Probabilities of Chat LLMs Are Miscalibrated but Still Predict Correctness on Multiple-Choice Q&A
This work addresses the problem of uncertainty estimation in LLMs for practitioners, showing that miscalibrated probabilities can still be useful for correctness prediction and selective abstention, though it is incremental in nature.
The study found that while the maximum softmax probabilities (MSPs) of 15 chat-fine-tuned LLMs are miscalibrated on multiple-choice Q&A, they still predict answer correctness, with wrong answers having smaller MSPs than correct ones for high-performing models, and demonstrated that using MSPs to abstain from uncertain answers improves performance with minimal labeled data.
We study 15 large language models (LLMs) fine-tuned for chat and find that their maximum softmax probabilities (MSPs) are consistently miscalibrated on multiple-choice Q&A. However, those MSPs might still encode useful uncertainty information. Specifically, we hypothesized that wrong answers would be associated with smaller MSPs compared to correct answers. Via rigorous statistical testing, we show that this hypothesis holds for models which perform well on the underlying Q&A task. We also find a strong direction correlation between Q&A accuracy and MSP correctness prediction, while finding no correlation between Q&A accuracy and calibration error. This suggests that within the current fine-tuning paradigm, we can expect correctness prediction but not calibration to improve as LLM capabilities progress. To demonstrate the utility of correctness prediction, we show that when models have the option to abstain, performance can be improved by selectively abstaining based on the MSP of the initial model response, using only a small amount of labeled data to choose the MSP threshold.