Vroom-Vroom at SHROOM-Visions: A Multi-Judge Committee for Detecting Hallucinated Spans in Vision-Language Outputs

Problem

This work addresses the challenge of detecting and classifying hallucinated character spans in outputs generated by vision-language models. The authors highlight the need for robust methods to identify inaccuracies in these outputs, which can significantly impact the reliability of applications utilizing such models. The paper is a preprint and has not undergone peer review.

Method

The authors employ several fine-tuned vision-language models as independent annotators to assess the outputs. The core technical contribution is the implementation of a character-level majority voting mechanism to combine span predictions from these models. This approach allows for a more reliable detection of hallucinated spans by leveraging the collective judgment of multiple models. Additionally, the authors explore the use of activation probes to gain insights into the decision-making processes of the models.

Results

The proposed method ranks first in three out of four languages tested, outperforming all other submissions in those categories. Furthermore, it consistently achieves podium placements across all languages and metrics evaluated, indicating a strong performance relative to competing approaches.

Limitations

The authors do not report any limitations in their study, suggesting confidence in their methodology and results. However, the absence of reported limitations may also indicate a lack of comprehensive evaluation across diverse datasets or scenarios.

Why it matters

The implications of this work are significant for the field of vision-language models, particularly in enhancing the reliability of outputs in real-world applications. By improving the detection of hallucinated spans, this research contributes to the development of more trustworthy AI systems, which is crucial for applications in areas such as automated content generation, assistive technologies, and interactive AI systems.