Vision-Language Models (VLMs) are AI systems capable of interpreting and understanding visual and textual inputs. However, while these models have seen significant progress, their performance in challenging settings is still somewhat limited. One notable weakness lies in the core of VLMs, represented by Large Language Models (LLMs), which can occasionally deliver inaccurate or harmful content.
These potential vulnerabilities, such as the generation of discriminatory statements or inadvertent disclosure of personal information, underline the importance of thorough stress-testing, including ‘red teaming’ scenarios, in the deployment of VLMs. The lack of a comprehensive benchmark for red teaming VLMs led to the development of The Red Teaming Visual Language Model (RTVLM) dataset.
Consisting of ten subtasks grouped into four categories—faithfulness, privacy, safety, and fairness—the RTVLM dataset serves as the first-ever red teaming benchmark for current VLMs. The application of Supervised Fine-tuning (SFT) to the RTVLM dataset led to significant improvements in the performance of the tested VLM models, emphasizing the positive impact of red teaming alignment on the robustness of these systems.
On testing, the researchers found that all ten of the top open-source VLMs showcased challenges when exposed to red teaming, with performance differences reaching up to 31% when compared to GPT-4V. This outcome, combined with the lack of red teaming alignment in present VLMs, highlighted the need for additional refinements.
The introduction of the RTVLM dataset represents an important step towards ensuring the reliable performance of existing VLMs. By providing crucial insights about the current limitations of these models—alongside solid recommendations for improvements—the dataset supports the ongoing pursuit of enhanced VLM robustness.
The full research paper is available for review. Supporters of the research are encouraged to follow the team’s updates on Twitter, join their ML SubReddit, Facebook community, Discord Channel, and LinkedIn Group, and subscribe to their newsletter. The team also maintains a Telegram Channel for additional updates.
