
Search: Finds a feature ID or a word in the interpretation.

Reading a feature

Use four pieces of evidence together

1. Ranked feature

The score shows how strongly the feature appears in the selected evidence stage. Compare scores only within the current selection.

2. Interpretation

The explanation summarizes the contexts in which the feature activates. The feature-type label and warning indicate how literally to read it.

3. Concrete examples

Value-choice and broader-text examples reveal whether the interpretation matches actual activations rather than only a short label.

4. Causal evidence

When available, intervention results show whether changing the feature or feature group changes support for an action or flips the choice.

Scope: This atlas is an auditable research interface, not a definitive moral profile of the model. Individual features can encode semantics, decision scaffolding, or dataset regularities, so conclusions should combine examples, warnings, and causal validation.