REAL: Resolving Knowledge Conflicts in Knowledge-Intensive Visual Question Answering via Reasoning-Pivot Alignment
Researchers have introduced REAL, a new framework designed to tackle knowledge conflicts in visual question answering systems. This approach utilizes a novel concept called a 'Reasoning-Pivot' to identify and resolve discrepancies arising from open-domain knowledge retrieval. The framework includes a training method for conflict detection and a decoding strategy to mitigate these issues, demonstrating significant performance improvements on various datasets. AI
IMPACT Introduces a novel method for improving the accuracy and reliability of AI systems that process visual and textual information with external knowledge.