Research map / Large language models / Retrieval and knowledge
Faithful in Steps: Improving Generalization and Citation in RAG via Query Decomposition
Yue Liu, Zhongying Ru, Shimin Di, Jipeng Zhang, Ruiyuan Zhang, Xiaofang Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DPO alignment with retrieval feedback to QD-R (ablation) | QDRAG-R 7B · MMQA | citation (AutoAIS): 53.8 -> 57.5 | Improved |
| add DPO alignment with retrieval feedback to QD-R (ablation) | QDRAG-R 7B · WebQA | citation (AutoAIS): 56.8 -> 59.9 | Improved |
| add query decomposition (D, R, S) to context reranking (ablation) | QDRAG-C · MMQA | citation (AutoAIS): 48.8 -> 52.6 | Improved |
| add query decomposition (D, R, S) to context reranking (ablation) | QDRAG-C · WebQA | citation (AutoAIS): 41.9 -> 45.8 | Improved |
| use query decomposition instead of query rewriting (multi-round, with DPO) (ablation) | QDRAG-R · WebQA | citation (AutoAIS): 48.4 -> 59.9 | Improved |
| fine-tune on the QD task instead of the RAGQ task (ablation) | QDRAG-R · InfoSeek (out-of-domain) | citation (AutoAIS): 68.4 -> 72.9 | Improved |
| fine-tune on the QD task instead of the RAGQ task (ablation) | QDRAG-R · E-VQA (out-of-domain) | citation (AutoAIS): 26.0 -> 33.8 | Improved |
| fine-tune on RAGQ task instead of leaving the base model unfinetuned | Qwen2.5VL-7B RAG · InfoSeek (out-of-domain) | LLM-judge: 36.7 -> 32.9 | Hurt |
| replace CLIP retriever with MMEmbed-7B retriever | QDRAG-C · MMQA | citation (AutoAIS): 52.6 -> 74.7 | Improved |
| apply sparse context filtering (reranking) rather than dense RAG | Qwen2.5VL-7B RAG · multi-hop MMQA and WebQA | answer accuracy | Hurt |
| reduce distractor images in the citation context | MLLM image citation · CIRCO and CIRR | precision and recall | Improved |
| use QDRAG-C instead of SearChain for decomposition cost | query decomposition · Query decomposition cost | cost per query (USD): 0.070 -> 0.019 | Improved |
Closest papers by meaning
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.88)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.88)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.87)
- Explore What LLM Does Not Know in Complex Question Answering (AAAI 2025 · similarity 0.87)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.87)
- Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
