How do you design few-shot examples to prevent label bias?

Short answer: Label bias in few-shot prompting occurs when an unbalanced proportion or order of target labels in exemplars causes the model to favor specific outputs; it is mitigated by balancing class representations, randomizing exemplar ordering, and maintaining uniform length across examples.

Detail

In-context learning is sensitive to example selection:

Key Causes of Few-Shot Bias

1. Majority Class Bias: If 4 out of 5 exemplars demonstrate "Positive" sentiment, the model disproportionately predicts "Positive" for ambiguous inputs.
2. Recency Bias: The model tends to repeat the label shown in the final exemplar ($N$-th example) right before the user prompt.
3. Format Asymmetry: Detailed explanations in some exemplars and brief answers in others cause inconsistent output generation depth.

Biased Exemplars:   [Doc1 -> Positive], [Doc2 -> Positive], [Doc3 -> Positive] (Biases prediction to Positive)
Balanced Exemplars: [Doc1 -> Positive], [Doc2 -> Negative], [Doc3 -> Neutral]  (Unbiased decision boundary)

Example

Python helper to construct balanced few-shot prompts:

from collections import Counter

def format_balanced_few_shot(examples: list[dict], target_query: str) -> str:
    # Verify label distribution balance
    labels = [ex["label"] for ex in examples]
    counts = Counter(labels)
    print("Label distribution in exemplars:", counts)

    formatted = "Classify customer feedback into [Bug, Feature Request, Question].\n\n"
    for ex in examples:
        formatted += f"Text: \"{ex['text']}\"\nCategory: {ex['label']}\n\n"

    formatted += f"Text: \"{target_query}\"\nCategory:"
    return formatted

Interview tips

• Discuss how permutation of example order can alter accuracy by up to 20-30% on sensitive classification tasks.
• Mention using vector search to select dynamically relevant few-shot exemplars (k-NN exemplar selection) for each incoming user query.

⬅ Back to Prompt Engineering · All topics