QDRAG
An attributable retrieval-augmented generation framework based on query decomposition. It decomposes complex questions into atomic sub-questions and applies reranking-based pruning to improve citation faithfulness and generalization capability.
Timeline (last 90 days)
A recent paper proposes the QDRAG framework, which systematically addresses RAGQ's over-citation, implicit entity recognition failure, and poor generalization problems through query decomposition.
The QDRAG framework contains two key steps: query decomposition and reranking-based pruning.
QDRAG uses a multimodal large language model (MLLM) to decompose the original question into several atomic sub-questions, each corresponding to an independent entity or relation.
QDRAG proposes two zero-shot query decomposition methods: QD-C (based on contextual knowledge) and QD-R (based on retrieval rewards).
The paper finds that fine-tuning on the query decomposition task exhibits better generalization than fine-tuning directly on the downstream RAGQ task.
QDRAG was evaluated on four multimodal question answering benchmarks, covering both in-domain and out-of-domain tests.
Experimental results show that QDRAG significantly outperforms all baselines on answer grounding and citation faithfulness, even surpassing Gemini-Pro.
QDRAG experiments show that it effectively reduces over-citation, improves the accuracy of implicit entity recognition, and maintains robust performance in cross-domain scenarios.
All Facts (8)
QDRAG experiments show that it effectively reduces over-citation, improves the accuracy of implicit entity recognition, and maintains robust performance in cross-domain scenarios.
50%UnverifiedThe QDRAG framework contains two key steps: query decomposition and reranking-based pruning.
50%UnverifiedQDRAG uses a multimodal large language model (MLLM) to decompose the original question into several atomic sub-questions, each corresponding to an independent entity or relation.
50%UnverifiedQDRAG proposes two zero-shot query decomposition methods: QD-C (based on contextual knowledge) and QD-R (based on retrieval rewards).
50%UnverifiedThe paper finds that fine-tuning on the query decomposition task exhibits better generalization than fine-tuning directly on the downstream RAGQ task.
50%UnverifiedQDRAG was evaluated on four multimodal question answering benchmarks, covering both in-domain and out-of-domain tests.
50%UnverifiedExperimental results show that QDRAG significantly outperforms all baselines on answer grounding and citation faithfulness, even surpassing Gemini-Pro.
50%UnverifiedA recent paper proposes the QDRAG framework, which systematically addresses RAGQ's over-citation, implicit entity recognition failure, and poor generalization problems through query decomposition.
50%