Large Language Models are Not Always Faithful in Explaining Their Political Decision Making

• Extends the concept of faithfulness from alignment research to the political decision-making context
• Operationalizes and compares faithfulness in three distinct evaluation settings: news credibility, source framing, and hate speech detection
• Evaluates consistency under manipulated chain-of-thought (CoT) and personas

Recommended citation: Quang Minh Nguyen*, Hayden Arnold*, Taegyoon Kim
* equal contribution
Download Paper | Download Slides