path: root/table.html

<table border="1" cellspacing="0" cellpadding="5">
  <thead>
    <tr>
      <th>Paper Title</th>
      <th>Approach</th>
      <th>Datasets Used</th>
      <th>Results</th>
      <th>Key Contributions</th>
    </tr>
  </thead>
  <tbody>
    <!-- Paper 1 -->
    <tr>
      <td>
        <ul>
          <li>A Neural Corpus Indexer for Document Retrieval (Wang et al., 2022)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>End-to-end seq2seq network with a Prefix-Aware Weight-Adaptive (PAWA) Decoder</li>
          <li>Query generation network and hierarchical k-means indexing</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>NQ320k</li>
          <li>TriviaQA</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>+21.4% relative improvement in Recall@1 on NQ320k</li>
          <li>+16.8% improvement in R-Precision on TriviaQA</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Unifies training and indexing</li>
          <li>Introduces a novel decoder and realistic query–document pair generation for enhanced retrieval performance</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 2 -->
    <tr>
      <td>
        <ul>
          <li>Active Retrieval Augmented Generation (Jiang et al., 2023)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Dynamic, iterative retrieval integrated into generation (FLARE)</li>
          <li>Detects low-confidence tokens and retrieves additional context</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Knowledge-intensive tasks (e.g., multihop QA, open-domain summarization)</li>
          <li>Specific datasets not detailed</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Significant performance improvements in complex, long-form generation tasks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Introduces a forward-looking, active retrieval mechanism</li>
          <li>Moves beyond static, single-time retrieval methods</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 3 -->
    <tr>
      <td>
        <ul>
          <li>Atlas Few-shot Learning with Retrieval Augmented Language Models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Dual-encoder retrieval combined with a sequence-to-sequence generator</li>
          <li>Joint pre-training of both components</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Natural Questions</li>
          <li>MMLU, KILT benchmarks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Over 42% accuracy on Natural Questions with only 64 training examples</li>
          <li>Outperforms larger models (e.g., PaLM) by 3%</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Demonstrates effective few-shot learning with minimal data</li>
          <li>Offers an adaptable document index</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 4 -->
    <tr>
      <td>
        <ul>
          <li>Benchmarking Large Language Models in Retrieval-Augmented Generation (Chen et al., 2024)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Evaluation framework (RGB) assessing retrieval quality in LLMs (e.g., ChatGPT, ChatGLM, Vicuna)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Evaluation tasks in English and Chinese under varying noise conditions</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Accuracy drop: e.g., ChatGPT from 96.33% to 76% with noise</li>
          <li>Multi-document integration challenges (accuracy drops to 43–55%)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Provides a rigorous benchmark for RAG settings</li>
          <li>Highlights error detection and rejection behaviors in LLMs</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 5 -->
    <tr>
      <td>
        <ul>
          <li>C-RAG Certified Generation Risks for Retrieval-Augmented Language Models (Kang et al., 2024)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Conformal risk analysis to certify generation risks</li>
          <li>Establishes an upper bound (“conformal generation risk”)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>AESLC</li>
          <li>CommonGen, DART, E2E</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Consistently lower conformal generation risks compared to non-retrieval models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Extends conformal prediction methods to RAG</li>
          <li>Provides a framework for risk certification in generation</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 6 -->
    <tr>
      <td>
        <ul>
          <li>Can Knowledge Graphs Reduce Hallucinations in LLMs: A Survey (Agrawal et al., 2024)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Survey categorizing KG integration methods into:</li>
          <li>&nbsp;&nbsp;&bull; Knowledge-aware inference</li>
          <li>&nbsp;&nbsp;&bull; Knowledge-aware learning</li>
          <li>&nbsp;&nbsp;&bull; Knowledge-aware validation</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Aggregated studies across multiple tasks</li>
          <li>No single dataset</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Up to 80% enhancement in answer correctness in certain settings</li>
          <li>Improved chain-of-thought reasoning</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Comprehensively categorizes KG-based augmentation methods</li>
          <li>Addresses hallucination reduction in LLMs</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 7 -->
    <tr>
      <td>
        <ul>
          <li>Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Dual-encoder dense vector representations for semantic matching</li>
          <li>Utilizes in-batch negative training</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Natural Questions</li>
          <li>Other open-domain QA benchmarks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Top-20 accuracy of 78.4% on Natural Questions (vs. 59.1% for BM25)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Introduces dense retrieval techniques</li>
          <li>Significantly improves semantic matching in QA systems</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 8 -->
    <tr>
      <td>
        <ul>
          <li>DocPrompting Generating Code by Retrieving the Docs (Zhou et al., 2023)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Retrieval of documentation to guide code generation</li>
          <li>Focuses on documentation rather than NL-code pairs</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>CoNaLa (Python)</li>
          <li>Curated Bash dataset</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>52% relative gain in pass@1</li>
          <li>30% relative gain in pass@10 on CoNaLa</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Highlights the importance of documentation retrieval</li>
          <li>Boosts code generation accuracy and generalization</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 9 -->
    <tr>
      <td>
        <ul>
          <li>Document Language Models, Query Models, and Risk Minimization for Information Retrieval<br>(Ponte & Croft, 1998; Berger & Lafferty, 1999; Lafferty & Zhai, 2001)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Combines unigram language models</li>
          <li>Statistical translation methods</li>
          <li>Markov chain query models and Bayesian risk minimization</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>TREC collections</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Significant improvements over traditional vector space models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Laid the foundation for integrating DLMs, QMs, and risk minimization</li>
          <li>Influenced modern retrieval methods</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 10 -->
    <tr>
      <td>
        <ul>
          <li>Evaluating Retrieval Quality in Retrieval-Augmented Generation (Salemi & Zamani, 2024)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>eRAG: Uses LLMs to generate document-level relevance labels</li>
          <li>Labels correlate with downstream performance</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Various downstream RAG tasks</li>
          <li>Exact datasets not specified</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Kendall’s tau increased from 0.168 to 0.494</li>
          <li>Up to 50× memory efficiency and 2.468× speedup</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Proposes a novel evaluation metric aligning retrieval quality with end-task performance</li>
          <li>Reduces computational overhead</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 11 -->
    <tr>
      <td>
        <ul>
          <li>Fine Tuning vs Retrieval Augmented Generation for Less Popular Knowledge</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Comparative analysis between fine tuning (FT) and RAG</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Not explicitly specified</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>RAG achieves higher accuracy for low-frequency entities</li>
          <li>Hybrid FT+RAG yields best results for smaller models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Highlights benefits of RAG over traditional fine tuning</li>
          <li>Effective for less popular or emerging knowledge</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 12 -->
    <tr>
      <td>
        <ul>
          <li>How Much Knowledge Can You Pack Into the Parameters of a Language Model (Roberts et al., 2020)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Fine-tuning with salient span masking (SSM) as a pre-training objective</li>
          <li>Applied to open-domain QA</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Natural Questions</li>
          <li>WebQuestions</li>
          <li>TriviaQA</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Larger models outperform smaller ones</li>
          <li>Significant performance gains with SSM</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Contrasts closed-book vs. open-book QA</li>
          <li>Demonstrates task-specific pre-training benefits</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 13 -->
    <tr>
      <td>
        <ul>
          <li>Learning Transferable Visual Models From Natural Language Supervision<br>(Radford et al., 2021; Brown et al., 2020; Deng et al., 2009)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Contrastive Language-Image Pre-training (CLIP)</li>
          <li>Joint image and text encoders</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>400M (image, text) pairs</li>
          <li>Evaluated on ImageNet and other benchmarks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Competitive zero-shot performance on ImageNet</li>
          <li>Robust to natural distribution shifts</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Bridges visual and textual modalities</li>
          <li>Enables transferable visual representations via contrastive learning</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 14 -->
    <tr>
      <td>
        <ul>
          <li>Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering<br>(Izacard & Grave, 2021; Roberts et al., 2020)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Fusion-in-Decoder: Independently encodes multiple passages</li>
          <li>Aggregates evidence in the decoder</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Natural Questions</li>
          <li>TriviaQA</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>State-of-the-art Exact Match scores</li>
          <li>Performance scales with more retrieved passages</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Combines retrieval with generation to synthesize evidence</li>
          <li>Improves open-domain QA accuracy</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 15 -->
    <tr>
      <td>
        <ul>
          <li>Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., 2023)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>HyDE: Two-step process</li>
          <li>Generates hypothetical documents using instruction-following LMs</li>
          <li>Applies unsupervised contrastive encoding</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Various tasks: web search, QA, fact verification (multi-language settings)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Outperforms existing unsupervised dense retrieval models</li>
          <li>Competitive with fine-tuned models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Enables effective zero-shot retrieval without explicit relevance labels</li>
          <li>Leverages hypothetical document generation</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 16 -->
    <tr>
      <td>
        <ul>
          <li>Re2G Retrieve, Rerank, Generate (Lewis et al., 2020; Guu et al., 2020)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Integrated framework combining retrieval, reranking, and generation</li>
          <li>Uses knowledge distillation</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Various tasks (exact datasets not specified)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Enhanced selection of relevant passages</li>
          <li>Improved overall performance across tasks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Unifies retrieval, reranking, and generation in an end-to-end framework</li>
          <li>Improves evidence selection</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 17 -->
    <tr>
      <td>
        <ul>
          <li>REALM Retrieval-Augmented Language Model Pre-Training<br>(Guu et al., 2020; Devlin et al., 2018; Raffel et al., 2019)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Two-step process: retrieval followed by masked language model prediction</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Natural Questions</li>
          <li>WebQuestions</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>4–16% absolute accuracy improvements on open-domain QA benchmarks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Integrates retrieval into pre-training</li>
          <li>Enhances prediction accuracy and model interpretability</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 18 -->
    <tr>
      <td>
        <ul>
          <li>REST Retrieval-Based Speculative Decoding<br>(He et al., 2024; Miao et al., 2023; Chen et al., 2023)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Uses a non-parametric retrieval datastore to construct draft tokens</li>
          <li>For speculative decoding</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>HumanEval</li>
          <li>MT-Bench</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>1.62× to 2.36× speedup in token generation</li>
          <li>Compared to standard autoregressive methods</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Improves generation speed without additional training</li>
          <li>Allows seamless integration with various LLMs</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 19 -->
    <tr>
      <td>
        <ul>
          <li>Retrieval Augmentation Reduces Hallucination in Conversation<br>(Roller et al., 2021; Maynez et al., 2020; Lewis et al., 2020b; Shuster et al., 2021; Dinan et al., 2019b; Zhou et al., 2021)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Integration of retrieval mechanisms into dialogue systems</li>
          <li>Fetches relevant documents for improved factuality</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Knowledge-grounded conversational datasets</li>
          <li>Specific names not provided</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Reduced hallucination rates</li>
          <li>Higher factual accuracy compared to standard models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Demonstrates more reliable, factually grounded conversational responses</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 20 -->
    <tr>
      <td>
        <ul>
          <li>Retrieval Augmented Code Generation and Summarization<br>(Parvez et al., 2021; Karpukhin et al., 2020; Feng et al., 2020; Guo et al., 2021; Ahmad et al., 2021)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>REDCODER framework: Two-step process combining retrieval with generation</li>
          <li>Uses pre-trained code models</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Code generation benchmarks (e.g., CoNaLa, CodeXGLUE)</li>
          <li>Evaluated via BLEU, Exact Match, CodeBLEU</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Significant improvements in BLEU, Exact Match, and CodeBLEU scores</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Enhances code generation and summarization</li>
          <li>Effectively retrieves relevant code snippets and integrates pre-trained models</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 21 -->
    <tr>
      <td>
        <ul>
          <li>Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks<br>(Lewis et al., 2020; Karpukhin et al., 2020; Guu et al., 2020; Petroni et al., 2019)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Combines retrieval with generative modeling to synthesize external knowledge</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Various knowledge-intensive benchmarks (e.g., Natural Questions, TriviaQA)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Outperforms extractive and closed-book models in accuracy and robustness</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Balances internal model knowledge with external retrieval</li>
          <li>Provides accurate and comprehensive answers</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 22 -->
    <tr>
      <td>
        <ul>
          <li>Retrieval-Enhanced Machine Learning (Zamani et al., 2022)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>REML framework: Joint optimization of a prediction model and a retrieval model</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Applied in domain adaptation and few-shot learning scenarios</li>
          <li>Datasets not specified</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Improves model generalization, scalability, and interpretability</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Offloads memorization to a retrieval system</li>
          <li>Supports dynamic updates to knowledge bases</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 23 -->
    <tr>
      <td>
        <ul>
          <li>Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (Asai et al., 2024)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Adaptive retrieval with self-reflection using “reflection tokens”</li>
          <li>Structured self-critique</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Open-domain QA, reasoning, and fact verification tasks</li>
          <li>Specific datasets not provided</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Outperforms state-of-the-art models in factuality and citation accuracy</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Introduces self-critique into the RAG pipeline</li>
          <li>Enables adaptive retrieval and improved output reliability</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 24 -->
    <tr>
      <td>
        <ul>
          <li>The Probabilistic Relevance Framework – BM25 and Beyond<br>(Robertson & Sparck Jones, 1977; Robertson et al., 1994; Robertson & Zaragoza, 2009; Sparck Jones et al., 2000)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Probabilistic relevance modeling using term frequency, inverse document frequency, and document length normalization</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>TREC collections and other ad-hoc retrieval tasks</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Demonstrated robust performance as a benchmark for relevance estimation</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Provides the theoretical foundation for modern IR systems</li>
          <li>Basis for the widely adopted BM25 scoring function</li>
        </ul>
      </td>
    </tr>
    <!-- Paper 25 -->
    <tr>
      <td>
        <ul>
          <li>TIARA Multi-grained Retrieval for Robust Question Answering over Large Knowledge Base<br>(Gu et al., 2021; Ye et al., 2021; Chen et al., 2021; Raffel et al., 2020; Devlin et al., 2019)</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Multi-grained retrieval integrating entity, exemplary logical form, and schema retrieval</li>
          <li>Uses constrained decoding</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>GrailQA</li>
          <li>WebQuestionsSP</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Significant improvements in compositional and zero-shot generalization</li>
          <li>Outperforms previous methods</li>
        </ul>
      </td>
      <td>
        <ul>
          <li>Addresses KBQA challenges by retrieving multiple granularities of context</li>
          <li>Enhances accuracy and reliability of logical form generation</li>
        </ul>
      </td>
    </tr>
  </tbody>
</table>