CEDA: Cross-modal Evaluation through Debate Agents for Robust Hallucination Detection скачать в хорошем качестве

CEDA: Cross-modal Evaluation through Debate Agents for Robust Hallucination Detection

The document introduces CEDA, a novel multimodal framework designed for robust hallucination detection in Large Language Model (LLM) outputs. It addresses the critical problem of LLMs generating factually incorrect or ungrounded content, especially severe in multimodal and high-stakes applications. CEDA employs a unique three-fold approach: a multi-agent debate system where agents critically examine and argue about the authenticity of generated content across modalities; a lightweight classifier head integrated with an LLM-as-a-judge for calibrated detection; and a confidence estimation mechanism to quantify uncertainty. This debate-based architecture allows for a more nuanced and contextual evaluation, offering improved generalizability, interpretability, and reliability compared to existing methods. Unlike prior work, CEDA uses dynamic prompting and cross-references information from various modalities, including images and text, for enhanced context. Extensive experiments on five benchmarks demonstrate CEDA's significant improvements over baseline hallucination detection methods. The framework also provides interpretable debate traces, enhancing the understanding of its reasoning. #CEDA #HallucinationDetection #MultiAgentDebate #MultimodalAI #LLMs #UncertaintyQuantification #AIResearch #amazon paper - https://www.amazon.science/publicatio... subscribe - https://t.me/arxivpaper donations: USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e created with NotebookLM