Announcing our new Paper: The Prompt Report, with Co-authors from OpenAI & Microsoft!
Check it out →
Learn Prompting🟢 Self-Calibration Prompting
Last updated on August 19, 2024 by Bhuwan Bhatt-
Importance of Self-Calibration: Self-calibration allows LLMs to evaluate their own answers, reducing misinformation.
-
Practical Implementation: You can use self-calibration by prompting the LLM to assess its own response for correctness.
-
Effectiveness in Larger Models: Larger and more complex models perform better at self-calibration, improving accuracy.
What is Self-Calibration Prompting?
If you ask any question to a Large Language Model (LLM), in most cases, it will likely generate an answer, be it correct or incorrect. This is undesirable as it can propagate incorrect information to the users. Recently, Air Canada had to compensate its customer for providing incorrect information regarding bereavement fares. Such actions can cost companies significant money and tarnish their reputation.
Self-Calibration prompting1 is a self-evaluation technique that asks the model to evaluate its output after generating it. Experiments show that, while challenging, models can self-evaluate their answers as either true or false.
How to Use Self-Calibration Prompting?
Employing self-calibration is a two-step process:
- Get the initial answer
- Ask the model whether the proposed answer is true or false
Let's see an example.
Step 1:
Ask the model, "Who is the first president of the United States?"
Step 2:
Ask the model to self-evaluate if the proposed answer is true or false.
You can also use few-shot prompting:
Step 1: Get the initial response
Step 2: Validate the response
What Are Self-Calibration Prompting Results?
- Model are apt to self-evaluate their own samples. In most cases, they can correctly identify whether their predictions are correct or incorrect.
Self-evaluation results for Lambada 52B1
- Larger models are better at self-calibration.
Larger models make fewer self-calibration errors1
Limitations of Self-Calibration
- The authors focus on pre-trained language models and exclude finetuned models. Hence, the technique may not work well for fine-tuned models.
Conclusion
The study clearly shows that LLMs are capable of evaluating their own response using a simple prompt. This can be used to minimize the number of false positive and false negative responses from the model. It also helps to establish confidence in the model.