Gabriel Alon

Gabriel Alon, Michael Kamfonas

A novel hack involving Large Language Models (LLMs) has emerged, exploiting adversarial suffixes to deceive models into generating perilous responses. Such jailbreaks can trick LLMs into providing intricate instructions to a malicious user for creating explosives, orchestrating a bank heist, or facilitating the creation of offensive content. By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. As we explored a broad range of regular (non-adversarial) prompt varieties, we concluded that false positives are a significant challenge for plain perplexity filtering. A Light-GBM trained on perplexity and token length resolved the false positives and correctly detected most adversarial attacks in the test set.

Michael Kamfonas, Gabriel Alon

Performance in natural language processing, and specifically for the question-answer task, is typically measured by comparing a modelś most confident (primary) prediction to golden answers (the ground truth). We are making the case that it is also useful to quantify how close a model came to predicting a correct answer even for examples that failed. We define the Golden Rank (GR) of an example as the rank of its most confident prediction that exactly matches a ground truth, and show why such a match always exists. For the 16 transformer models we analyzed, the majority of exactly matched golden answers in secondary prediction space hover very close to the top rank. We refer to secondary predictions as those ranking above 0 in descending confidence probability order. We demonstrate how the GR can be used to classify questions and visualize their spectrum of difficulty, from persistent near successes to persistent extreme failures. We derive a new aggregate statistic over entire test sets, named the Golden Rank Interpolated Median (GRIM) that quantifies the proximity of failed predictions to the top choice made by the model. To develop some intuition and explore the applicability of these metrics we use the Stanford Question Answering Dataset (SQuAD-2) and a few popular transformer models from the Hugging Face hub. We first demonstrate that the GRIM is not directly correlated with the F1 and exact match (EM) scores. We then calculate and visualize these scores for various transformer architectures, probe their applicability in error analysis by clustering failed predictions, and compare how they relate to other training diagnostics such as the EM and F1 scores. We finally suggest various research goals, such as broadening data collection for these metrics and their possible use in adversarial training.