Self-Critique Loop

The foundational rule set against which outputs are evaluated. This is the AI's 'constitution'—a formal, machine-readable list of principles covering safety, ethics, legality, and task-specific constraints (e.g., 'do not provide harmful instructions', 'ensure factual accuracy'). The specificity and clarity of these principles directly determine the loop's effectiveness.

The initial draft output created by the primary language model in response to a user query. This is the 'first pass' response that will be subjected to critique. Its quality and scope set the starting point for the revision process. In advanced implementations, the model may generate multiple candidate proposals for comparative evaluation.

The evaluation engine that analyzes the proposal. This can be:

• The same base model prompted to act as a critic.
• A specialized classifier fine-tuned to detect principle violations.
• A separate evaluator model with distinct capabilities.

Its core function is to identify flaws, such as factual inaccuracies, safety violations, logical inconsistencies, or deviations from formatting rules, and provide specific, actionable feedback.

The corrective action phase where the initial proposal is updated based on the critique. The system must integrate the feedback to produce a revised output that addresses the identified issues while preserving relevant, correct information. This often involves a new generation call with an augmented prompt that includes the original query, the flawed proposal, and the critique.

The decision logic that determines whether the loop continues or terminates. It defines stopping criteria, such as:

• A maximum number of iterations (e.g., 3 cycles).
• The critique model finding zero violations.
• The revision yielding diminishing returns in improvement scores.

This component prevents infinite loops and manages computational cost.

The telemetry system that records the loop's internal operations for transparency and governance. It logs the original proposal, the generated critique, the revised output, and the principle(s) invoked. This creates an audit trail essential for debugging, compliance (e.g., EU AI Act), and validating that the self-critique process was executed correctly.

The overarching framework that defines the self-critique loop. It governs AI behavior by training models to adhere to a predefined set of core principles or a 'constitution'. The self-critique loop is the primary mechanism by which a model operationalizes these principles, evaluating and revising its own outputs against them before final generation.

A scalable alignment technique that often utilizes a self-critique loop. Instead of human feedback, an AI 'critic' model—guided by a constitution—generates preferences used to train the main model. The self-critique loop can be seen as a single, inference-time instance of this broader training paradigm, where the model acts as its own critic.

A downstream, programmatic check that complements the self-critique loop. While self-critique is an internal, model-driven process, output verification is an external safety net. It involves automatically scanning the final generated text for compliance with safety, factual accuracy, and formatting rules after the self-critique and generation process is complete.

A potential outcome of a self-critique loop. If the model's self-evaluation determines that a compliant revision is impossible or that the query fundamentally violates its principles, it triggers a refusal. An explainable refusal extends this by providing a clear, principle-based justification, which is a direct product of the critique step.

The technique that supplies the principles for the self-critique loop. It involves explicitly including the constitutional rules within the model's system prompt or in-context instructions. This text directly guides the model's internal critique process, telling it what to evaluate its outputs against during the loop.

The underlying machine learning task that powers the evaluation phase of a self-critique loop. The model must judge its own proposed output, which requires an internalized understanding of what constitutes a 'preferred' or 'aligned' response according to its constitution. This capability is often developed through training on preference data, as in RLHF or RLAIF.