Efficient LLM Auditing Using Fewer Than 20 Questions

By Muhammad OsamaReviewed by Joel ScanlonSep 19 2024

With just a handful of binary questions, researchers can now accurately distinguish AI models, a breakthrough that promises to revolutionize intellectual property audits and enhance AI transparency.

​​​​​​​Study: The 20 questions game to distinguish large language models. Image Credit: Krot_Studio / Shutterstock

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

In a recent article published on the arXiv preprint* server, researchers introduced a new approach for distinguishing large language models (LLMs) using a small set of benign binary questions in a black-box setting.

The method was designed to differentiate LLMs efficiently and accurately, achieving near 100% accuracy with fewer than 20 questions, depending on the approach used.

This technique is particularly useful for practical audits like detecting model leaks or assessing convergence, making it a valuable tool for professionals in the field.

Background

LLMs are widely employed for tasks such as language translation, text summarization, and chatbots. Trained on extensive text data, these models produce human-like responses to a variety of prompts.

However, their complexity and lack of transparency pose a challenge in determining if two LLMs are identical. This is a crucial aspect of auditing, especially in cases of model theft or intellectual property disputes, making the research highly relevant.

The challenge of differentiating LLMs is particularly pronounced when they provide identical responses to some prompts, even when they are not the same model. This differentiation is crucial for evaluating their accuracy on specific regulatory prompts.

Despite rapid advancements and state-of-the-art performance, the growing complexity of LLMs raises concerns about transparency. This lack of clarity makes understanding how LLMs make decisions, particularly concerning high-stakes applications, challenging.

About the Research

This paper proposed an innovative method to differentiate LLMs using fewer than 20 benign binary questions.

The authors formalized the problem using mathematical frameworks and established a baseline by randomly selecting questions from benchmark datasets, achieving nearly 100% accuracy with just 20 questions.

They introduced two effective questioning heuristics: the Separability Heuristic, which selects questions that maximize model separation, and the Recursive Similarity Heuristic, which constructs a sequence of questions that are as dissimilar as possible from previous ones.

Both heuristics were designed to optimize question selection, minimizing the number of questions needed for differentiation.

The approach is based on the idea that a good question should split the set of models into two equal groups, maximizing the number of differentiated pairs.

The Recursive Similarity Heuristic is particularly effective in ensuring that successive questions differ significantly, preventing redundant questioning. A query is considered optimal if it divides the models into two equal groups based on their answers.

The methodology involves several key steps. First, the researchers selected 22 LLMs and a set of binarized questions from HuggingFace, a popular natural language processing platform.

They then assess their heuristics' performance using a Monte Carlo approach to approximate the true negatives by sampling from the model distribution.

Key Findings

The outcomes showed that the proposed heuristics outperformed random question selection in distinguishing between LLMs.

The Separability Heuristic achieved an average accuracy of 95% with just 6 questions, while the Recursive Similarity Heuristic reached 95% accuracy with only 5 questions. The study also confirmed that the number of questions needed to differentiate models increases logarithmically with the number of models.

The study provided more insights into LLM behavior, showing that models' responses to questions were not independent. Some questions proved to be easier, answered correctly by most models, while others were significantly more difficult.

Additionally, their performance on these questions was linked to their performance on other tasks, such as language translation and text summarization.

Furthermore, the authors visualized the proximity of all 22 LLMs using a t-SNE plot, showing that models from the same family tend to cluster together, revealing similar behavior due to shared training data or architecture.

Their results suggest that the proposed heuristics can effectively distinguish between LLMs with high accuracy, even when the number of questions is limited.

Applications

This research has significant implications for auditing AI models, especially in model theft or intellectual property disputes.

The proposed heuristics can accurately distinguish between LLMs, even with a limited number of questions. They can also be used to check LLM convergence on specific regulatory prompts.

The findings have broader relevance for creating more transparent and understandable AI models. By differentiating LLMs with a small set of questions, the research provides valuable insights into model behavior, helping identify potential biases or performance gaps.

This method could also improve how AI model performance is evaluated, particularly in high-stakes applications such as legal or regulatory settings.

Conclusion

In summary, the novel approach proved effective for distinguishing LLMs using a small set of binary questions. The presented heuristics outperformed random selection, demonstrating high accuracy while reducing the number of questions.

This has important implications for auditing AI models, particularly in legal contexts involving model theft or intellectual property claims.  The results indicate that this method can improve AI model transparency and accountability.

Future work should focus on expanding the framework to a broader range of models and differentiating similar models, such as those with varying training parameters or datasets.

Additionally, exploring the method’s robustness with non-deterministic models would offer deeper insights into its effectiveness and limitations in real-world scenarios.

Overall, this study highlighted the importance of developing effective auditing techniques for the growing field of LLMs.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.