Technical Accuracy and Factual Precision

Technical accuracy and factual precision in AI citation mechanics and ranking factors represent critical evaluation dimensions that assess the reliability and trustworthiness of AI-generated content by ensuring correct attribution of information to verifiable sources, maintaining factual consistency across outputs, and ranking information based on credibility and relevance 12. These concepts serve as foundational pillars for large language models (LLMs) and retrieval-augmented generation (RAG) systems, directly impacting how AI systems prioritize and present information across academic, commercial, and public domains 34. As AI systems increasingly mediate information access and knowledge synthesis, ensuring technical accuracy in citation mechanisms becomes paramount for preventing misinformation propagation, maintaining scholarly integrity, and building user trust in automated knowledge systems 57.

Overview

The emergence of technical accuracy and factual precision as critical concerns in AI systems stems from the fundamental challenge of hallucination in large language models—the generation of plausible-sounding but factually incorrect or unsupported information 1012. Early language models, while impressive in their fluency and coherence, frequently produced outputs that lacked grounding in verifiable sources, limiting their utility for knowledge-intensive tasks where accuracy is paramount 611. This limitation became particularly evident as these systems began deployment in high-stakes domains such as healthcare, legal research, and academic scholarship, where factual errors could have serious consequences.

The practice has evolved significantly with the development of retrieval-augmented generation frameworks that combine neural retrieval with conditional generation 1211. These systems address the fundamental challenge by anchoring generated text to verifiable sources retrieved from large document corpora, enabling both improved factual accuracy and explicit citation of supporting evidence. The introduction of specialized datasets like FEVER (Fact Extraction and VERification) and evaluation frameworks such as Attributed Question Answering (AQA) has provided standardized benchmarks for assessing citation quality and factual consistency 58. More recent innovations include systems like WebGPT and GopherCite, which explicitly train models to search, browse, and cite sources through human feedback, representing a shift toward treating citation generation as a first-class modeling objective rather than a post-hoc addition 34.

Key Concepts

Grounding

Grounding refers to the process of anchoring generated text to verifiable sources, ensuring that AI outputs are supported by retrievable evidence rather than relying solely on patterns learned during pre-training 12. This concept represents a fundamental shift from purely generative approaches to hybrid systems that maintain explicit connections between outputs and source materials.

Example: A medical AI assistant responding to a query about diabetes treatment generates the statement "Metformin is typically the first-line medication for type 2 diabetes management." The system grounds this claim by retrieving and citing a passage from the American Diabetes Association's clinical practice guidelines, providing both the specific guideline document and the relevant section number, allowing healthcare professionals to verify the recommendation against authoritative sources before applying it in clinical decision-making.

Hallucination Detection

Hallucination detection encompasses techniques for identifying when AI systems generate false, unsupported, or fabricated information that appears plausible but lacks basis in training data or retrieved sources 1012. This involves both automated verification mechanisms and confidence scoring to flag potentially unreliable outputs.

Example: A legal research AI generates a response citing "Johnson v. State Supreme Court (2019)" as precedent for a particular interpretation of contract law. The hallucination detection system queries the legal database and finds no case matching this citation, flagging it as a potential fabrication. The system then either removes the citation, replaces it with a verified alternative, or explicitly indicates to the user that the claim could not be verified, preventing reliance on non-existent legal precedent.

Attribution Accuracy

Attribution accuracy measures the precision with which generated claims are linked to their actual sources, ensuring that citations correctly identify the documents and passages that support specific statements 78. This goes beyond merely providing citations to verifying that the cited sources actually contain the attributed information.

Example: An AI research assistant helping with a literature review generates the statement "Recent studies show that transformer models achieve 95% accuracy on sentiment analysis tasks" and attributes this to "Devlin et al. (2019) BERT paper." The attribution accuracy component verifies that the cited paper actually reports this specific metric for sentiment analysis, rather than for a different task like question answering, and confirms that the 95% figure appears in the results section, ensuring the citation genuinely supports the claim as stated.

Citation Recall and Precision

Citation recall measures the proportion of claims in generated text that receive appropriate citations, while citation precision assesses the accuracy of provided citations 712. Together, these metrics evaluate both the completeness and correctness of citation behavior in AI systems.

Example: A scientific writing assistant generates a three-paragraph summary of climate change research containing twelve factual claims. Citation recall analysis reveals that ten of the twelve claims include citations (83% recall), while citation precision verification confirms that eight of the ten citations accurately support their associated claims (80% precision). This analysis identifies that two claims lack citations entirely and two citations point to sources that don't fully support the stated claims, highlighting specific areas for improvement.

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation represents an architectural paradigm that combines neural retrieval mechanisms with text generation, allowing models to access external knowledge bases during the generation process 1211. This approach addresses knowledge limitations by retrieving relevant documents before or during generation.

Example: A customer service AI receives a technical question about a specific software feature released last month. Rather than relying solely on potentially outdated training data, the RAG system first retrieves the three most relevant passages from the current product documentation and recent release notes. It then generates a response that synthesizes information from these retrieved passages, providing inline citations to specific documentation sections and including direct quotations for critical technical specifications, ensuring the response reflects the most current product capabilities.

Factual Consistency Scoring

Factual consistency scoring employs natural language inference (NLI) models and specialized verification algorithms to assess whether generated claims align with source materials, quantifying the degree of semantic agreement between outputs and evidence 510. This provides explicit confidence measures for factual accuracy.

Example: A news summarization system generates the statement "The Federal Reserve raised interest rates by 0.75 percentage points" based on a retrieved news article. The factual consistency scorer uses an NLI model to compare this generated claim against the source text, which states "The Fed implemented a three-quarter point rate increase." The system assigns a high consistency score (0.92 out of 1.0) recognizing these as semantically equivalent statements, while a claim stating "The Federal Reserve raised rates by 1 percentage point" would receive a low score, flagging a factual discrepancy.

Source Credibility Ranking

Source credibility ranking involves ordering retrieved documents and potential citations based on multiple factors including publication venue, author reputation, citation authority, recency, and methodological quality 89. This ensures that AI systems prioritize more reliable sources when multiple options are available.

Example: An academic research assistant retrieving information about vaccine efficacy encounters five potential sources: a peer-reviewed study in The Lancet with 2,400 citations, a preprint server article with preliminary results, a blog post by a medical professional, a news article summarizing research, and a social media thread. The credibility ranking algorithm assigns highest weight to the peer-reviewed Lancet study based on venue prestige, citation count, and peer review status, followed by the preprint (recent but unreviewed), then the professional blog, news article, and finally the social media content, ensuring the generated response prioritizes the most authoritative evidence.

Applications in Knowledge-Intensive AI Systems

Academic Research Assistance

Technical accuracy and factual precision are applied in academic research tools that help scholars conduct literature reviews, identify relevant studies, and synthesize findings across multiple papers 68. Systems like Elicit and Consensus implement domain-specific ranking that prioritizes peer-reviewed sources and evaluates study methodology quality, providing researchers with evidence-based summaries that include precise citations to original research papers. These applications employ specialized verification to ensure that claims about study findings, sample sizes, and statistical significance accurately reflect the source publications, maintaining scholarly integrity in AI-assisted research workflows.

Healthcare Information Systems

In medical AI applications, factual precision becomes critical as incorrect information can directly impact patient safety 9. Clinical decision support systems implement multi-layered verification that cross-references generated recommendations against authoritative medical databases, clinical practice guidelines, and peer-reviewed literature. For instance, a system providing drug interaction warnings retrieves information from pharmaceutical databases, verifies contraindications against multiple sources, and cites specific sections of prescribing information, enabling healthcare providers to validate recommendations before clinical application. These systems often implement conservative confidence thresholds, refusing to generate responses when certainty falls below safety-critical levels.

Legal Research and Analysis

Legal AI systems like CaseText's CoCounsel apply technical accuracy principles through jurisdiction-aware citation and validation against official legal databases 4. These applications must handle complex requirements including temporal validity of precedents, jurisdictional variations in law, and hierarchical authority of different court decisions. The systems implement specialized ranking that considers factors such as whether a case has been overturned, the level of the deciding court, and the relevance of the legal issue to the current query, ensuring that generated legal analysis cites applicable and current authority.

Conversational Search Engines

Modern conversational AI systems like Perplexity AI and Bing Chat (now Copilot) integrate web search with natural language generation, providing inline citations with source previews 310. These applications implement real-time retrieval across web-scale corpora, ranking sources based on relevance, recency, and domain authority. The systems generate responses with numbered citations linked to specific sentences, allowing users to immediately verify claims by accessing source materials. This application context requires balancing comprehensiveness with response latency, often employing hierarchical retrieval strategies that perform initial broad searches followed by focused re-ranking of top candidates.

Best Practices

Implement Multi-Stage Verification Pipelines

Robust citation systems should employ layered verification that combines multiple validation approaches rather than relying on a single mechanism 510. The rationale is that different verification methods catch different types of errors: NLI models excel at semantic consistency checking, database lookups verify entity existence, and citation format validators ensure proper attribution structure.

Implementation Example: A biomedical AI system implements a three-stage verification pipeline. First, it uses an NLI model trained on medical literature to assess whether generated claims are supported by retrieved passages (semantic verification). Second, it queries structured medical databases like PubMed to confirm that cited publications exist and match the stated metadata (existence verification). Third, it employs a specialized fact-checking model trained on the FEVER dataset to classify the relationship between claims and evidence as "supported," "refuted," or "insufficient information" (claim verification). Only outputs passing all three stages are presented to users without warnings, while those failing any stage are flagged for human review.

Prioritize Source Diversity in Ranking Algorithms

Ranking systems should explicitly incorporate source diversity metrics to avoid over-reliance on frequently-cited sources and reduce echo chamber effects 89. This practice recognizes that comprehensive understanding often requires synthesizing perspectives from multiple sources, including emerging research that may not yet have accumulated extensive citations.

Implementation Example: A scientific literature assistant implements a diversity-aware ranking function that combines relevance scores with source diversity penalties. After retrieving candidate papers, the system clusters them by research group, methodology, and publication venue. The ranking algorithm then applies a diversity bonus to papers from underrepresented clusters, ensuring that the top-ranked results include methodologically diverse approaches rather than multiple papers from the same research group using similar methods. For a query about climate modeling, this ensures the system surfaces both observational studies and simulation-based approaches, rather than exclusively prioritizing one methodology.

Provide Explicit Uncertainty Quantification

Systems should communicate confidence levels and uncertainty explicitly rather than presenting all outputs with equal apparent certainty 712. This practice enables users to appropriately calibrate their trust and apply additional scrutiny to lower-confidence claims, particularly important in high-stakes domains.

Implementation Example: A legal research AI generates responses with color-coded confidence indicators and explicit uncertainty statements. High-confidence claims (>0.9 consistency score, supported by multiple authoritative sources) appear in standard text with citations. Medium-confidence claims (0.7-0.9 score, supported by single sources or with minor inconsistencies) include a yellow indicator and language like "According to available sources..." Low-confidence claims (<0.7 score) are presented with orange highlighting and explicit caveats: "Limited evidence suggests... [citation], but this claim could not be verified across multiple sources." Claims that cannot be verified are either omitted or presented with red highlighting and clear warnings about uncertainty.

Maintain Temporal Awareness and Citation Freshness

Citation systems should track publication dates, update frequencies, and temporal validity of information, prioritizing current sources for time-sensitive topics while recognizing foundational work for established concepts 89. This practice addresses the challenge that information currency requirements vary dramatically across domains and query types.

Implementation Example: A medical information system implements temporal ranking that adjusts based on query classification. For queries about treatment protocols or drug approvals, the system strongly weights sources from the past two years and flags older citations with publication dates, recognizing that clinical guidelines evolve rapidly. For queries about fundamental biological mechanisms, the system balances classic foundational papers with recent research, explicitly noting when citing older sources: "The structure of DNA was first described by Watson and Crick (1953), with subsequent research confirming..." The system maintains a background process that periodically re-retrieves information for cached queries, detecting when newer sources supersede previously cited materials.

Implementation Considerations

Tool and Format Choices

Implementing technical accuracy systems requires selecting appropriate retrieval architectures, embedding models, and verification frameworks based on specific use case requirements 1211. Dense retrieval systems using bi-encoder architectures (separate encoders for queries and documents) offer computational efficiency for large-scale applications, enabling sub-second retrieval across millions of documents through approximate nearest neighbor search. However, cross-encoder architectures that jointly encode query-document pairs provide superior accuracy for re-ranking top candidates, suggesting a two-stage approach for applications where accuracy justifies additional latency.

Example: A legal research platform serving thousands of concurrent users implements a hybrid architecture. Initial retrieval uses a bi-encoder model (based on the REALM architecture) to efficiently search across 50 million legal documents, returning the top 100 candidates in under 200 milliseconds. These candidates then pass through a cross-encoder re-ranker that jointly processes the query and each document, producing more accurate relevance scores. The final top 10 results undergo citation extraction and verification before presentation. This architecture balances the need for comprehensive corpus coverage with accuracy requirements, while maintaining acceptable response times for interactive use.

Audience-Specific Customization

Citation presentation and verification rigor should adapt to audience expertise and use case requirements 67. Expert users in specialized domains may prefer comprehensive citations with direct access to source materials, while general audiences benefit from simplified presentations that highlight key sources without overwhelming detail. High-stakes applications require more conservative confidence thresholds and extensive verification, while exploratory research tools may accept lower certainty to maximize coverage.

Example: A medical AI system implements three presentation modes. For healthcare professionals, it provides detailed citations including PubMed IDs, study design classifications, sample sizes, and confidence intervals, enabling rapid assessment of evidence quality. For patients seeking health information, it presents simplified summaries with statements like "According to multiple medical studies..." and provides links to patient-friendly source materials from authoritative organizations like the Mayo Clinic or NIH. For researchers conducting systematic reviews, it offers an advanced mode with comprehensive citation metadata, study quality scores, and tools to export citations in standard formats like BibTeX, recognizing that different audiences have different verification needs and information literacy levels.

Organizational Maturity and Context

Implementation approaches should align with organizational technical capabilities, existing infrastructure, and risk tolerance 910. Organizations with mature machine learning operations may implement custom-trained verification models and sophisticated ranking algorithms, while those earlier in AI adoption may leverage pre-built solutions and third-party APIs. Regulatory requirements in domains like healthcare and finance may mandate specific verification standards and audit trails.

Example: A financial services firm implementing an AI research assistant for investment analysts faces strict regulatory requirements for information accuracy and auditability. The organization implements a conservative architecture that maintains complete provenance tracking, logging every retrieval, generation, and verification step with timestamps and model versions. All generated outputs undergo mandatory human review before use in client-facing materials, with reviewers having access to detailed audit trails showing which sources influenced each claim. The system integrates with the firm's existing compliance infrastructure, automatically flagging outputs that cite sources not on the approved vendor list or that discuss regulated topics requiring additional review. This implementation prioritizes regulatory compliance and risk management over response speed, reflecting the organization's context and maturity level.

Computational Resource Allocation

Technical accuracy systems must balance verification thoroughness against computational costs and latency requirements 111. Dense retrieval over large corpora, cross-encoder re-ranking, and NLI-based verification all consume significant computational resources, potentially creating bottlenecks for high-throughput applications. Strategic resource allocation focuses expensive verification on high-stakes or uncertain claims while using lighter-weight methods for routine cases.

Example: A customer service AI platform handling millions of daily queries implements tiered verification based on query classification and confidence scores. Simple factual queries about business hours or return policies use cached responses with pre-verified citations, requiring minimal computation. Technical product questions trigger retrieval and basic citation matching but skip expensive NLI verification. Complex queries about warranty coverage or technical specifications—identified as high-stakes through keyword detection—receive full verification including multi-model ensemble checking and human-in-the-loop review for low-confidence outputs. This tiered approach allocates 80% of computational budget to the 20% of queries where accuracy is most critical, optimizing resource utilization while maintaining appropriate verification rigor.

Common Challenges and Solutions

Challenge: Persistent Hallucination in Low-Resource Domains

AI systems frequently hallucinate when generating content about rare entities, recent events not well-represented in training data, or specialized domains with limited available sources 610. This occurs because language models learn to generate plausible-sounding text based on statistical patterns, and when retrieval fails to find relevant sources, models may fabricate information that fits expected patterns. The challenge intensifies for emerging topics where authoritative sources may not yet exist or for niche domains where relevant documents are scattered across non-indexed sources.

Solution:

Implement explicit knowledge boundary detection that identifies when queries fall outside the system's reliable knowledge domain 712. Train classifiers to recognize low-confidence scenarios based on features like retrieval score distributions (when top results have low similarity scores), source scarcity (few retrieved documents), and domain classification (queries about topics underrepresented in the index). When these conditions are detected, systems should either refuse to generate responses, provide explicit uncertainty warnings, or pivot to alternative strategies like suggesting related queries where better information is available.

Example: A scientific research assistant detects a query about a protein discovered in a paper published last week. Retrieval returns only the single original paper with no corroborating sources. The system recognizes this as a high-hallucination-risk scenario and responds: "I found limited information about this protein in one recent publication [citation]. Because this is newly reported research without independent verification, I cannot provide a comprehensive summary. I recommend reviewing the original paper directly and checking for subsequent citations in the coming months as the research community evaluates these findings." This approach maintains user trust by acknowledging limitations rather than generating potentially inaccurate speculative content.

Challenge: Contradictory Information Across Sources

Real-world information sources frequently contain contradictions, whether due to evolving scientific understanding, differing methodological approaches, or genuine disagreement among experts 58. AI systems struggle to navigate these contradictions, often either selecting one source arbitrarily, attempting to synthesize incompatible claims, or failing to acknowledge the disagreement, all of which can mislead users.

Solution:

Implement contradiction detection and explicit disagreement surfacing that identifies when retrieved sources provide conflicting information and presents these contradictions transparently to users 510. Use NLI models to detect entailment, contradiction, and neutral relationships between claims from different sources. When contradictions are identified, generate responses that acknowledge the disagreement, present the different perspectives with their respective citations, and when possible, provide context about why sources might disagree (different time periods, methodologies, or theoretical frameworks).

Example: A health information system receives a query about the effectiveness of a particular dietary supplement. Retrieval finds that clinical trials show minimal effects while observational studies suggest benefits. The contradiction detection system identifies this discrepancy and generates: "Research on this supplement shows mixed results. Randomized controlled trials have found minimal effects on the measured outcomes [citation 1, citation 2], while observational studies report positive associations [citation 3, citation 4]. This discrepancy may reflect differences in study design, with observational studies potentially capturing long-term effects or confounding factors not present in shorter clinical trials. Consult with a healthcare provider to evaluate whether this supplement is appropriate for your specific situation." This response educates users about the evidence landscape rather than presenting false certainty.

Challenge: Citation Format and Style Variation

Different domains, publications, and organizations employ varying citation formats (APA, MLA, Chicago, legal citation styles, etc.), and users expect AI systems to follow appropriate conventions for their context 78. Systems must handle not only format variation but also domain-specific citation practices, such as legal citations that include subsequent history or scientific citations that specify figure or table numbers.

Solution:

Implement modular citation formatting systems with domain-specific templates and style guides that can be selected based on user preferences or automatic domain detection 34. Maintain structured citation metadata (authors, title, publication, date, DOI, page numbers) in a normalized format throughout the pipeline, then apply style-specific formatting only at the presentation layer. This separation allows the same underlying citation data to be rendered in multiple formats without requiring separate retrieval or verification processes.

Example: An academic writing assistant maintains citations in a structured JSON format internally: {"authors": ["Smith, J.", "Jones, M."], "year": 2023, "title": "...", "journal": "...", "volume": 45, "pages": "123-145", "doi": "..."}. When generating output for a psychology paper, it formats this as APA style: "Smith, J., & Jones, M. (2023). Title. Journal Name, 45, 123-145. https://doi.org/..." For a humanities paper, the same data renders as MLA: "Smith, John, and Mary Jones. 'Title.' Journal Name, vol. 45, 2023, pp. 123-145." Users can switch between styles instantly, and the system can export citations to reference management software in standard formats like BibTeX or RIS, providing flexibility without sacrificing citation accuracy.

Challenge: Computational Cost and Latency Constraints

Comprehensive verification involving dense retrieval over large corpora, cross-encoder re-ranking, multiple NLI model evaluations, and fact-checking against structured databases can require seconds or even minutes of computation per query 1911. This latency is unacceptable for interactive applications where users expect sub-second responses, creating tension between thoroughness and usability.

Solution:

Implement hierarchical and selective verification strategies that allocate computational resources based on claim importance, user context, and confidence levels 211. Use fast approximate methods for initial filtering and confidence estimation, reserving expensive verification for high-stakes claims or low-confidence outputs. Employ caching strategies that store verified responses for common queries, and implement progressive disclosure where initial responses appear quickly with basic citations, followed by enhanced verification that completes in the background.

Example: A legal research platform implements a three-tier verification strategy. Tier 1 (50ms budget) uses cached results for common queries and performs fast semantic similarity matching for citations. Tier 2 (500ms budget) adds cross-encoder re-ranking and basic citation validation for novel queries. Tier 3 (5000ms budget) includes comprehensive multi-source verification, precedent chain validation, and jurisdiction-specific checking, reserved for queries flagged as high-stakes (detected through keywords like "malpractice," "liability," "criminal") or when Tier 2 confidence scores fall below 0.8. Users see Tier 1 results immediately, with a progress indicator showing "Verifying citations..." as Tier 2 and 3 processing completes. This progressive approach provides immediate value while ensuring thorough verification for critical applications.

Challenge: Evaluation Dataset Quality and Coverage

Assessing citation accuracy and factual precision requires high-quality evaluation datasets with expert annotations, but creating such datasets is expensive and time-consuming 5612. Existing datasets like FEVER focus on specific domains or claim types, and may not generalize to specialized applications. Automated metrics can miss nuanced errors that human evaluators would catch, while human evaluation doesn't scale to continuous monitoring of production systems.

Solution:

Implement hybrid evaluation strategies that combine automated metrics for continuous monitoring with periodic expert human evaluation on representative samples 812. Develop domain-specific evaluation datasets through targeted annotation efforts focusing on high-value use cases and common error modes. Use active learning approaches to identify challenging cases where automated metrics and human judgments disagree, focusing annotation resources on these informative examples. Establish ongoing feedback loops where user corrections and expert reviews continuously improve evaluation datasets.

Example: A medical AI system implements a comprehensive evaluation framework. Automated metrics (citation precision, factual consistency scores, retrieval relevance) run on every query, with results logged to a monitoring dashboard that tracks trends over time. Weekly, the system samples 100 queries stratified by confidence scores and medical specialty, sending them to domain expert physicians for detailed review. Experts assess whether citations support claims, whether important caveats are included, and whether the overall response would be helpful in clinical decision-making. Cases where automated metrics indicated high confidence but experts identified errors are added to a "challenging cases" dataset used for model improvement. This hybrid approach provides both scalable continuous monitoring and deep quality assessment, enabling iterative system improvement while maintaining production quality standards.

References

  1. Guu, K., Lee, K., Tung, Z., Pasupat, P., & Chang, M. (2020). REALM: Retrieval-Augmented Language Model Pre-Training. arXiv. https://arxiv.org/abs/2005.11401
  2. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv. https://arxiv.org/abs/2005.11401
  3. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2021). WebGPT: Browser-assisted question-answering with human feedback. arXiv. https://arxiv.org/abs/2112.09332
  4. Menick, J., Trebacz, M., Mikulik, V., Aslanides, J., Song, F., Chadwick, M., Glaese, M., Young, S., Campbell-Gillingham, L., Irving, G., & McAleese, N. (2022). Teaching language models to support answers with verified quotes. arXiv. https://arxiv.org/abs/2203.11147
  5. Thorne, J., Vlachos, A., Christodoulopoulos, C., & Mittal, A. (2018). FEVER: a large-scale dataset for Fact Extraction and VERification. arXiv. https://arxiv.org/abs/1803.05355
  6. Krishna, K., Roy, A., & Iyyer, M. (2021). Hurdles to Progress in Long-form Question Answering. Proceedings of NAACL-HLT 2021. https://aclanthology.org/2021.naacl-main.149/
  7. Gao, T., Yen, H., Yu, J., & Chen, D. (2023). Enabling Large Language Models to Generate Text with Citations. arXiv. https://arxiv.org/abs/2305.14627
  8. Bohnet, B., Tran, V., Versley, Y., Hartung, M., Kuznetsov, I., Daxenberger, J., Eckle-Kohler, J., Klinger, R., Mieskes, M., Castelli, M., Chernodub, A., Cleuziou, G., Dias, G., Erbs, N., Ermakova, L., Fernández, R., Gienapp, L., Hagen, M., Hou, Y., ... & Gurevych, I. (2022). Attributed Question Answering: Evaluation and Modeling. arXiv. https://arxiv.org/abs/2301.10226
  9. Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., van den Driessche, G., Lespiau, J., Damoc, B., Clark, A., de Las Casas, D., Guy, A., Mensch, J., Hennigan, T., Reed, S., Glaese, M., Sifre, L., Cassirer, A., Brock, A., ... & Sifre, L. (2021). Improving language models by retrieving from trillions of tokens. Nature. https://www.nature.com/articles/s41586-021-03819-2
  10. Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. arXiv. https://arxiv.org/abs/2211.09110
  11. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Proceedings of NeurIPS 2020. https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html
  12. Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W., Koh, P. W., Iyyer, M., Zettlemoyer, L., & Hajishirzi, H. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. arXiv. https://arxiv.org/abs/2304.09848