2306.04528v4.pdf- “Understanding the Robustness of Large Language Models: PromptBench Analysis” – “Unveiling the Vulnerabilities of Large Language Models to Adversarial Prompts” – “Crafting More Robust Prompts: Insights from PromptBench Analysis” – “Examining the Transferability of Adversarial Prompts in Large Language Models” – “Enhancing Prompt Robustness: Practical Recommendations from PromptBench Study”

– Large Language Models (LLMs) have gained popularity in various tasks.
– LLMs consist of a prompt and (optionally) a sample for analysis.
– Existing work has evaluated the robustness of LLMs to adversarial samples.
– Some studies show that current LLMs are not robust to adversarial and out-of-distribution samples.
– Existing studies on robustness to adversarial samples may not be applicable in certain scenarios.
– A single prompt can be used for multiple samples in some tasks.

– The paper mentions Large Language Models (LLMs), not specifically GPTs.

– LLMs are not robust to adversarial prompts.
– Attention visualization helps analyze the reason behind the lack of robustness.
– Word frequency analysis provides guidance for prompt engineering.
– PromptBench is open-sourced for robust LLMs research.

– PromptBench is a benchmark to measure the robustness of Large Language Models (LLMs) to adversarial prompts.
– Adversarial prompts mimic user errors and evaluate the impact on LLM outcomes.
– LLMs are not robust to adversarial prompts.
– The study includes 4,788 adversarial prompts evaluated over 8 tasks and 13 datasets.
– The paper provides analysis and recommendations for prompt composition.

– LLMs are not robust to adversarial prompts.
– Attention visualization was used to analyze the reason behind this.
– Frequent words were analyzed to provide guidance for prompt engineering.
– PromptBench will be open-sourced for robust LLMs research.

– Full results of the three tasks: Trivia Creative Writing, Codenames Collaborative, and Logic Grid Puzzle can be found in Tables 5, 6, and 7, respectively.

– The researchers created a test to see how well language models handle different types of prompts.
– They made 4,788 prompts that were designed to trick the models.
– The models were tested on 8 different tasks, like understanding emotions and translating languages.
– The researchers found that the models were not very good at handling the tricky prompts.
– They also looked at why the models struggled and gave suggestions for making better prompts.