Self-Feedback | HyperKit.ai

Definition

Self-feedback is a technique where an LLM generates feedback on its own output, then uses that feedback to improve subsequent responses. This creates an internal improvement loop without requiring external evaluators.

Unlike self-critique (which identifies problems), self-feedback provides specific, actionable suggestions that guide refinement.

Key Concepts

Generation: Initial response to the task
Feedback: Self-generated critique with improvement suggestions
Refinement: Updated response incorporating feedback
Iteration: Multiple feedback-refinement cycles

Examples

Process

Self-Feedback Loop

SELF-FEEDBACK PROCESS:

TASK: "Write a function to check if a string is a palindrome"

STEP 1 - INITIAL GENERATION:
def is_palindrome(s):
    return s == s[::-1]

STEP 2 - SELF-FEEDBACK:
"Reviewing my solution:

Strengths:
+ Concise and readable
+ Correct basic logic

Weaknesses:
- Doesn't handle case sensitivity ('Racecar' fails)
- Doesn't ignore spaces/punctuation
- No input validation
- No docstring

Suggestions:
1. Normalize case with .lower()
2. Remove non-alphanumeric characters
3. Add type hints and docstring
4. Handle edge cases (empty string, None)"

STEP 3 - REFINED GENERATION:
def is_palindrome(s: str) -> bool:
    """
    Check if a string is a palindrome.

    Ignores case, spaces, and punctuation.

    Args:
        s: The string to check

    Returns:
        True if palindrome, False otherwise
    """
    if not s:
        return True

    # Normalize: lowercase, alphanumeric only
    cleaned = ''.join(c.lower() for c in s if c.isalnum())
    return cleaned == cleaned[::-1]

IMPROVEMENT: Much more robust and production-ready!

Implementation

Self-Feedback Prompting

SELF-FEEDBACK PROMPT TEMPLATE:

# Stage 1: Generate
generate_prompt = """
Task: {task}

Provide your best solution:
"""

# Stage 2: Self-Feedback
feedback_prompt = """
Task: {task}

Your solution:
{solution}

Provide detailed feedback on your solution:

1. STRENGTHS: What works well?

2. WEAKNESSES: What could be improved?

3. SPECIFIC SUGGESTIONS: List concrete changes
   (not vague - be specific!)

4. PRIORITY: Which improvement matters most?

Format your feedback as actionable items.
"""

# Stage 3: Refine
refine_prompt = """
Task: {task}

Original solution:
{solution}

Feedback received:
{feedback}

Now provide an improved solution that addresses
the feedback. Explain what you changed.
"""

# Implementation
def self_feedback_loop(task, iterations=2):
    solution = llm.generate(generate_prompt.format(task=task))

    for i in range(iterations):
        feedback = llm.generate(feedback_prompt.format(
            task=task, solution=solution
        ))
        solution = llm.generate(refine_prompt.format(
            task=task, solution=solution, feedback=feedback
        ))

    return solution

RESULTS:
┌────────────────────┬─────────┬──────────────┐
│ Task Type          │ Initial │ After S-F    │
├────────────────────┼─────────┼──────────────┤
│ Code generation    │ 72%     │ 85% (+13%)   │
│ Writing quality    │ 3.2/5   │ 4.1/5 (+0.9) │
│ Math word problems │ 68%     │ 76% (+8%)    │
└────────────────────┴─────────┴──────────────┘

Interactive Exercise

✎

Generate Self-Feedback

Initial solution for "Explain machine learning to a 10-year-old":

"Machine learning is when computers use algorithms to learn patterns from data and make predictions."

Provide self-feedback with specific improvement suggestions.

Pro Tips

Feedback should be specific and actionable, not vague
2-3 iterations usually sufficient; diminishing returns after
Different "personas" for feedback can catch different issues
Combine with external feedback for best results

Definition

Key Concepts

Examples

Interactive Exercise

Related Terms