How Political Stances Affect the Reasoning Ability of Large Language Models: A Deep Study on AI Alignment Bias

This study explores the changes in the reasoning ability of large language models after inducing left/right political stances through three methods: role-play prompting, activation steering, and LoRA fine-tuning. Key findings include: political alignment affects the quality of the model on neutral reasoning tasks; in value-laden tasks, the model tends to handle controversial topics with its aligned stance; and there exists a "collapse threshold" (when alignment intensity exceeds a threshold, reasoning ability drops off a cliff). The study also provides an interactive results browser to show details of the impact.

With the widespread application of large language models (LLMs), their non-neutrality has attracted attention. Core question: How does the model's reasoning ability change after actively inducing a specific political stance? This study has academic value and practical significance for AI safety and alignment research, helping to understand and control the boundaries of AI behavior.

Three methods are used to induce political alignment: 1. Role-play prompting: Let the model play a role with a specific political tendency through system prompts (no weight modification required); 2. Activation steering: Dynamically adjust output by adding vectors to activation values of specific layers during reasoning; 3. LoRA fine-tuning: Parameter-efficient fine-tuning using low-rank adaptation technology, keeping most parameters unchanged while learning political stances.

Focused on three RQs: RQ1: Political alignment affects neutral reasoning tasks (BBH task performance varies by method and intensity); RQ2: In value-laden tasks, the model tends to handle controversial topics with its aligned stance; RQ3: There exists a collapse threshold—when alignment intensity exceeds the threshold, reasoning ability drops off a cliff (e.g., repeated output, logical breaks).

An online interactive browser is provided (link: https://0ssamaak0.github.io/political-alignment-reasoning/) with three views: 1. Discovery Tour: Guides browsing of research findings and links to evidence; 2. Example Browser: Displays model responses, supporting multi-dimensional filtering and search; 3. Intensity Explorer: Visualizes the relationship between alignment intensity and metrics (accuracy, collapse, etc.), and marks the threshold point.

Provides empirical data for the AI alignment field, indicating that political stances profoundly affect the model's reasoning mechanism, which has warning significance for building fair and reliable AI. For researchers: Provides methods to quantify the effect of alignment interventions; For policymakers/deployers: Reminds of the far-reaching impact of technical choices (e.g., fine-tuning data, system prompts). The study open-sources code, data, and the browser, laying the foundation for subsequent research and being a key step in controlling AI behavior bias.