Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering (FiD, 2021)
This entry reviews the arXiv paper Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering by Gautier Izacard and Edouard Grave — the work behind the well-known Fusion-in-Decoder (FiD) architecture.
Metadata
| Field | Value | |-------|-------| | Title | Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering | | Authors | Gautier Izacard, Edouard Grave | | Source | arXiv:2007.01282 | | Category | Ranking for Search / RAG |
Key points
- Proposes an open-domain QA architecture that couples passage retrieval with a seq2seq generative reader, rather than relying on extractive span prediction alone.
- Retrieved passages (from Wikipedia, using DPR-style dense retrieval combined with BM25) are encoded independently, and their representations are fused in the decoder to condition answer generation — hence "Fusion-in-Decoder."
- Independent per-passage encoding makes the approach scalable to many passages (dozens to hundreds), since encoding can be batched and cached.
- Reported state-of-the-art results on Natural Questions and TriviaQA benchmarks at the time of publication.
- The paper sits at the intersection of retrieval-augmented generation (RAG) and large-scale search systems, addressing how to allocate responsibility between retrieval, reranking, and generation.
- Classical cascades (retrieve → rerank → generate) remain mainstream, but LLM-era and agentic paradigms increasingly treat when and how many times to retrieve as a learnable decision.
- Evaluation has shifted from static nDCG toward task success rate, citation accuracy, and multi-hop reasoning chain completeness.
- Latency & cost: control retrieval depth with early stopping, cache popular queries, and budget p99 latency.
- Quality: verify that offline gains translate to online user satisfaction via interleaving experiments and human audits.
- Safety: open retrieval can introduce poisoning or bias; consider source whitelisting and output filtering.
- Data hygiene: index versioning, PII scrubbing, and rollback-capable embedding versions.
- Experimental scale bounded by compute budgets; benchmark distributions may differ from real user traffic.
- English-centric data leaves cross-lingual generalization open.
- Future work includes better test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommender applications.
- Deep Learning to Rank in Industrial Search Engines, Recommender Systems
- Multi-Objective Recommendation in the Era of Generative AI: A Survey
- A Generative Re-ranking Model for List-level Multi-objective Optimization
- A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE
- Adaptive Neural Ranking Framework: Toward Maximized Business Goal
- Original paper: https://arxiv.org/abs/2007.01282
Method pipeline
1. Question encoding and retrieval: retrieve top-k relevant passages using a combination of sparse (BM25) and dense (dual-encoder) retrieval. 2. Per-passage encoding: each retrieved passage is concatenated with the question and encoded separately by the model encoder. 3. Fusion in decoder: all passage representations attend to each other within the decoder, enabling multi-hop information aggregation before autoregressive answer generation. 4. Answer generation: the model generates the answer as free text, supporting both extractive and abstractive answers.