Home AI News Expressing Calibrated Uncertainty: GPT-3 Model Learns to Verbalize Probability of Answers

Expressing Calibrated Uncertainty: GPT-3 Model Learns to Verbalize Probability of Answers

GPT-3 Model Expresses Uncertainty in Natural Language

A breakthrough has been made in AI technology with the development of the GPT-3 model that is capable of expressing uncertainty about its answers using natural language. This means that the model can not only provide an answer to a question but also convey a level of confidence in that answer, such as “90% confidence” or “high confidence.” These confidence levels are well calibrated, meaning they accurately reflect the probability of the answer being correct.

What sets the GPT-3 model apart is that it remains moderately calibrated even when faced with distribution shift, which refers to changes in the data it encounters during testing. This shows that the model is sensitive to uncertainty in its own answers and doesn’t solely rely on imitating human examples. In fact, this is the first time a model has demonstrated the ability to express calibrated uncertainty about its answers in natural language.

To ascertain the calibration of the model’s uncertainty, a suite of tasks called CalibratedMath was created for testing purposes. This suite compares the verbalized probability, which is uncertainty expressed in words, to uncertainty extracted from model logits. It has been found that both forms of uncertainty are capable of maintaining calibration even under distribution shift.

Furthermore, it has been discovered that GPT-3’s ability to generalize calibration is dependent on pre-trained latent representations. These representations are closely linked to epistemic uncertainty over the model’s answers. This correlation plays a crucial role in enabling the model to accurately express and maintain uncertainty.

In conclusion, the GPT-3 model’s proficiency in expressing calibrated uncertainty about its answers using natural language is groundbreaking. Its ability to remain moderately calibrated even under distribution shift and reliance on pre-trained latent representations gives it a significant advantage in the field of AI. This advancement opens up new possibilities for the development of AI systems that can effectively communicate uncertainty in a human-like manner.