{
  "version": 1,
  "date": "2026-09-22",
  "operator": "VibeResearch 产品维护者",
  "source_kind": "STRUCTURED",
  "model": "deepseek-flash",
  "template": "materials_v32",
  "sources": [
    {
      "id": "dpr",
      "title": "Dense Passage Retrieval for Open-Domain Question Answering",
      "authors": [
        "Vladimir Karpukhin",
        "Barlas Oguz",
        "Sewon Min",
        "Patrick Lewis",
        "Ledell Wu",
        "Sergey Edunov",
        "Danqi Chen",
        "Wen-tau Yih"
      ],
      "year": 2020,
      "doi": "10.18653/v1/2020.emnlp-main.550",
      "url": "https://aclanthology.org/2020.emnlp-main.550/",
      "pdf_url": "https://aclanthology.org/2020.emnlp-main.550.pdf",
      "version": "EMNLP 2020 proceedings, pages 6769–6781",
      "license": "CC-BY-4.0",
      "license_url": "https://creativecommons.org/licenses/by/4.0/",
      "rights_evidence": "https://aclanthology.org/faq/copyright/",
      "pdf_sha256": "2abfc7991519be2b8ce30215905516fdb593218b7bf59e50ece0d8a088a7c4d5",
      "structure_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
    },
    {
      "id": "fid",
      "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
      "authors": [
        "Gautier Izacard",
        "Edouard Grave"
      ],
      "year": 2021,
      "doi": "10.18653/v1/2021.eacl-main.74",
      "url": "https://aclanthology.org/2021.eacl-main.74/",
      "pdf_url": "https://aclanthology.org/2021.eacl-main.74.pdf",
      "version": "EACL 2021 proceedings, pages 874–880",
      "license": "CC-BY-4.0",
      "license_url": "https://creativecommons.org/licenses/by/4.0/",
      "rights_evidence": "https://aclanthology.org/faq/copyright/",
      "pdf_sha256": "a0b8708f35c78bc5799322a91e042f33233a5e48226c400843082c8a76b01175",
      "structure_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
    }
  ],
  "runs": {
    "native_transfers": 1,
    "initial_builds": 2,
    "failed_embedding_attempts_per_initial_build": 3,
    "maintenance_retries": 2,
    "reused_original_structure": true,
    "comparison_model_tasks": 1,
    "reader_model_tasks": 2
  },
  "comparison": {
    "question": "仅根据所选两篇论文的轻量结构化条目，比较研究目标、检索器或生成器所处环节、评测指标及其适用范围。每个事实给出具体条目引用。重点检查：DPR 的 top-k 段落检索准确率，能否与 FiD 的答案 exact match 直接比较？资料没有提供的数字写“未找到证据”；不要推断一种方法在所有场景全面更好。",
    "prompt_version": "structured-comparison-v3",
    "result": {
      "rows": [
        {
          "paper": "dpr",
          "sample": {
            "text": "DPR 管线：双编码器 BERT 检索器（离线 FAISS 索引 21,015,324 个 100 词段落，运行时取 top k），并含 BERT reader 的端到端 QA 评测；未涉及化学或物理材料。",
            "citations": [
              "P1.E2",
              "P1.E10",
              "P1.E35",
              "P1.E18"
            ]
          },
          "cells": [
            {
              "text": "提出并评测 Dense Passage Retriever（DPR）：两个独立 BERT（base, uncased）把问题与段落编码为 768 维向量、按点积检索，并与 BM25 比较 top-k 检索和端到端 QA。",
              "status": "REPORTED",
              "citations": [
                "P1.E2",
                "P1.E12",
                "P1.E66"
              ],
              "column_id": "c1"
            },
            {
              "text": "主体在检索环节：passage encoder 离线编码全部段落并用 FAISS 建索引，question encoder 运行时取 top k；另含两阶段管线中的 BERT reader 做答案抽取。",
              "status": "REPORTED",
              "citations": [
                "P1.E35",
                "P1.E18",
                "P1.E10"
              ],
              "column_id": "c2"
            },
            {
              "text": "Top-20/Top-100 段落检索准确率；端到端 exact match（EM）；检索吞吐（questions/s）；延迟与索引构建时间；消融准确率。",
              "status": "REPORTED",
              "citations": [
                "P1.E42",
                "P1.E29",
                "P1.E47"
              ],
              "column_id": "c3"
            },
            {
              "text": "Top-k 准确率 = top 20/100 检索段落中含答案的问题占比；EM = 归一化答案串与参考答案匹配，采用 Chen et al. 2017、Lee et al. 2019 的轻微归一化；吞吐与延迟在指定服务器/单张 32GB GPU 测量。",
              "status": "REPORTED",
              "citations": [
                "P1.E44",
                "P1.E76",
                "P1.E43"
              ],
              "column_id": "c4"
            },
            {
              "text": "五个 QA 数据集 NQ、TriviaQA、WQ、TREC、SQuAD v1.1；英文 Wikipedia 2018-12-20 转储，21,015,324 个 100 词段落；测试集报主结果，NQ dev 做消融；含 Single/Multi 训练。",
              "status": "REPORTED",
              "citations": [
                "P1.E3",
                "P1.E10",
                "P1.E50"
              ],
              "column_id": "c5"
            },
            {
              "text": "NQ top-20 78.4、top-100 85.4（BM25 59.1/73.7）；TriviaQA 79.4/85.0；WQ 73.2/81.4；TREC 79.8/89.1；SQuAD 63.2/77.2；端到端 EM（Single DPR，NQ）41.5。",
              "status": "REPORTED",
              "citations": [
                "P1.E47",
                "P1.E77"
              ],
              "column_id": "c6"
            },
            {
              "text": "同时报告检索级 top-k 准确率与端到端答案级 EM，并与 BM25、ORQA、REALM 等比较；适用对象是开放域 QA 的检索及端到端管线。",
              "status": "REPORTED",
              "citations": [
                "P1.E42",
                "P1.E47",
                "P1.E57"
              ],
              "column_id": "c7"
            }
          ]
        },
        {
          "paper": "fid",
          "sample": {
            "text": "FiD 管线：BM25 或 DPR 检索 100 词非重叠段落（检索 100 段、截断 250 word pieces），再经 T5 初始化的 Fusion-in-Decoder 生成答案；NQ/TriviaQA 用 DPR 检索，SQuAD 用 BM25。",
            "citations": [
              "P2.E2",
              "P2.E12",
              "P2.E13",
              "P2.E31"
            ]
          },
          "cells": [
            {
              "text": "研究检索增强生成式开放域 QA：把 BM25 或 DPR 检索到的段落交给 T5 初始化的 Fusion-in-Decoder 生成答案，考察 EM/F1 及随检索段落数、训练段落数的变化。",
              "status": "REPORTED",
              "citations": [
                "P2.E2",
                "P2.E12",
                "P2.E75"
              ],
              "column_id": "c1"
            },
            {
              "text": "主要在生成/读取环节：FiD 编码器独立编码每段+标题+问题（question:/title:/context:），解码器联合注意所有段落表示并生成答案；检索使用现成 BM25 或 DPR。",
              "status": "REPORTED",
              "citations": [
                "P2.E12",
                "P2.E59",
                "P2.E4"
              ],
              "column_id": "c2"
            },
            {
              "text": "Exact Match（EM）与 F1：NQ EM、TriviaQA open 与 hidden test EM、SQuAD Open EM/F1；另报告检索段落数、训练段落数和 GPU 小时。",
              "status": "REPORTED",
              "citations": [
                "P2.E48",
                "P2.E60",
                "P2.E77"
              ],
              "column_id": "c3"
            },
            {
              "text": "EM：生成答案经小写、去冠词/标点/重复空白等归一化后与任一可接受答案匹配；TriviaQA 大写答案用 Python title 方法归一化；SQuAD Open 另报 F1；按验证集 EM 每 500 步选最优 checkpoint。",
              "status": "REPORTED",
              "citations": [
                "P2.E50",
                "P2.E14",
                "P2.E61"
              ],
              "column_id": "c4"
            },
            {
              "text": "NaturalQuestions（open-domain 版本，丢弃答案超过 5 token 的样本）、TriviaQA（unfiltered）、SQuAD v1.1；NQ/TriviaQA 用 DPR 检索，SQuAD 用 BM25；检索 100 段、截断 250 word pieces、贪心解码。",
              "status": "REPORTED",
              "citations": [
                "P2.E31",
                "P2.E58",
                "P2.E13"
              ],
              "column_id": "c5"
            },
            {
              "text": "FiD large：NQ 51.4 EM；TriviaQA open/hidden 67.6/80.1 EM；SQuAD Open 56.7 EM/63.2 F1。FiD base：NQ 48.2 EM；TriviaQA 65.0/77.1 EM；SQuAD 53.4/60.6。检索段落 10→100：TriviaQA +6%、NQ +3.5%。",
              "status": "REPORTED",
              "citations": [
                "P2.E48",
                "P2.E32",
                "P2.E77"
              ],
              "column_id": "c6"
            },
            {
              "text": "所列条目报告端到端答案级 EM/F1，覆盖检索与生成整条管线。",
              "status": "REPORTED",
              "citations": [
                "P2.E48",
                "P2.E60"
              ],
              "column_id": "c7"
            }
          ]
        }
      ],
      "title": "DPR 与 FiD 的研究目标、检索/生成环节及评测指标可比性",
      "caveats": [
        {
          "text": "NQ 数据集版本可能不一致：FiD 明确用 open-domain 版本并丢弃答案超过 5 token 的样本；DPR 所列条目未说明其 NQ 是否使用同一过滤版本，因此 NQ EM 数字的可比条件不完整。",
          "citations": [
            "P2.E31",
            "P2.E58",
            "P1.E43",
            "P1.E50"
          ]
        },
        {
          "text": "FiD 的 EM 混合了检索环节：NQ/TriviaQA 用 DPR 检索、SQuAD 用 BM25；DPR 的 EM 用自身双编码器检索。两篇的 EM 不是同一检索器/生成器组合的对照。",
          "citations": [
            "P2.E19",
            "P2.E31",
            "P2.E59",
            "P1.E18",
            "P1.E35"
          ]
        },
        {
          "text": "结构化层限制：FiD 所列条目未单列其检索器的 top-k 段落检索准确率，DPR 所列条目未报告 FiD 生成器的 EM；两篇数字来自不同测量流程，不宜直接排名或推断一种方法在所有场景全面更好。",
          "citations": [
            "P2.E48",
            "P2.E60",
            "P1.E42",
            "P1.E47"
          ]
        }
      ],
      "columns": [
        {
          "id": "c1",
          "label": "研究目标"
        },
        {
          "id": "c2",
          "label": "系统所处环节"
        },
        {
          "id": "c3",
          "label": "主要评测指标"
        },
        {
          "id": "c4",
          "label": "指标定义与测量流程"
        },
        {
          "id": "c5",
          "label": "数据集与适用范围"
        },
        {
          "id": "c6",
          "label": "关键报告数值"
        },
        {
          "id": "c7",
          "label": "指标层级与适用对象"
        }
      ],
      "findings": [
        {
          "text": "研究目标侧重不同：DPR 训练并评测稠密检索器，主指标是 top-k 段落检索准确率，并用 BERT reader 顺带评测端到端 QA；FiD 则把检索结果交给 T5 生成器，主指标是答案级 EM/F1。",
          "citations": [
            "P1.E2",
            "P1.E12",
            "P1.E66",
            "P2.E2",
            "P2.E12",
            "P2.E75"
          ]
        },
        {
          "text": "所处环节不同：DPR 处于检索环节，离线用 passage encoder 与 FAISS 索引 21,015,324 个段落，运行时取 top k；FiD 处于生成环节，独立编码各段落、由解码器联合融合并生成答案。",
          "citations": [
            "P1.E35",
            "P1.E18",
            "P2.E12",
            "P2.E59",
            "P2.E76"
          ]
        },
        {
          "text": "DPR 的 top-k 段落检索准确率不能与 FiD 的答案 EM 直接比较：前者是 top 20/100 检索段落中含答案的问题占比，属检索级；后者是生成答案串与参考答案匹配，属答案级。",
          "citations": [
            "P1.E44",
            "P1.E76",
            "P2.E50",
            "P2.E14"
          ]
        },
        {
          "text": "即使同为 EM，DPR 端到端 EM 用 BERT reader 在 top 100 段中选段并抽答案（NQ 41.5），FiD EM 用 T5 FiD 生成（large NQ 51.4）；读取器、检索器、训练设置与段落数不同，不能归因于单一方法。",
          "citations": [
            "P1.E47",
            "P1.E81",
            "P2.E48",
            "P2.E85",
            "P2.E13"
          ]
        },
        {
          "text": "适用范围有限制：DPR 在 NQ、TriviaQA、WQ、TREC 的 top-20 高于 BM25，但 SQuAD top-20 为 63.2 低于 BM25 的 68.8；FiD 的 NQ/TriviaQA 报告随段落数增加提升，SQuAD Open 则用 BM25 检索。",
          "citations": [
            "P1.E47",
            "P1.E77",
            "P1.E71",
            "P2.E31",
            "P2.E77"
          ]
        }
      ]
    },
    "entries": [
      {
        "entry_id": "P1.E2",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Abstract",
        "field": "Abstract",
        "content": "This paper introduces Dense Passage Retriever (DPR), a dual-encoder framework that uses two independent BERT (base, uncased) networks to map questions and passages into 768-dimensional vectors and retrieves passages by dot-product similarity (Section 3.1). Trained on question-passage pairs using in-batch negatives plus one BM25 hard negative, DPR outperforms Lucene-BM25 by 9%–19% absolute in top-20 retrieval accuracy across open-domain QA datasets (Abstract; Section 5.1). On Natural Questions, DPR achieves 78.4% top-20 retrieval accuracy vs. 59.1% for BM25 and 65.2% vs. 42.9% top-5 accuracy (Introduction; Section 5.1; Table 2). End-to-end QA with a BERT reader reaches 41.5% exact match on NQ, compared with 33.3% for ORQA, and establishes state-of-the-art results on multiple benchmarks (Section 6.2; Table 4). The mechanism is that dense encodings capture lexical variations and semantic relationships that sparse term-matching misses, while FAISS MIPS indexing enables efficient retrieval (995.0 questions/s for top-100 vs. 23.7 questions/s per CPU thread for BM25/Lucene) (Introduction; Section 5.4).",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E3",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Materials and Formulation",
        "field": "Primary Material(s)",
        "content": "Not applicable (no physical material; data resources include English Wikipedia dump Dec. 20, 2018, 21,015,324 passages of 100 words, and QA datasets NQ, TriviaQA, WQ, TREC, SQuAD v1.1)",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E10",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Materials and Formulation",
        "field": "Details",
        "content": "The paper does not report material identities, ratios, concentrations, purity, or formulation steps. Computational/data resources: English Wikipedia dump from Dec. 20, 2018, preprocessed with DrQA code and split into 100-word passages, resulting in 21,015,324 passages; each passage prepended with the article title and [SEP] token (Section 4.1). QA datasets: NQ, TriviaQA, WQ, TREC, SQuAD v1.1 (Section 4.2; Table 1). Model components: two independent BERT (base, uncased) encoders producing [CLS] representations, d=768; dot-product similarity; FAISS index (Section 3.1). No chemical solvents, substrates, or surface treatments are involved.",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E12",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Main Method(s)",
        "content": "DPR training via dual-encoder metric learning; Wikipedia data preprocessing; positive/negative passage mining; FAISS indexing; BM25+DPR reranking; BERT reader training; joint retriever-reader ablation",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E18",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Details",
        "content": "Data preprocessing: English Wikipedia dump Dec. 20, 2018; DrQA preprocessing removes tables, infoboxes, lists, and disambiguation pages; split articles into disjoint 100-word passages, 21,015,324 total; prepend title + [SEP] (Section 4.1). Positive passage selection: for TREC, WQ, TriviaQA use highest-ranked BM25 passage containing the answer; discard question if none in top 100 retrieved passages; for SQuAD and NQ match/replace gold passage with corresponding passage in candidate pool; discard failed matches (Section 4.2). DPR training: dual-encoder metric learning; in-batch negatives with batch size 128 and one additional BM25 negative per question; up to 40 epochs for large datasets (NQ, TriviaQA, SQuAD) and 100 epochs for small datasets (TREC, WQ); learning rate 1e-5 using Adam, linear scheduling with warm-up, dropout 0.1 (Section 5). Multi-dataset encoder combines all datasets excluding SQuAD (Section 5). BM25+DPR reranking: obtain top-2000 passages from BM25 and DPR, rerank union with BM25(q,p)+λ·sim(q,p), λ=1.1 chosen on dev (Section 5). Reader training: sample one positive and m̃−1=23 negative passages from top 100 retrieved passages per question, m̃=24; batch size 16 large, 4 small; tune k on dev; experiments on eight 32GB GPUs (Section 6.1). Inference: passage encoder applied to all passages and indexed offline with FAISS; question encoder at run-time; retrieve top k with closest embeddings (Section 3.1). Joint-training variant: fix passage encoder, update question encoder; top 100 passages per question; batch 16; effectively 1,600 passages per question for retriever loss; reader uses 24 passages selected from top 5 positive and top 30 negative from top 100 (Appendix D). Index build: embedding 21M passages ~8.8 hours on 8 GPUs; FAISS index build on 21M vectors 8.5 hours on single server; Lucene inverted index ~30 minutes (Section 5.4).",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E29",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Structure and Characterization",
        "field": "Results",
        "content": "DPR top-20 NQ 78.4 vs BM25 59.1; top-100 NQ 85.4 vs BM25 73.7; in-batch negatives improve over 1-of-N; one BM25 hard negative improves; dot product and L2 comparable, cosine inferior; triplet loss similar; cross-dataset loss 3–5 points; DPR 995.0 q/s vs BM25 23.7 q/s; index build 8.8 h/8 GPUs and 8.5 h FAISS; Lucene ~30 min; qualitative examples show complementary strengths",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E35",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Device Fabrication",
        "field": "Stack/Assembly",
        "content": "Two-stage framework: (1) retriever selects candidate passages; (2) BERT reader examines top k passages, assigns passage selection score, extracts answer span; DPR stack: passage encoder E_P indexes all passages with FAISS offline, question encoder E_Q maps question at run-time, dot-product similarity retrieves top k; reader uses BERT (base, uncased) and top k retrieved passages (Sections 1, 3.1, 6.1)",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E42",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Metrics",
        "content": "Top-20/Top-100 passage retrieval accuracy; end-to-end exact match (EM); retrieval throughput (questions/s); latency; index build time; ablation accuracies",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E43",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Test Conditions",
        "content": "Five QA datasets (NQ, TriviaQA, WQ, TREC, SQuAD); test sets for main results; NQ development set for ablations; top k = 20/100; retrieval top 100 for reader; k=50 optimal for NQ; server Intel Xeon CPU E5-2698 v4 @ 2.20GHz and 512GB memory for speed; eight 32GB GPUs for training; exact match after minor normalization as in Chen et al. 2017 and Lee et al. 2019",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E44",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Data Reduction",
        "content": "Top-k accuracy: fraction of questions for which top-k retrieved passages contain an answer span; exact match: normalized answer string matches reference; run-time measured on specified server; latency approximately 20 ms for 100 passages on single 32GB GPU",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E47",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Details",
        "content": "Retrieval metrics: top-k accuracy (k=20,100) = percentage of top 20/100 retrieved passages that contain the answer (Table 2). Test sets: NQ, TriviaQA, WQ, TREC, SQuAD. Main results: DPR vs BM25 (Single): NQ top-20 78.4 vs 59.1; top-100 85.4 vs 73.7; TriviaQA 79.4 vs 66.9 and 85.0 vs 76.7; WQ 73.2 vs 55.0 and 81.4 vs 71.1; TREC 79.8 vs 70.9 and 89.1 vs 84.1; SQuAD 63.2 vs 68.8 and 77.2 vs 80.0 (Table 2). Multi DPR: NQ 79.4/86.0; TriviaQA 78.8/84.7; WQ 75.0/82.9; TREC 89.1/93.9; SQuAD 51.6/67.6. BM25+DPR sometimes improves (e.g., TREC Single top-20 85.2; Multi top-100 94.1). End-to-end QA EM: DPR Single NQ 41.5, TriviaQA 56.8, WQ 34.6, TREC 25.9, SQuAD 29.8; Multi DPR NQ 41.5, TriviaQA 56.8, WQ 42.4, TREC 49.4, SQuAD 24.1; BM25+DPR Multi NQ 38.8, TriviaQA 57.9, WQ 41.1, TREC 50.6, SQuAD 35.8 (Table 4). ORQA NQ 33.3, TriviaQA 45.0, WQ 36.4, TREC 30.1, SQuAD 20.2; REALM_Wiki NQ 39.2, WQ 40.2, TREC 46.8; REALM_News NQ 40.4, WQ 40.7, TREC 42.9 (Table 4). Run-time: DPR 995.0 q/s top 100 vs BM25/Lucene 23.7 q/s per CPU thread; embedding 21M passages ~8.8 h on 8 GPUs; FAISS index 8.5 h single server; Lucene ~30 min (Section 5.4). Reader latency ~20 ms for top 100 passages on single 32GB GPU; k=50 optimal for NQ, k=10 only marginal loss 40.8 vs 41.5 EM (Section 6.2). Ablations: sample efficiency 1k examples already beats BM25; in-batch negatives; one BM25 negative best; two does not help further; gold vs distant supervision 1 point lower; L2 comparable, cosine inferior; triplet loss similar; cross-dataset 3-5 points loss (Sections 5.2, Appendix A/B, Tables 3,5,6, Figure 1). Joint training NQ 39.8 EM (Appendix D).",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E50",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Data",
        "content": "English Wikipedia dump Dec. 20, 2018: 21,015,324 passages of 100 words, title + [SEP]; QA datasets NQ, TriviaQA, WQ, TREC, SQuAD v1.1. Train/dev/test counts: NQ 79,168/8,757/3,610; TriviaQA 78,785/8,837/11,313; WQ 3,417/361/2,032; TREC 1,353/133/694; SQuAD 78,713/8,886/10,570 (Table 1). Positive passage selection: highest BM25 passage containing answer for TREC/WQ/TriviaQA; gold match for SQuAD/NQ; discard if no answer in top 100. Multi-dataset training excludes SQuAD. Sample efficiency tested from 1k to 59k examples.",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E57",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Comparative study",
        "content": "The paper explicitly compares DPR with BM25, BM25+DPR, ORQA, REALM, and other QA systems, and ablates training schemes.",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E66",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Claim",
        "content": "Dense retrieval with a properly trained dual encoder can outperform sparse BM25 for open-domain QA; in-batch negatives and one BM25 hard negative are critical; better retrieval precision improves QA; dense retrieval captures semantic/lexical variation but may miss rare salient phrases.",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E71",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Details",
        "content": "Claims: (1) Dense retrieval can outperform TF-IDF/BM25 for open-domain QA when trained properly; evidence: DPR beats BM25 on most datasets (Table 2), with 9%-19% absolute top-20 gain (Abstract). (2) In-batch negatives are a key training mechanism; they increase effective training pairs and improve accuracy; evidence: Table 3. (3) One BM25 hard negative improves performance substantially; two do not help further; evidence: Table 3. (4) Dot product/L2 are comparable; cosine inferior; triplet loss similar; evidence: Table 6, Appendix B. (5) DPR generalizes cross-dataset with 3-5 point loss but still beats BM25; evidence: Section 5.2. (6) Higher retrieval precision translates to higher end-to-end QA accuracy; evidence: Table 4, Section 6.2. (7) Dense retrieval captures semantic/lexical variation but can miss rare salient phrases; evidence: Table 7, Section 5.3. Limitations: SQuAD is an exception due to annotator lexical overlap and biased training distribution (Section 5.1); dense index build is expensive (Section 5.4); no additional pretraining needed (Abstract, Section 6.2).",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E76",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Tables",
        "field": "Tables",
        "content": "Table 2: Top-20 & Top-100 retrieval accuracy on test sets, measured as the percentage of top 20/100 retrieved passages that contain the answer. Single and Multi denote that our Dense Passage Retriever (DPR) was trained using individual or combined training datasets (all the datasets excluding SQuAD).",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E77",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Tables",
        "field": "Tables",
        "content": "| Training | Retriever | Top-20 NQ | Top-20 TriviaQA | Top-20 WQ | Top-20 TREC | Top-20 SQuAD | Top-100 NQ | Top-100 TriviaQA | Top-100 WQ | Top-100 TREC | Top-100 SQuAD |\n|----------|-----------|-----------|-----------------|-----------|-------------|--------------|------------|------------------|------------|--------------|---------------|\n| None | BM25 | 59.1 | 66.9 | 55.0 | 70.9 | 68.8 | 73.7 | 76.7 | 71.1 | 84.1 | 80.0 |\n| Single | DPR | 78.4 | 79.4 | 73.2 | 79.8 | 63.2 | 85.4 | 85.0 | 81.4 | 89.1 | 77.2 |\n| Single | BM25+DPR | 76.6 | 79.8 | 71.0 | 85.2 | 71.5 | 83.8 | 84.5 | 80.5 | 92.7 | 81.3 |\n| Multi | DPR | 79.4 | 78.8 | 75.0 | 89.1 | 51.6 | 86.0 | 84.7 | 82.9 | 93.9 | 67.6 |\n| Multi | BM25+DPR | 78.0 | 79.9 | 74.7 | 88.5 | 66.2 | 83.9 | 84.4 | 82.3 | 94.1 | 78.6 |",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P1.E81",
        "title": "Dense Passage Retrieval for Open-Domain Question Answering",
        "section": "Tables",
        "field": "Tables",
        "content": "| Training | Model | NQ | TriviaQA | WQ | TREC | SQuAD |\n|----------|-------|----|----------|----|------|-------|\n| Single | BM25+BERT (Lee et al., 2019) | 26.5 | 47.1 | 17.7 | 21.3 | 33.2 |\n| Single | ORQA (Lee et al., 2019) | 33.3 | 45.0 | 36.4 | 30.1 | 20.2 |\n| Single | HardEM (Min et al., 2019a) | 28.1 | 50.9 | - | - | - |\n| Single | GraphRetriever (Min et al., 2019b) | 34.5 | 56.0 | 36.4 | - | - |\n| Single | PathRetriever (Asai et al., 2020) | 32.6 | - | - | - | 56.5 |\n| Single | REALM_Wiki (Guu et al., 2020) | 39.2 | - | 40.2 | 46.8 | - |\n| Single | REALM_News (Guu et al., 2020) | 40.4 | - | 40.7 | 42.9 | - |\n| Single | BM25 | 32.6 | 52.4 | 29.9 | 24.9 | 38.1 |\n| Single | DPR | 41.5 | 56.8 | 34.6 | 25.9 | 29.8 |\n| Single | BM25+DPR | 39.0 | 57.0 | 35.2 | 28.0 | 36.7 |\n| Multi | DPR | 41.5 | 56.8 | 42.4 | 49.4 | 24.1 |\n| Multi | BM25+DPR | 38.8 | 57.9 | 41.1 | 50.6 | 35.8 |",
        "content_sha256": "03257efefa034007f813e64a968324ab43c5981828647bbc3fe2e6521d1f2b5b"
      },
      {
        "entry_id": "P2.E2",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Abstract",
        "field": "Abstract",
        "content": "This paper investigates retrieval-augmented generative models for open-domain question answering, combining BM25 or DPR passage retrieval with a sequence-to-sequence Fusion-in-Decoder model initialized from T5. The encoder processes each retrieved passage independently with the question and special tokens, and the decoder attends over the concatenated encoder representations to generate the answer. Fusion-in-Decoder (large) obtains 51.4 EM on NaturalQuestions, 67.6/80.1 EM on TriviaQA open/hidden test sets, and 56.7 EM/63.2 F1 on SQuAD Open (Table 1). Increasing retrieved passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5%, which the authors interpret as evidence that sequence-to-sequence models efficiently aggregate and combine evidence from multiple passages (Figure 3; Section 4). Training with fewer passages decreases accuracy, but finetuning with 100 passages for 1000 steps reaches 46.0 EM on NaturalQuestions using 147 GPU hours versus 425 GPU hours when training on 100 passages (Table 2; Section 4).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E4",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Materials and Formulation",
        "field": "Components/Additives",
        "content": "Retrieval methods: BM25 (Apache Lucene with default parameters; SpaCy tokenization) and DPR (two BERT networks; dot product ranking; FAISS approximate nearest neighbors). Generative model: T5-based Fusion-in-Decoder with encoder-decoder evidence fusion. Evidence: Section 3 Method (Retrieval, Reading).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E12",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Main Method(s)",
        "content": "Two-step open-domain QA: (1) passage retrieval with BM25 or DPR; (2) generative Fusion-in-Decoder model. Fusion-in-Decoder processes each retrieved passage and its title concatenated with the question independently in the encoder, adds special tokens `question:`, `title:`, `context:`, and the decoder attends over the concatenation of all retrieved-passage representations. Evidence: Section 3 Method (Retrieval, Reading); Figure 2.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E13",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Parameters",
        "content": "T5 base = 220M parameters; T5 large = 770M parameters. Fine-tuned per dataset with Adam, constant learning rate 1e-4, dropout 10%, 10k gradient steps, batch size 64, using 64 Tesla V100 32Gb. Models evaluated every 500 steps; best selected on validation Exact Match. Retrieve 100 passages unless stated otherwise; truncate passages to 250 word pieces. Passages are 100 words and do not overlap. For NQ and TriviaQA retrieval uses DPR; for SQuAD retrieval uses BM25. Greedy decoding for answer generation. Evidence: Section 3 Method; Section 4 Experiments (Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E14",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Post-Processing",
        "content": "Exact Match normalization: lowercasing and removing articles, punctuation, and duplicated whitespace. TriviaQA uppercase answers normalized by converting all letters to lowercase except the first letter of each word, using the Python `title` string method. During training on NaturalQuestions and SQuAD, target is sampled among the list of answers; for TriviaQA, the unique human-generated answer is used. Evidence: Section 4 Experiments (Evaluation, Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E19",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Comparative study",
        "content": "The paper compares model sizes, retrieval methods, and training-passage counts. Key comparisons:\n- T5 base vs. T5 large: FiD base obtains NQ 48.2 EM; FiD large obtains NQ 51.4 EM (Table 1).\n- Training passages 5, 10, 25, 50, 100: reducing training passages decreases accuracy; finetuning with 100 passages for 1000 steps reduces the gap (Table 2).\n- DPR vs. BM25: DPR is used for NQ/TriviaQA; BM25 is used for SQuAD, following Karpukhin et al. (2020) (Section 4 Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E31",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure and Characterization",
        "field": "Measurement Conditions",
        "content": "Datasets: NaturalQuestions (open-domain version discarding answers with more than 5 tokens), TriviaQA (unfiltered), SQuAD v1.1. Wikipedia dumps: Dec. 20, 2018 for NQ/TriviaQA; Dec. 21, 2016 for SQuAD. Following Lee et al. (2019), validation is used as test and 10% of training is kept for validation. Passages are 100 words, non-overlapping. Retrieve 100 passages unless stated otherwise and truncate to 250 word pieces. NQ/TriviaQA use DPR; SQuAD uses BM25. Greedy decoding. TriviaQA results reported on open-domain test set and hidden test set. Evidence: Section 4 Experiments (Datasets, Technical details); Table 1 caption.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E32",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure and Characterization",
        "field": "Results",
        "content": "FiD base: NQ EM 48.2; TriviaQA open EM 65.0; TriviaQA hidden EM 77.1; SQuAD Open EM 53.4, F1 60.6. FiD large: NQ EM 51.4; TriviaQA open EM 67.6; TriviaQA hidden EM 80.1; SQuAD Open EM 56.7, F1 63.2. Increasing passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5%. Training with fewer passages reduces accuracy; finetuning with 100 passages for 1000 steps reaches 46.0 EM on NaturalQuestions with 147 GPU hours vs. 425 GPU hours when training on 100 passages. Evidence: Table 1; Table 2; Figure 3; Section 4.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E48",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Metrics",
        "content": "Exact Match (EM) and F1. NaturalQuestions EM; TriviaQA EM on open-domain and hidden test sets; SQuAD Open EM and F1. FiD base: NQ 48.2 EM; TriviaQA 65.0/77.1 EM; SQuAD Open 53.4 EM/60.6 F1. FiD large: NQ 51.4 EM; TriviaQA 67.6/80.1 EM; SQuAD Open 56.7 EM/63.2 F1. Evidence: Table 1.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E50",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Data Reduction",
        "content": "Exact Match: generated answer is correct if it matches any acceptable answer after normalization, which consists of lowercasing and removing articles, punctuation, and duplicated whitespace. F1 is also used for SQuAD Open. TriviaQA uppercase answers normalized by converting all letters to lowercase except the first letter of each word using the Python `title` string method. Best checkpoint selected on validation Exact Match. Evidence: Section 4 Experiments (Evaluation, Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E58",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Data",
        "content": "NaturalQuestions (open-domain version obtained by discarding answers with more than 5 tokens); TriviaQA (unfiltered version); SQuAD v1.1. Wikipedia dumps: Dec. 20, 2018 for NQ/TriviaQA; Dec. 21, 2016 for SQuAD. Following Lee et al. (2019), validation is used as test and 10% of training set is kept for validation. Preprocessing follows Chen et al. (2017) and Karpukhin et al. (2020), leading to 100-word non-overlapping passages. Retrieve 100 passages unless stated otherwise and truncate to 250 word pieces. Evidence: Section 4 Experiments (Datasets, Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E59",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Model",
        "content": "T5-based Fusion-in-Decoder. Two sizes: base 220M parameters and large 770M parameters, initialized from pretrained T5 models available in HuggingFace Transformers. Encoder processes each passage and title concatenated with the question independently, with special tokens `question:`, `title:`, `context:`; decoder attends over the concatenation of all retrieved-passage representations. Retrieval: BM25 (Apache Lucene default parameters; SpaCy tokenization) or DPR (two BERT networks; dot product ranking; FAISS approximate nearest neighbors). DPR used for NQ/TriviaQA; BM25 used for SQuAD. Evidence: Section 3 Method; Section 4 Experiments (Technical details).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E60",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Evaluation",
        "content": "Exact Match (EM) and F1. Best checkpoint selected on validation Exact Match every 500 steps. NaturalQuestions EM; TriviaQA EM on open-domain and hidden test sets; SQuAD Open EM/F1. FiD large: NQ 51.4 EM; TriviaQA 67.6/80.1 EM; SQuAD Open 56.7 EM/63.2 F1. FiD base: NQ 48.2 EM; TriviaQA 65.0/77.1 EM; SQuAD Open 53.4 EM/60.6 F1. Training-passage ablation and passage-scaling study also reported. Evidence: Table 1; Table 2; Figure 3; Section 4 Experiments.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E61",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Machine Learning and Data-Driven Study",
        "field": "Decision/UQ",
        "content": "No uncertainty quantification, Bayesian optimization, or active learning is described. Model selection uses validation Exact Match; training uses Adam with constant learning rate 1e-4, dropout 10%, 10k gradient steps, batch size 64, 64 Tesla V100 32Gb; greedy decoding at inference. Finetuning with 100 passages for 1000 steps is used to reduce the accuracy gap with lower compute. Evidence: Section 4 Experiments (Technical details, Impact of the number of training passages).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E75",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Claim",
        "content": "Retrieval-augmented generative QA with Fusion-in-Decoder benefits from retrieving many passages and scales with the number of retrieved passages due to decoder-side evidence fusion.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E76",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Evidence Chain",
        "content": "Retrieve passages with BM25 or DPR → encode each passage and title concatenated with the question independently in the encoder → decoder attends over the concatenation of all passage representations → generate answer; independent encoding keeps computation linear in the number of passages, while joint decoding aggregates evidence from multiple passages. Evidence: Section 3 Method (Reading); Figure 2.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E77",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Quantitative Support",
        "content": "FiD large: NQ 51.4 EM; TriviaQA open/hidden 67.6/80.1 EM; SQuAD Open 56.7 EM/63.2 F1 (Table 1). Increasing retrieved passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5% (Figure 3; Section 4 Scaling with number of passages). Training with fewer passages decreases accuracy; finetuning with 100 passages for 1000 steps reaches 46.0 EM on NQ with 147 GPU hours vs. 425 GPU hours when training on 100 passages (Table 2; Section 4 Impact of the number of training passages).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "P2.E85",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Tables",
        "field": "Tables",
        "content": "| Model | NQ EM | TriviaQA EM (open domain test set) | TriviaQA EM (hidden test set) | SQuAD Open EM | SQuAD Open F1 |\n|-------|-------|-----------------------------------|-------------------------------|---------------|---------------|\n| DrQA (Chen et al., 2017) | Not specified | Not specified | Not specified | 29.8 | Not specified |\n| Multi-Passage BERT (Wang et al., 2019) | Not specified | Not specified | Not specified | 53.0 | 60.9 |\n| Path Retriever (Asai et al., 2020) | 31.7 | Not specified | Not specified | 56.5 | 63.8 |\n| Graph Retriever (Min et al., 2019b) | 34.7 | 55.8 | Not specified | Not specified | Not specified |\n| Hard EM (Min et al., 2019a) | 28.8 | 50.9 | Not specified | Not specified | Not specified |\n| ORQA (Lee et al., 2019) | 31.3 | 45.1 | Not specified | 20.2 | Not specified |\n| REALM (Guu et al., 2020) | 40.4 | Not specified | Not specified | Not specified | Not specified |\n| DPR (Karpukhin et al., 2020) | 41.5 | 57.9 | Not specified | 36.7 | Not specified |\n| SpanSeqGen (Min et al., 2020) | 42.5 | Not specified | Not specified | Not specified | Not specified |\n| RAG (Lewis et al., 2020b) | 44.5 | 56.1 | 68.0 | Not specified | Not specified |\n| T5 (Roberts et al., 2020) | 36.6 | Not specified | 60.5 | Not specified | Not specified |\n| GPT-3 few shot (Brown et al., 2020) | 29.9 | Not specified | 71.2 | Not specified | Not specified |\n| Fusion-in-Decoder (base) | 48.2 | 65.0 | 77.1 | 53.4 | 60.6 |\n| Fusion-in-Decoder (large) | 51.4 | 67.6 | 80.1 | 56.7 | 63.2 |",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      }
    ]
  },
  "reader": {
    "question": "只依据当前论文的轻量结构化内容，用不超过250字说明方法怎样使用多个检索段落，以及实验结果的适用边界。每个事实附具体条目引用。如果现有条目不足以回答，就明确说明。能否据此断言所有任务中检索段落越多越好？请区分论文报告和你的推断。",
    "answer": "方法：BM25/DPR检索段落[Rb7f8295ef5d44c93.E12]；FiD各段独立编码，解码器对所有段落表示拼接注意力后生成答案[Rb7f8295ef5d44c93.E12]。独立编码计算随段数线性增长，联合解码融合多段证据[Rb7f8295ef5d44c93.E76][Rb7f8295ef5d44c93.E21]。\n\n适用边界：结果限于NQ、TriviaQA、SQuAD Open；默认100段，NQ/TriviaQA用DPR、SQuAD用BM25，截断250词片，贪心解码[Rb7f8295ef5d44c93.E31][Rb7f8295ef5d44c93.E49]。论文报告10→100段使TriviaQA+6%、NQ+3.5%，多数抽取模型约10–20段达峰[Rb7f8295ef5d44c93.E32][Rb7f8295ef5d44c93.E56]；当前资料未提供SQuAD缩放或>100段趋势[Rb7f8295ef5d44c93.E31][Rb7f8295ef5d44c93.E32]。\n\n不能断言所有任务中越多越好。论文报告仅NQ/TriviaQA增益[Rb7f8295ef5d44c93.E2][Rb7f8295ef5d44c93.E77]；现有条目不足以推广到所有任务或无限增加，此为我的推断[Rb7f8295ef5d44c93.E31]。",
    "entries": [
      {
        "entry_id": "Rb7f8295ef5d44c93.E2",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Abstract",
        "field": "Abstract",
        "content": "This paper investigates retrieval-augmented generative models for open-domain question answering, combining BM25 or DPR passage retrieval with a sequence-to-sequence Fusion-in-Decoder model initialized from T5. The encoder processes each retrieved passage independently with the question and special tokens, and the decoder attends over the concatenated encoder representations to generate the answer. Fusion-in-Decoder (large) obtains 51.4 EM on NaturalQuestions, 67.6/80.1 EM on TriviaQA open/hidden test sets, and 56.7 EM/63.2 F1 on SQuAD Open (Table 1). Increasing retrieved passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5%, which the authors interpret as evidence that sequence-to-sequence models efficiently aggregate and combine evidence from multiple passages (Figure 3; Section 4). Training with fewer passages decreases accuracy, but finetuning with 100 passages for 1000 steps reaches 46.0 EM on NaturalQuestions using 147 GPU hours versus 425 GPU hours when training on 100 passages (Table 2; Section 4).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E12",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Main Method(s)",
        "content": "Two-step open-domain QA: (1) passage retrieval with BM25 or DPR; (2) generative Fusion-in-Decoder model. Fusion-in-Decoder processes each retrieved passage and its title concatenated with the question independently in the encoder, adds special tokens `question:`, `title:`, `context:`, and the decoder attends over the concatenation of all retrieved-passage representations. Evidence: Section 3 Method (Retrieval, Reading); Figure 2.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E21",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Synthesis and Processing",
        "field": "Concluding remarks",
        "content": "The synthesis and processing pipeline is a retrieval-then-generation framework, not a physical synthesis route. The main controllable variables are retrieval method, number of retrieved passages, model size, and number of training passages. The paper reports that processing passages independently in the encoder keeps computation linear in the number of passages, while joint decoding enables evidence fusion.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E31",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure and Characterization",
        "field": "Measurement Conditions",
        "content": "Datasets: NaturalQuestions (open-domain version discarding answers with more than 5 tokens), TriviaQA (unfiltered), SQuAD v1.1. Wikipedia dumps: Dec. 20, 2018 for NQ/TriviaQA; Dec. 21, 2016 for SQuAD. Following Lee et al. (2019), validation is used as test and 10% of training is kept for validation. Passages are 100 words, non-overlapping. Retrieve 100 passages unless stated otherwise and truncate to 250 word pieces. NQ/TriviaQA use DPR; SQuAD uses BM25. Greedy decoding. TriviaQA results reported on open-domain test set and hidden test set. Evidence: Section 4 Experiments (Datasets, Technical details); Table 1 caption.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E32",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure and Characterization",
        "field": "Results",
        "content": "FiD base: NQ EM 48.2; TriviaQA open EM 65.0; TriviaQA hidden EM 77.1; SQuAD Open EM 53.4, F1 60.6. FiD large: NQ EM 51.4; TriviaQA open EM 67.6; TriviaQA hidden EM 80.1; SQuAD Open EM 56.7, F1 63.2. Increasing passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5%. Training with fewer passages reduces accuracy; finetuning with 100 passages for 1000 steps reaches 46.0 EM on NaturalQuestions with 147 GPU hours vs. 425 GPU hours when training on 100 passages. Evidence: Table 1; Table 2; Figure 3; Section 4.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E49",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Test Conditions",
        "content": "Retrieve 100 passages unless stated otherwise; truncate to 250 word pieces; NQ/TriviaQA use DPR; SQuAD uses BM25; greedy decoding. NaturalQuestions open-domain version discards answers with more than 5 tokens. TriviaQA uses unfiltered version. SQuAD v1.1. Following Lee et al. (2019), validation used as test and 10% of training kept for validation. Models fine-tuned per dataset with Adam, constant learning rate 1e-4, dropout 10%, 10k gradient steps, batch size 64, 64 Tesla V100 32Gb. Evaluate every 500 steps and select best on validation Exact Match. Evidence: Section 4 Experiments (Datasets, Evaluation, Technical details); Table 1 caption.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E56",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Device Performance Characterization",
        "field": "Concluding remarks",
        "content": "Fusion-in-Decoder sets new state-of-the-art results on NaturalQuestions and TriviaQA at the time of writing. Its performance improves with more retrieved passages, unlike most extractive models that peak around 10–20 passages. The system also benefits from retrieval compared with a much larger closed-book T5 model, indicating that explicit text memory is competitive for knowledge retrieval.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E76",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Evidence Chain",
        "content": "Retrieve passages with BM25 or DPR → encode each passage and title concatenated with the question independently in the encoder → decoder attends over the concatenation of all passage representations → generate answer; independent encoding keeps computation linear in the number of passages, while joint decoding aggregates evidence from multiple passages. Evidence: Section 3 Method (Reading); Figure 2.",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      },
      {
        "entry_id": "Rb7f8295ef5d44c93.E77",
        "title": "Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering",
        "section": "Structure–Property Relationship and Mechanistic Claims",
        "field": "Quantitative Support",
        "content": "FiD large: NQ 51.4 EM; TriviaQA open/hidden 67.6/80.1 EM; SQuAD Open 56.7 EM/63.2 F1 (Table 1). Increasing retrieved passages from 10 to 100 improves TriviaQA by 6% and NaturalQuestions by 3.5% (Figure 3; Section 4 Scaling with number of passages). Training with fewer passages decreases accuracy; finetuning with 100 passages for 1000 steps reaches 46.0 EM on NQ with 147 GPU hours vs. 425 GPU hours when training on 100 passages (Table 2; Section 4 Impact of the number of training passages).",
        "content_sha256": "b7f8295ef5d44c932c8e87306d1119015c4021c98b02235d22531298d56a39ed"
      }
    ],
    "tools_called": 0,
    "desktop_mobile_excerpt_match": true
  },
  "first_reader_attempt": {
    "answer": "论文报告：方法用BM25或DPR检索段落，FiD将每段与标题、问题独立编码，解码器对全部段落表示联合注意力后生成答案[R...E12][R...E76]；默认检索100段、截断250词片[R...E13]。评测限于NQ、TriviaQA、SQuAD Open，NQ/TriviaQA用DPR，SQuAD用BM25[R...E31][R...E58]。在NQ/TriviaQA上，10→100段分别提升3.5%和6%[R...E2][R...E77]；训练段数5→100时，NQ EM 37.8→46.5，TriviaQA 58.1→64.7[R...E20][R...E87]。\n\n我的推断：不能断言所有任务中检索段落越多越好。论文报告仅覆盖上述数据集、检索器和设置，且当前条目未提供SQuAD等任务的段落数缩放结果[R...E31][R...E32]。",
    "accepted": false,
    "reason": "模型把引用缩写为 R...，不可展开；保留原始输出，提示词补强后使用相同问题复测"
  },
  "audit": {
    "reviewer": "VibeResearch 产品维护者",
    "date": "2026-09-22",
    "scope": "逐项检查公开正文采用的判断；完整原始分析并不因此获得科学正确性认证",
    "checks": [
      {
        "id": "retrieval-vs-answer",
        "claim": "DPR 的 top-k 检索准确率不能与 FiD 的答案 EM 直接排大小",
        "entries": [
          "P1.E44",
          "P1.E76",
          "P2.E50",
          "P2.E14"
        ],
        "source": "DPR §2，PDF 第2页；FiD §4 Evaluation，PDF 第3页",
        "status": "SUPPORTED",
        "note": "DPR 分母是问题数；部分结构化段落措辞有歧义，定义以 §2 为准。"
      },
      {
        "id": "two-numbers",
        "claim": "78.4 是 Single DPR 的 NQ top-20；51.4 是 FiD large 的 NQ EM",
        "entries": [
          "P1.E77",
          "P2.E85"
        ],
        "source": "DPR Table 2，PDF 第5页；FiD Table 1，PDF 第3页",
        "status": "SUPPORTED",
        "note": "不能因 78.4 大于 51.4 就宣称 DPR 更好。"
      },
      {
        "id": "dpr-also-em",
        "claim": "DPR 也报告端到端 EM，NQ 41.5；两篇并非一切指标都不可比",
        "entries": [
          "P1.E81",
          "P2.E85"
        ],
        "source": "DPR Table 4，PDF 第7页；FiD Table 1，PDF 第3页",
        "status": "SUPPORTED_WITH_CONDITIONS",
        "note": "同为 EM 可记录论文报告差异，不是控制变量消融，不能归因于某一个部件。"
      },
      {
        "id": "reader-method",
        "claim": "FiD 各段与问题分别编码，解码器对拼接后的表示做注意力",
        "entries": [
          "Rb7f8295ef5d44c93.E12",
          "Rb7f8295ef5d44c93.E76"
        ],
        "source": "FiD Figure 2，PDF 第2页；§3 Reading，PDF 第3页",
        "status": "SUPPORTED",
        "note": "应说拼接表示，不是拼接注意力；网页人工说明使用准确表述，模型原句不改。"
      },
      {
        "id": "reader-conditions",
        "claim": "默认检索100段，截断250词片；NQ/TriviaQA用DPR，SQuAD用BM25",
        "entries": [
          "Rb7f8295ef5d44c93.E31",
          "Rb7f8295ef5d44c93.E49"
        ],
        "source": "FiD §4 Technical details，PDF 第3–4页",
        "status": "SUPPORTED",
        "note": "均有 unless stated otherwise 的例外，不能推广到每个消融实验。"
      },
      {
        "id": "reader-scaling",
        "claim": "测试范围内10到100段有增益，不支持所有任务中越多越好",
        "entries": [
          "Rb7f8295ef5d44c93.E32",
          "Rb7f8295ef5d44c93.E77"
        ],
        "source": "FiD Figure 3，PDF 第4页；Scaling with number of passages，PDF 第4–5页",
        "status": "SUPPORTED_WITH_CONDITIONS",
        "note": "论文写6%和3.5%；不要另行推导相对增幅或所有任务趋势。不可推广是维护者/模型的范围判断，不是作者原句。"
      },
      {
        "id": "comparison-top100",
        "claim": "原始主要判断第4条把DPR端到端结果描述为top100阅读",
        "entries": [
          "P1.E47",
          "P1.E81"
        ],
        "source": "DPR §6.2，PDF 第8页",
        "status": "NEEDS_CORRECTION",
        "note": "论文称 NQ 最优 k=50。top100候选、训练采样和最终读者 k 不应混为一谈；原分析保留，不作为已核对结论转述。"
      },
      {
        "id": "comparison-dataset-version",
        "claim": "原始边界第1条说NQ版本可能不一致",
        "entries": [
          "P1.E43",
          "P1.E50",
          "P2.E31",
          "P2.E58"
        ],
        "source": "DPR §4.2；FiD §4 Datasets",
        "status": "UNRESOLVED",
        "note": "结构化层未说明完全一致，不足以证明实际数据版本不同。保留为待核实，不能宣传成已确认差异。"
      },
      {
        "id": "comparison-retriever-combination",
        "claim": "原始边界第2条称两篇不是同一检索器/生成器组合",
        "entries": [
          "P1.E35",
          "P2.E59"
        ],
        "source": "DPR §3/§6；FiD §3/§4",
        "status": "SUPPORTED_WITH_CONDITIONS",
        "note": "读取/生成部件不同，但 NQ/TriviaQA 的 FiD 也使用 DPR，不能改写成检索器一定不同。"
      },
      {
        "id": "source-denominator",
        "claim": "结构化 P1.E47/P1.E76 的 top-k 定义有歧义",
        "entries": [
          "P1.E44",
          "P1.E47",
          "P1.E76"
        ],
        "source": "DPR §2，PDF 第2页",
        "status": "EXTRACTION_LIMITATION",
        "note": "应核对问题中有多少能在前k段找到答案，不是统计所有检索段落中含答案段落的比例。不改输入来美化本轮。"
      },
      {
        "id": "reader-extractive-peaks",
        "claim": "回答中的多数抽取模型10–20段达峰",
        "entries": [
          "Rb7f8295ef5d44c93.E56"
        ],
        "source": "FiD Scaling with number of passages，PDF 第5页",
        "status": "SUPPORTED_WITH_CONDITIONS",
        "note": "这是作者对相关工作的有限观察，原文带 seems，不应读作对所有现代抽取模型的保证。"
      }
    ],
    "not_claimed": [
      "所有原始分析判断均正确",
      "跨学科完整提取已通过",
      "引用可点击等于科学结论已核实",
      "提示词补强能保证所有后续输出编号正确"
    ]
  },
  "screenshots": [
    {
      "name": "zotero-menu.png",
      "sha256": "1c6348fa1858ac98b552794da5e1dc8d55c2f294f670f0f31ae42922e058e2a5"
    },
    {
      "name": "browser-approval.png",
      "sha256": "3c37e4b0725ddc709f700a88472edb5dd275881a34b009bf186ee6df8041debc"
    },
    {
      "name": "browser-imported.png",
      "sha256": "717fa911bf3251a607666c7cc23e98b7f25d3f14c89da12b7813eb823dc0f323"
    },
    {
      "name": "comparison-question.png",
      "sha256": "dcfcd2a843835f9ac6f7722c20f65e759a97982501841c9f7c1a087c2b0ff919"
    },
    {
      "name": "comparison-result.png",
      "sha256": "1e45fbc9274a2d769f1608750933d1b74b9b5e343f1e87a6f77808c200388a78"
    },
    {
      "name": "comparison-evidence.png",
      "sha256": "44cafd3c300f40c6283356d6b5bba5f9e3f636f7d67b9e0effea0150fd422892"
    },
    {
      "name": "reader-question.png",
      "sha256": "9cf145b84c2db7578ce8a603ffc4ce730dedcf281a9c95658d3a89bbc3e42b6d"
    },
    {
      "name": "reader-desktop-citation.png",
      "sha256": "64ec7b50fc92215dc9592f9f1a6232bd3dd8f3f4ee2185dc30c0f3485040a29d"
    },
    {
      "name": "reader-mobile-citation.png",
      "sha256": "82bb9c8ccd1bc36e1ef793c3ccece4f02f8c4b92ce3705e78e0ea73eaa88b405"
    }
  ]
}