liveBook · Manning

Policy alignment

As explained in LLM Evaluation and Alignment, The Foundational Ideas

Policy alignment makes a model follow explicit operational rules, procedures, priorities, and external requirements.

Policy alignment makes a model follow specific, actionable rules: procedures, instruction priorities, and institutional or legal requirements. It is distinct from alignment to values or communication style, but the three dimensions can conflict. section 1.2.1

A behavioral specification can make policies explicit through a chain of command and rules for categories such as illegal activity. Constitutional AI can encode procedural rules as principles, but applying them to novel cases depends on the critic’s interpretation. section 8.4.3 section 8.1.4

Tool output and instruction priority

A policy can require treating tool output as data rather than instructions, so it cannot override user or system instructions. section 8.4.3

Figure 1.9 Three pillars of alignment and how they relate to each other.

The three-pillar diagram situates policy alignment alongside principled and personality alignment. section 1.2.1

How LLM Evaluation and Alignment, The Foundational Ideas covers it

How deep it goes
Distinguishes explicit chain-of-command rules from rules encoded as principles, noting interpretation risk without detailing an implementation.
Examples
Uses illegal-activity refusals, tool-output handling, and company procedures or legal requirements to illustrate policy rules.
What you can do after
Apply it by writing explicit instruction priorities and compliant or refusal procedures for defined request categories.
How it connects
Connects policy alignment to principled alignment, personality alignment, RLHF, Constitutional AI, and the chain of command.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Evaluation and Alignment, The Foundational Ideas. The section links point to the parts of the book it draws on.