Skip to content

Definition

A workflow where one LLM call generates a response and another provides evaluation and feedback in a loop, iteratively refining the output.

Key points

  • Analogous to a human writer iterating with editorial feedback.
  • When to use: when there are clear evaluation criteria AND "LLM responses can be demonstrably improved" through iterative refinement.

Why it matters for the exam

  • "Evaluator-optimizer" is a named workflow pattern; recognize the generator + evaluator loop and its precondition of clear, actionable evaluation criteria.

Common gotchas

  • Start simple; only add agentic complexity when it demonstrably improves outcomes — the loop only pays off when feedback measurably improves responses.

See also

Sources