Self-Refine | HyperKit.ai

Definition

Self-Refine is a technique where the model iteratively improves its output through a generate-critique-refine loop. After initial generation, the model critiques its own response, then produces an improved version based on that feedback.

This process can repeat multiple times until the output meets quality criteria or shows no further improvement.

Key Concepts

Generate: Produce initial output for the task
Critique: Identify weaknesses and areas for improvement
Refine: Create improved version addressing the critique
Iterate: Repeat until quality threshold met

Examples

Complete Loop

Self-Refine in Action

Task: "Write a professional email declining a meeting"

ITERATION 1:
Generate: "I can't come to the meeting. Too busy."

Critique: "Too abrupt and unprofessional. Missing:
- Polite greeting, reason for declining, alternative
  suggestions, professional closing."

Refine: "Hi, Thanks for the invite. Unfortunately I have
a conflict. Could we reschedule? Best, [Name]"

ITERATION 2:
Critique: "Better but still brief. Could be warmer,
specify willingness to find alternative time."

Refine: "Hi [Name], Thank you for thinking of me for this
meeting. Unfortunately, I have a prior commitment at that
time. I'd value the opportunity to connect—would any time
later this week work for you? Please let me know your
availability and I'll do my best to accommodate.
Best regards, [Your name]"

ITERATION 3:
Critique: "Professional, warm, and actionable. No major
improvements needed."

Final Output: [Iteration 2 response]

Implementation

Self-Refine Algorithm

def self_refine(task, model, max_iterations=3):
    # Initial generation
    output = model.generate(task)

    for i in range(max_iterations):
        # Critique current output
        critique = model.generate(f"""
            Critique this response for the task "{task}":
            Response: {output}

            List specific weaknesses and improvements needed.
        """)

        # Check if refinement needed
        if "no major improvements" in critique.lower():
            break

        # Refine based on critique
        output = model.generate(f"""
            Task: {task}
            Previous response: {output}
            Critique: {critique}

            Write an improved response addressing the critique.
        """)

    return output

Interactive Exercise

✎

Practice Self-Refine

Given this initial output and critique, write a refined version:

Task: "Explain APIs to a beginner"
Initial: "An API is an Application Programming Interface."
Critique: "Just states the acronym. Missing: analogy, practical example, why it matters."

Pro Tips

Set clear stopping criteria to avoid infinite loops
Include the original task in each iteration for context
Track improvements to detect when refinement plateaus
3-5 iterations usually sufficient for most tasks

Definition

Key Concepts

Examples

Interactive Exercise

Related Terms