contrastive examples
**Contrastive examples** in prompt engineering is the technique of providing the language model with **both positive (correct) and negative (incorrect) demonstrations** — showing not just what good output looks like, but also what bad output looks like and why, enabling the model to learn sharper decision boundaries for the target task.
**Why Contrastive Examples Work**
- Standard few-shot prompting shows only positive examples — the model sees what to do, but not what to avoid.
- **Contrastive examples** add negative demonstrations — "here is a wrong answer and why it's wrong" — helping the model understand the **boundaries** between correct and incorrect responses.
- This is especially valuable for tasks with **subtle distinctions** where the model might otherwise confuse similar categories or make common errors.
**Contrastive Example Format**
```
Good example:
Input: "The battery lasts all day"
Label: Positive
Why: Describes a desirable product feature.
Bad example:
Input: "The battery lasts all day"
Label: Negative
Why WRONG: Despite mentioning "lasts," this is a
positive statement about battery life, not negative.
```
**When to Use Contrastive Examples**
- **Fine-Grained Classification**: Distinguishing between closely related categories — e.g., sarcasm vs. genuine praise, factual claims vs. opinions.
- **Error Correction**: When the model consistently makes a specific type of mistake — show the mistake explicitly and explain why it's wrong.
- **Boundary Cases**: Tasks with ambiguous edge cases — contrastive pairs on either side of the decision boundary help the model calibrate.
- **Style Requirements**: Show both the desired writing style AND common style mistakes to avoid.
**Contrastive Prompting Strategies**
- **Paired Examples**: For each positive example, provide a closely matched negative example — same topic or structure, but different correct label.
- **Near-Miss Examples**: Show examples that are almost correct but wrong in a specific way — teaches the model what subtle features matter.
- **Error Annotation**: Include an explanation of WHY the negative example is wrong — the reasoning helps the model internalize the distinction.
- **Before/After Pairs**: Show a bad output and its corrected version — teaches the model what transformations to apply.
**Benefits**
- **Accuracy**: Contrastive examples can improve classification accuracy by **5–15%** on difficult tasks compared to positive-only few-shot prompting.
- **Reduced Ambiguity**: Explicitly showing the boundary between categories reduces misclassification of edge cases.
- **Error Awareness**: The model learns to actively avoid common mistakes rather than just mimicking correct patterns.
**Practical Tips**
- Don't use too many negative examples — a ratio of 1 negative per 2–3 positive examples works well.
- Make negative examples **realistic** — they should represent actual mistakes the model might make, not obviously wrong cases.
- Always explain WHY the negative example is wrong — unexplained negatives can confuse the model.
Contrastive examples are a **high-impact prompt engineering technique** — by teaching the model what to avoid alongside what to produce, they create sharper, more discriminating few-shot learners.