Research map / Probabilistic and scientific modeling / Molecules and proteins
End-to-End Training of Multi-Document Reader and Retriever for Open-Domain Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train the multi-document reader and retriever end-to-end using the EMDR² objective instead of stage-wise FiD training | FiD with T5-base reader and DPR retriever · NQ dev | EM: 47.3 -> 50.4 | Helped |
| Train the multi-document reader and retriever end-to-end using the EMDR² objective instead of stage-wise FiD training | FiD with T5-base reader and DPR retriever · TriviaQA dev | EM: 65.5 -> 71.1 | Helped |
| Train the multi-document reader and retriever end-to-end using the EMDR² objective instead of stage-wise FiD training | FiD with T5-base reader and DPR retriever · WebQ dev | EM: 46.0 -> 49.9 | Helped |
| Use 50 retrieved documents rather than one (ablation) | FiD (DPR retriever, T5 reader) · NQ dev | EM: 37.3 -> 47.3 | Helped |
| Use 50 retrieved documents rather than one (ablation) | FiD (DPR retriever, T5 reader) · TriviaQA dev | EM: 50.8 -> 65.5 | Helped |
| Use 50 retrieved documents rather than one (ablation) | FiD (DPR retriever, T5 reader) · WebQ dev | EM: 40.2 -> 46.0 | Helped |
| Add supervised DPR training after unsupervised MSS retriever initialization (ablation) | EMDR² · NQ dev | EM: 50.4 -> 50.5 | No clear effect |
| Add supervised DPR training after unsupervised MSS retriever initialization (ablation) | EMDR² · TriviaQA dev | EM: 71.1 -> 71.2 | No clear effect |
| Add supervised DPR training after unsupervised MSS retriever initialization (ablation) | EMDR² · WebQ dev | EM: 49.9 -> 49.9 | No clear effect |
| Initialize the reader with MSS pre-training rather than T5 (ablation) | EMDR² with an MSS-initialized retriever · WebQ dev | EM: 47.7 -> 49.9 | Helped |
| Use the factorized retrieval-probability objective L_alt-1 instead of EMDR² (ablation) | End-to-end reader-retriever QA · NQ dev | EM: 50.4 -> 14.1 | Hurt |
| Use the factorized retrieval-probability objective L_alt-1 instead of EMDR² (ablation) | End-to-end reader-retriever QA · TriviaQA dev | EM: 71.1 -> 11.9 | Hurt |
Closest papers by meaning
- A Statistical Framework for Data-dependent Retrieval-Augmented Models (ICML 2024 · similarity 0.88)
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.87)
- Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.86)
- DSAS: A Universal Plug-and-Play Framework for Attention Optimization in Multi-Document Question Answering (NeurIPS 2025 · similarity 0.86)
- Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.86)
- UniGen: A Unified Generative Framework for Retrieval and Question Answering with Large Language Models (AAAI 2024 · similarity 0.86)
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.