Relationship to Instruction Following

While both Constitutional AI (CAI) and standard Instruction Fine-Tuning (IFT) aim to shape the behavior of Large Language Models (LLMs), they operate through distinct mechanisms and target different aspects of model alignment. Understanding their relationship is important for appreciating the specific contribution of CAI within the broader toolkit for LLM development.

IFT primarily focuses on teaching the LLM to follow explicit instructions accurately and effectively. It's a form of supervised learning where the model is trained on a dataset typically composed of pairs: (instruction, desired_output). The goal is to minimize the difference between the model's generated response and the provided desired_output for a given instruction. This approach is highly effective for imbuing models with specific skills, enabling them to perform tasks like summarization, translation, question answering, or adopting particular personas or formats as directed. The supervision signal is direct: produce this output for this input instruction.

Constitutional AI, particularly in its initial supervised learning (SL) phase discussed in this chapter, takes a different approach. Instead of optimizing for adherence to specific input-output examples provided by humans, CAI optimizes for adherence to a set of explicit, written principles, the constitution. The objective is not just task execution, but behavioral modification guided by these principles.

Here's a breakdown of the primary distinctions:

Objective and Supervision Source

• IFT: The objective is task competence and instruction adherence. The supervision comes from curated or human-generated examples of correct instruction execution. The model learns a direct mapping from an instruction prompt to a desired response format or content.
• CAI (SL Phase): The objective is principle adherence. The supervision arises from an AI-driven self-correction process. The model first generates an initial response to a prompt. Then, guided by the constitution, an AI system (or potentially the same LLM prompted appropriately) critiques this response based on the principles and subsequently revises it. The fine-tuning dataset consists of examples derived from this process, often training the model to predict the revised response given the initial prompt. The learning signal is indirect, mediated by the application of constitutional principles during the critique and revision steps.

Nature of Control

• IFT: Offers direct control over how the model should respond to specific types of instructions present in the training data. It excels at teaching the model what to do in defined scenarios.
• CAI: Provides a more generalized form of behavioral control. It aims to instill overarching rules or constraints (like "be harmless" or "avoid biased statements") that should apply across a wide range of inputs, even those not seen during training. It focuses more on how the model should behave generally, rather than specifying exact outputs for every possible instruction.

Data Generation Process

• IFT: Data generation often involves significant human effort in writing instructions and corresponding high-quality outputs, or sophisticated methods for synthesizing such pairs. Quality control of these pairs is a central concern.
• CAI: Data generation relies on the effectiveness of the AI critique and revision models operating under the guidance of the constitution. Human effort is shifted towards designing a comprehensive and effective constitution and potentially overseeing or refining the AI feedback process itself. This process can generate vast amounts of training data reflecting the desired principles without requiring humans to specify the "correct" output for every single prompt.

Generalization Properties

• IFT: Generalization depends heavily on the breadth and quality of the instruction dataset. The model might struggle to follow instructions or exhibit desired behaviors for prompts substantially different from the training distribution.
• CAI: Aims for broader behavioral generalization based on principles. If the constitution captures the desired behavioral traits effectively, the model should ideally exhibit these traits even when faced with novel situations or prompts, as the underlying principles guide its response generation process.

Complementarity

It's important to recognize that IFT and CAI are not mutually exclusive; they can be complementary techniques. An LLM might first undergo extensive instruction fine-tuning to acquire a wide range of skills and the ability to follow directions. Subsequently, CAI methods (both the supervised phase described here and the reinforcement learning phase discussed later) can be applied to align the model's behavior with specific ethical or safety principles outlined in a constitution, refining how it executes those instructions. This layering allows for building models that are both capable and aligned with desired norms.

In essence, while IFT teaches the model to follow orders, CAI (in its SL phase) begins the process of teaching the model an internal set of rules to guide its actions, derived from the provided constitution. This distinction forms the foundation for understanding how CAI contributes uniquely to the challenge of scalable LLM alignment.