Are the Values of LLMs Structurally Aligned with Humans? A Causal Perspective

Welcome! This codebase accompanies the ACL2025 paper Are the Values of LLMs Structurally Aligned with Humans? A Causal Perspective and is based on SAELens.

1. Python Environment Setup

pip install -r requirements_vsa.txt

2. Directory Structure Setup

Set up the following directory structure outside the main project directory:

• .
  • ├── model_data
  • │ ├── google
  • │ │ └── gemma-2b-it
  • │ ├── jbloom
  • │ │ ├── Gemma-2b-IT-Residual-Stream-SAEs
  • │ ├── meta-llama
  • │ │ └── Meta-Llama-3-8B-Instruct
  • │ ├── Juliushanhanhan
  • │ │ └── llama-3-8b-it-res
  • └── SAELens
  • │ └── value_data
  • │ └── value_orientation.csv

3. Execution Instructions

Generate Data

Run the following notebook to generate data with different role and SAE settings for all values:

tutorials/value_causal_graph.ipynb

Analyze Data

After generating the result CSV files, use the following notebook for data analysis by loading the CSV files:

tutorials/value_causal_graph_analysis.ipynb