Research map / Large language models / Retrieval and knowledge
MergePRAG: Orthogonal Merging of Passage-experts for Multi-hop Parametric RAG
Xuebing Liu, Shanbao Qiao, Roseline Nyange, Dongwook Min, Hyun Kim, Seung-Hoon Na
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 |
|---|---|---|---|
| use orthogonal (Gram-Schmidt) sequence merging instead of TIES merging (ablation) | MergePRAG+ |SP|=1 · HotpotQA | EM: 46.40 -> 48.80 | Improved |
| use orthogonal sequence merging instead of additive merging (ablation) | MergePRAG+ |SP|=1 · HotpotQA | EM: 36.20 -> 48.80 | Improved |
| use orthogonal sequence merging instead of arithmetic mean (ablation) | MergePRAG+ |SP|=1 · HotpotQA | EM: 48.20 -> 48.80 | No clear effect |
| use orthogonal sequence merging instead of concat merging (ablation) | MergePRAG+ |SP|=1 · HotpotQA | EM: 48.20 -> 48.80 | No clear effect |
| use orthogonal merging for inner and sequence merging instead of arithmetic mean (ablation) | MergePRAG+ |SP|=4 · HotpotQA | EM: 51.40 -> 52.40 | Improved |
| use hypernetwork parameterization (MergePRAG+) instead of finetuning [sq->sa] | MultihopRAG LLaMA3.1-8B · HotpotQA | EM: 43.70 -> 47.40 | Improved |
| finetune with gold passages instead of without passages | MultihopRAG LLaMA3.1-8B · HotpotQA | EM: 43.70 -> 40.10 | Hurt |
| add hypernetwork passage injection to MultihopRAG | MultihopRAG LLaMA3.1-8B · HotpotQA | EM: 37.80 -> 47.40 | Improved |
| add in-context passages to MergePRAG (MergePRAG+) (ablation) | MergePRAG LLaMA3.1-8B · HotpotQA | EM: 28.40 -> 48.80 | Improved |
| increase retrieved passages per hop from 1 to 4 (ablation) | MergePRAG+ · HotpotQA | EM: 48.80 -> 52.40 | Improved |
| increase number of key-value vectors k from 1 to 16 (ablation) | MergePRAG+ |SP|=4 · HotpotQA | EM: 49.00 -> 52.40 | Improved |
| compare MergePRAG+ against R3-RAG on a harder dataset | MergePRAG+ |SP|=4 LLaMA3.1-8B · MuSiQue (E5) | EM: 21.20 -> 16.70 | Hurt |
Closest papers by meaning
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
- Channel Merging: Preserving Specialization for Merged Experts (AAAI 2025 · similarity 0.88)
- Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.87)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.87)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 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.
