Predicting the Performance of Black-box LLMs through Self-Queries
This addresses the challenge of ensuring reliability in AI systems using black-box LLMs, with incremental improvements in interpretability and security.
The paper tackles the problem of predicting when black-box LLMs make mistakes by extracting features from follow-up prompts and response probabilities, enabling reliable instance-level performance prediction that often outperforms white-box methods and can detect adversarial influences or model misrepresentations.
As large language models (LLMs) are increasingly relied on in AI systems, predicting when they make mistakes is crucial. While a great deal of work in the field uses internal representations to interpret model behavior, these representations are inaccessible when given solely black-box access through an API. In this paper, we extract features of LLMs in a black-box manner by using follow-up prompts and taking the probabilities of different responses as representations to train reliable predictors of model behavior. We demonstrate that training a linear model on these low-dimensional representations produces reliable and generalizable predictors of model performance at the instance level (e.g., if a particular generation correctly answers a question). Remarkably, these can often outperform white-box linear predictors that operate over a model's hidden state or the full distribution over its vocabulary. In addition, we demonstrate that these extracted features can be used to evaluate more nuanced aspects of a language model's state. For instance, they can be used to distinguish between a clean version of GPT-4o-mini and a version that has been influenced via an adversarial system prompt that answers question-answering tasks incorrectly or introduces bugs into generated code. Furthermore, they can reliably distinguish between different model architectures and sizes, enabling the detection of misrepresented models provided through an API (e.g., identifying if GPT-3.5 is supplied instead of GPT-4o-mini).