Course contentsShow
← Machine Learning and Deep Learning
Lesson 1942 of 3,53841. Structured Output and Tool CallingFree lesson

Balancing Critique Specificity

Finding the right level of detail in critique to enable effective refinement without over-correction.

Balancing Critique Specificity

What You'll Learn

You'll learn how to calibrate the granularity of self-critique feedback so models improve effectively without getting stuck in over-correction loops or ignoring important issues.

The Goldilocks Problem of Critique

When a language model critiques its own output (or you design prompts for critique), you face a fundamental trade-off: too vague and the model can't act on the feedback; too specific and it may hyper-focus on minor details while missing broader issues—or make unnecessary changes that break what was already working.

Think of it like editing a draft essay. A critique saying "improve this" is useless. But marking every comma placement while ignoring structural problems wastes effort and creates churn without real improvement.

Finding the Sweet Spot

High-level critique works well for:

  • Early refinement iterations
  • Identifying strategic problems (logic, completeness, relevance)
  • Avoiding unnecessary token costs

Detailed critique works well for:

  • Later refinement stages after major issues are fixed
  • Format compliance and precise schema requirements
  • Edge cases and subtle errors

The optimal strategy often uses progressive specificity: start broad, then zoom in as major issues resolve. Your critique prompt might ask: "First identify any fundamental errors, then examine details only if the foundation is sound."

Practical Heuristics

  • Critique at the level where change is actionable
  • Avoid listing minor issues when major flaws exist
  • Match specificity to your stopping criteria (lesson 1941)
  • If refinement stalls or oscillates, your critique may be too granular

Key Takeaway: Effective self-critique balances breadth and depth—starting with strategic issues before drilling into details—preventing wasted refinement cycles while ensuring meaningful improvement at each iteration.