This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Evaluating and Improving Robustness in Large Language Models: A Survey and Future Directions
Created by
Haebom
Author
Kun Zhang, Le Wu, Kui Yu, Guangyi Lv, Dacao Zhang
Outline
This paper presents a comprehensive survey of the robustness of large-scale language models (LLMs), which have been rapidly developing and being applied in various fields recently. The robustness of LLMs means that they ensure consistent, accurate, and stable content generation under unexpected application scenarios, such as malicious prompts, limited noisy domain data, and out-of-distribution (OOD) applications. In this paper, we present a formal definition of LLM robustness and systematically organize and review it from three perspectives: 1) Adversarial robustness: handling intentionally manipulated prompts (such as noisy prompts, long contexts, and data attacks); 2) OOD robustness: handling unexpected real-world application scenarios (such as OOD detection, zero-shot transfers, and hallucinations); and 3) Evaluation of robustness: summarizing novel evaluation datasets, metrics, and tools to verify the robustness of LLMs. Finally, we discuss future research directions and opportunities and provide a list of related studies and searchable projects ( https://github.com/zhangkunzk/Awesome-LLM-Robustness-papers) .
Takeaways, Limitations
•
Takeaways:
◦
We systematize the research field by presenting a comprehensive concept and methodology for the robustness of LLM.
◦
We comprehensively analyze LLM robustness research from three major perspectives: adversarial robustness, OOD robustness, and robustness evaluation.
◦
We summarize and introduce new evaluation datasets, metrics, and tools for LLM robustness research.
◦
It provided useful information to researchers by suggesting future research directions and providing a list of related studies and projects.
•
Limitations:
◦
This paper is a review of previous research and does not present new experimental results.
◦
It may not be possible to fully cover all aspects of LLM robustness. New research is constantly emerging, so it may not always reflect the most up-to-date information.
◦
The presented project requires ongoing maintenance and updates.
