Self-Reflection / Critique

LLMs make mistakes. They hallucinate facts, miss edge cases, produce logically inconsistent reasoning, and generate outputs that satisfy the surface form of a request without meeting its underlying requirements. Self-reflection tackles this by inserting a critique step into the agent loop: after the primary generation, the model (or a separate critique model) evaluates the output against explicit quality criteria and either approves it, requests a revision, or flags it for human review.

The pattern has multiple implementation variants. **Self-critique** uses a single model to evaluate its own output — effective and cheap, but limited because the model's blind spots are often consistent across generation and critique. **Cross-model critique** uses a different model (or a fine-tuned critic model) to evaluate the primary model's output — more robust but more expensive. **Constitutional critique** checks outputs against a defined rubric of rules (accuracy, completeness, tone, compliance) and generates structured feedback that drives targeted revisions.

For enterprise deployments, self-reflection is particularly valuable in content generation (ensuring regulatory compliance before publication), code generation (catching logical errors before execution), financial analysis (verifying calculations and assumptions), and customer communication (tone and accuracy review). The design tradeoff is cost: each reflection cycle adds at least one additional LLM call. Calibrate reflection depth to task criticality — use lightweight checks for routine outputs and deep multi-round reflection for high-stakes deliverables.