A new research paper introduces the "Lost at the End" effect, demonstrating that multimodal retrieval-augmented question answering systems exhibit a primacy bias, unlike pure-text models which show a "lost-in-the-middle" effect. This means information presented at the beginning of retrieved passages is significantly more likely to be utilized by the system than information at the end. The study tested this on three open-source 7B/8B VLM readers and found that placing the correct answer at the start of the context improved performance by 16 to 26 points compared to placing it at the end. The researchers suggest that interventions targeting the reader model's prompt slot are necessary to address this bias, as retrieval-side fixes did not mitigate the issue. AI
IMPACT Highlights a significant bias in how multimodal AI systems process retrieved information, suggesting a need for reader-side interventions to improve performance.
RANK_REASON The cluster contains a research paper published on arXiv detailing a new finding about AI model behavior.
- 7B/8B VLM
- arXiv
- Knowledge-based visual question answering
- Liu et al. (2024)
- Lost at the End: Primacy Bias in Multimodal Retrieval-Augmented Question Answering
- Multimodal KB-VQA
- Wikipedia
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
