Model Alignment

Definitions

1. The extent to which a model's behavior matches intended goals, human values, policy constraints, or task requirements. Model alignment is a broad concern spanning safety, steerability, and usefulness.

   In plain English: How well a model's behavior matches the goals and rules people want it to follow.

   Example: "They treated refusal quality as part of model alignment because the assistant needed to decline unsafe requests without becoming unusably restrictive."

Related Terms

• Alignment Tax
• Context Extension
• Data Mixture
• Dense Model
• Frontier Model
• Goal Misgeneralization
• Goodhart's Law AI
• Human Preference
• Internal Representation
• Large Context Model
• Likelihood
• Model Fingerprint
• Power-Seeking
• Preference Model
• Self-Critique
• Specification Gaming
• Student Model