CoChat
Sign Up
Cited by 164Topic ModelingNatural Language Processing TechniquesInformation Retrieval and Search Behavior

RARR: Researching and revising what language models say, using language models

Open access
Gao, Luyu, Dai, Zhuyun, Pasupat, Panupong, Chen, Anthony, Chaganty, Arun Tejasvi, Fan, Yicheng, Zhao, Vincent, Lao, Ni +3 more·2023·Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
DOI 10.18653/v1/2023.acl-long.910
Open this paper in CoChat — free accountOpen PDF

Abstract

Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, Kelvin Guu. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

AI snapshot

Written from the paper’s full text.

TL;DR

RARR is a model-agnostic research-and-revise system that retrieves web evidence for an LM-generated passage and post-edits unsupported content while preserving the original text. Across factoid QA, reasoning chains, and dialog, RARR improves attribution more robustly than prior edit models and preserves original intent over 90% of the time.

Key findings

  • On PaLM NQ outputs, RARR raises auto-AIS from 45.6 to 54.9 and human AIS from 35.4 to 43.4, with intent preservation 90.0, Levenshtein preservation 89.6, combined preservation 83.1, and F1_AP 57.0.
  • On PaLM SQA outputs, RARR raises auto-AIS from 37.6 to 45.1 and AIS from 24.5 to 31.5, with intent preservation 92.6, Levenshtein 89.9, combined 84.6, and F1_AP 45.9.
  • On LaMDA QReCC outputs, RARR raises auto-AIS from 18.8 to 29.4 and AIS from 13.2 to 28.3, with intent preservation 95.6, Levenshtein 80.2, combined 78.1, and F1_AP 41.5.
  • RARR is the only method reported to perform robustly across all three datasets; EFEC and LaMDA preserve original intent only 6–40% of the time, while RARR preserves intent over 90% of the time. Editing increases attribution by up to 13% absolute while changing only 10–20% of the text.

Methods

  • The paper formalizes Editing for Attribution: given a generated passage x, output a revised passage y plus an attribution report A of evidence snippets. It evaluates attribution with sentence-level AIS and an automated NLI-based auto-AIS, and preservation with human intent preservation, Levenshtein similarity, and their combination.
  • RARR uses few-shot prompting with PaLM 540B: comprehensive question generation, Google Search retrieval (K=5 pages, top J=1 evidence per query), an agreement model to detect disagreement, an edit model to revise only when needed, and selection of at most M=5 evidence snippets for the attribution report.
  • Experiments use 150 development and 150 test passages for each generation-model/source-dataset combination: PaLM and GPT-3 on NQ and SQA, and LaMDA and GPT-3 on QReCC. Baselines include EFEC and LaMDA Research; human evaluation and automated metrics are both reported.

Limitations

  • Stated: attribution and preservation may not deserve equal weight depending on the application; the metrics do not capture all aspects of attribution, such as self-evident sentences, varying assertion scope, directness, or implicatures; preservation metrics do not cover discourse or logical coherence, and heavily flawed inputs may require changes that the metrics penalize.
  • Stated: RARR is not fully general—it is not prepared for poetry or long documents without prompt changes—and it tends to preserve rather than delete claims it cannot attribute, even when those are hallucinations. It is also computationally costly because it prompts a large language model.
  • Evident from the text: the main experiments are based on a single run, though Appendix A reports variance over 3 random runs; the paper also notes that evidence trustworthiness and conflicting evidence are not addressed by the current implementation or evaluation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2023.acl-long.910

Cites

60 of 63 references verified0 citation errors · 3 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

28 of 31 cited sources checked · 19 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    The NLI model from Honovich et al. (2022) correlates well with AIS scores.

    “We utilize the natural language inference (NLI) model from Honovich et al. (2022), which correlates well with AIS scores.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Honovich et al. (2022) source

      Section 3.3 and Appendix B describe an NLI model trained by Honovich et al. (2022), supporting the model identification. Section 4 reports its performance using ROC AUC on factual-consistency datasets, but the copy does not report a correlation with AIS scores (searched: AIS scores, Attributable to Identified Sources). (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    State-of-the-art LMs still frequently produce biased, misleading, or unsupported content, colloquially called 'hallucinations'.

    “Despite these incredible advances, state-of-theart LMs still frequently produce biased, misleading, or unsupported content, colloquially called”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Maynez et al., 2020 source

      Section 2.1 defines hallucinations as summary content unsupported by or misrepresenting the source document, and the assessment in Section 5.2 finds substantial hallucinations in the summarization systems studied. The paper does not establish that language models generally frequently produce biased content; its findings concern abstractive summarization. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites Menick et al., 2022 source

      The opening passage (p. 1) says large language models can hallucinate convincing nonsense, and the Conclusion (p. 18) says they often produce hallucinated facts. The Limitations section (p. 18) discusses misleading, cherry-picked quotations. The paper does not establish that language models frequently produce biased content; its discussion of bias concerns supporting sources such as Wikipedia. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Premisecitation

    Even retrieval-augmented models still do not guarantee attribution.

    “Even retrievalaugmented models (Guu et al., 2020;Lewis et al., 2020), which retrieve relevant documents and then condition on them to generate text, still do not guarantee attribution.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Guu et al., 2020 source
    • cites Lewis et al., 2020 source
  • Premisecitation

    In real-world scenarios such as the one RARR operates in, relevant evidence may not be provided, necessitating retrieval.

    “In real-world scenarios such as the one which RARR operates in, relevant evidence may not be provided, necessitating retrieval (Fan et al., 2020;Piktus et al., 2021).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Fan et al., 2020 source
    • cites Piktus et al., 2021 source
  • Premisecitation

    Even if a revision preserves intent, it may still make superfluous modifications, such as reordering words, changing textual style, or including unnecessary additional information.

    “However, even if a revision preserves intent, it may still make superfluous modifications, such as reordering words, changing textual style, or including unnecessary additional information (Thorne and Vlachos, 2021).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Thorne and Vlachos, 2021 source

      Section R2—Supported by Evidence says corrections should target errors in the original claim rather than introduce unrelated information; Section 7 Evaluation also checks whether unrelated information was added. These passages support the concern about unnecessary additions, but the copy does not discuss intent-preserving revisions that needlessly reorder words or change textual style. No page numbers are available in the indexed passages. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • May combine two claims. 31
  • Classified as probably not a substantial claim. 36

  • May combine two claims.

    Premisecitation

    Prior work has shown that retrieval-augmented models generate text that either includes additional information outside the retrieved documents, ignores the documents altogether, or even contradicts the documents.

    “Prior work has shown that retrieval-augmented models generate text that either includes additional information outside the retrieved documents (Dziri et al., 2022), ignores the documents altogether (Krishna et al., 2021), or even contradicts the documents (Longpre et al., 2021).”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites Dziri et al., 2022 source
    • cites Krishna et al., 2021 source
    • cites Longpre et al., 2021 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    LaMDA generates responses in three steps: generate a base response; generate search queries from the base response; generate a revised response conditioned on the base response and retrieved evidence.

    “LaMDA (Thoppilan et al., 2022) generates responses in three steps: 1) generate a “base response”; 2) generate search queries from the base response; 3) generate a “revised response” conditioned on the base response and retrieved evidence.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Thoppilan et al., 2022 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    EFEC uses a neural retrieval model to retrieve from Wikipedia.

    “EFEC uses a neural retrieval model (Karpukhin et al., 2020) to retrieve from Wikipedia”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Karpukhin et al., 2020 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    FRUIT implements an editor that is fine-tuned on Wikipedia edit history with the goal of updating outdated information.

    “FRUIT (Logan IV et al., 2022) and PEER (Schick et al., 2022) both implement an editor that is fine-tuned on Wikipedia edit history with the goal of updating outdated information and collaborative writing respectively.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Logan IV et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Schick et al., 2022 source

  • Classified as probably not a substantial claim.

    Premisecitation

    Language models can generate reasoning chains to answer complex questions.

    “Language models can generate reasoning chains to answer complex questions (Wei et al., 2022).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Wei et al., 2022 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A strategy similar to comprehensive question generation has been employed to train text-planning models.

    “A similar strategy has been employed to train textplanning models (Narayan et al., 2022).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Narayan et al., 2022 source

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Rashkin et al. (2021) proposed Attributable to Identified Sources (AIS), a human evaluation framework which considers a binary notion of attribution.

    “Previously, Rashkin et al. (2021) proposed Attributable to Identified Sources (AIS), a human evaluation framework which considers a binary notion of attribution.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Rashkin et al. (2021) source

  • Classified as probably not a substantial claim.

    Premisecitation

    The approach is inspired by works on fact-checking where simple research-and-revise workflows are effective at attributing or correcting unattributed claims made by humans.

    “The approach is inspired by works on fact-checking 2 where simple research-and-revise workflows are effective at attributing or correcting unattributed claims made by humans (Thorne et al., 2018;Schuster et al., 2021;Thorne and Vlachos, 2021).”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites Thorne et al., 2018 source

      Section 3.2 (Task 2—Claim Labeling; page not supplied) describes annotators finding Wikipedia evidence to support or refute human-generated claims. This supports citing FEVER as fact-checking work, but the paper does not establish that a research-and-revise workflow is effective at attributing or correcting unattributed human claims. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites Schuster et al., 2021 source

      Section 4.1 (page unavailable) says Wikipedia users verify and correct others’ revisions, and §3.1 describes collecting factual edits. The paper does not establish that simple research-and-revise workflows are effective, or address attributing previously unattributed human claims. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites Thorne and Vlachos, 2021 source

      The Retrieved Evidence section reports that the system retrieves evidence and uses it to correct claims, with satisfactory corrections for some evaluated instances. The Conclusions and Future Work section says the evaluation used mutations introduced by human annotators and may not represent errors in real-world documents. The paper does not demonstrate attribution of unattributed claims, so that part and the broader implication about human-made claims are not supported; page numbers are unavailable in our copy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    EFEC fine-tunes a T5-based model to revise text conditioned on multiple evidence snippets using both semi-supervised and fully-supervised approaches.

    “EFEC fine-tunes a T5-based model to revise text conditioned on multiple evidence snippets using both semi-supervised and fully-supervised approaches.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Thorne and Vlachos, 2021 source

      Section 5.3 says the paper fine-tunes a T5 model to generate corrections from masked claims and evidence, and Section 4.1 specifies multiple retrieved evidence sentences. The paper calls its main training approach “distant supervision,” not semi-supervision; it also fine-tunes a separate fully supervised T5 model as a performance ceiling (Implementation, “Fully Supervised Ceiling”). (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    EFEC's fully-supervised approach performed best in their experiments.

    “We compare against their fully-supervised approach, which performed best in their experiments.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Thorne and Vlachos, 2021 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Question-guided agreement between a passage and evidence was previously explored by Honovich et al. (2021).

    “This form of question-guided agreement was previously explored by Honovich et al. (2021).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Honovich et al., 2021 source

  • May combine two claims.

    Premisecitation

    LaMDA is a language model for dialog that performs revision by training on human annotations.

    “An exception is LaMDA (Thoppilan et al., 2022), a language model for dialog that performs revision by training on human annotations.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Thoppilan et al., 2022 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    PEER implements an editor that is fine-tuned on Wikipedia edit history with the goal of collaborative writing.

    “FRUIT (Logan IV et al., 2022) and PEER (Schick et al., 2022) both implement an editor that is fine-tuned on Wikipedia edit history with the goal of updating outdated information and collaborative writing respectively.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Logan IV et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Schick et al., 2022 source

  • May combine two claims.

    Findingown data

    RARR significantly improves attribution while otherwise preserving the original input to a much greater degree than previously explored edit models.

    “When applied to the output of several state-of-the-art LMs on a diverse set of generation tasks, we find that RARR significantly improves attribution while otherwise preserving the original input to a much greater degree than previously explored edit models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingown data

    Results on GPT-3 passages show similar trends.

    “Results on GPT-3 passages show similar trends (Appendix A).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    In the Georgia/California peach example, RARR using CQGen needs to sacrifice more preservation score to edit the text, leading to a lower F1 AP score.

    “In this case, RARR using CQGen needs to sacrifice more preservation score to edit the text, leading to a lower F1 AP score.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In terms of F1 AP, RARR is the only method that performs robustly across all three datasets, and significantly outperforms prior methods on NQ and SQA.

    “In terms of F1 AP , RARR is the only method that performs robustly across all three datasets, and significantly outperforms prior methods on NQ and SQA.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Excluding all of Wikipedia from web search results causes the attribution performance of sentences-as-queries to drop significantly, while CQGen is more robust.

    “The attribution performance of sentences-as-queries drops significantly, while CQGen is more robust.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The answers generated by LaMDA and GPT-3 for QReCC tend to be context-dependent, featuring pronouns and implicit references.

    “The answer tends to be context-dependent, featuring pronouns and implicit references.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    After editing, RARR can increase attribution by up to 13% absolute, while changing only 10-20% of the text.

    “After editing, RARR can increase attribution by up to 13% absolute, while changing only 10-20% of the text.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    More revision leads to less preservation score and spurious changes to the text passage.

    “As expected, more revision leads to less preservation score and spurious changes to the text passage, as demonstrated in Figure 8.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    RARR not only preserves short answer accuracy on NQ but actually improves it by roughly 5%.

    “We find that RARR not only preserves the short answer accuracy but actually improves it by roughly 5%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    RARR significantly improves attribution while preserving most of the original text.

    “RARR significantly improves attribution while preserving most of the original text.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    In datasets where the claim contains multiple sentences (NQ and SQA), EFEC's behavior yields low preservation scores and results in outputs that are less informative.

    “However, in datasets where the claim contains multiple sentences (NQ and SQA), such a behavior yields low preservation scores, and also results in outputs that are less informative.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    Not all passages in the experiments are supported by Wikipedia articles.

    “however, not all passages in our experiments are supported by Wikipedia articles.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    EFEC reduces the passage x into a single sentence.

    “EFEC reduces the passage x into a single sentence.”
    • Also stated in: § E Details on the dataset · plain
      “EFEC reduces the passage x into a single sentence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Findingnone given

    LaMDA changes the writing style.

    “LaMDA changes the writing style.”
    • Also stated in: § E Details on the dataset · plain
      “LaMDA changes the writing style.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    EFEC frequently attempts to summarize the entire passage into one sentence, or drops later sentences.

    “EFEC frequently attempts to summarize the entire passage into one sentence, or drops later sentences.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Editing is crucial to improve attribution: if only retrieve evidence to support the original response x without editing, attribution ranges from the low 10s to mid 30s.

    “We also see that editing is crucial to improve attribution: if we only retrieve evidence to support the original response x without editing, attribution ranges from the low 10s to mid 30s.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The generated long-form passages from PaLM and GPT-3 on NQ are mostly coherent but often contain factual errors.

    “The resulting passages are mostly coherent but often contain factual errors.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The design of RARR allows for the research stage to restrict its search over a user-specified corpus, based on what the user deems trustworthy.

    “While we do not provide a means for judging trustworthiness, the design of RARR does allow for the research stage to restrict its search over a userspecified corpus, based on what the user deems trustworthy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    RARR tends to preserve rather than delete claims that it cannot attribute.

    “Second, RARR tends to preserve rather than delete claims that it cannot attribute.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    RARR's performance is low on examples with mathematical reasoning, as these are beyond the capability of the edit model with the current prompt.

    “We also found that RARR's performance is low on examples with mathematical reasoning, as these are beyond the capability of the edit model with our current prompt.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Sometimes the whole output is based on a false premise and needs to be completely rewritten, in which case RARR cannot satisfactorily edit due to the revision threshold.

    “Sometimes the whole output is based on a false premise and needs to be completely rewritten, in which case RARR cannot satisfactorily edit due to our revision threshold (Section 3.2).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    AIS scores from human correlate well with auto-AIS scores, with some bias for non-attributed sentences to be judged as attributed by auto-AIS.

    “We found that the AIS scores from human correlate well with auto-AIS scores, with some bias for non-attributed sentences to be judged as attributed by auto-AIS.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    When RARR is not 100% successful in making text consistent with retrieved evidence, the revised text will be partially attributed.

    “When RARR is not 100% successful in making text consistent with retrieved evidence, the revised text will be partially attributed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    The model is computationally costly, since it is based on prompting a large language model.

    “Finally, our model is computationally costly, since it is based on prompting a large language model.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The editor occasionally produces large edits that bring the new revision close to e but far from the current y.

    “The editor occasionally produces large edits that bring the new revision close to e but far from the current y.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    EFEC and LaMDA face the issue of low attribution on news summaries but instead trade preservation for attribution and rewrite the text to a different topic.

    “EFEC and LaMDA also face this issue, but instead trade preservation for attribution and rewrite the text to a different topic.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    RARR improves attribution of the explanations on all four categories of MMLU, although the increases are relatively small.

    “proves attribution of the explanations on all four categories of MMLU, although the increases are relatively small.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Human and automatic metrics have strong Pearson correlation (attribution = 0.74; preservation = 0.62).

    “We conducted correlation studies between human and automatic metrics and found strong Pearson correlation (attribution = 0.74; preservation = 0.62).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given

    This work does not address the question of whether a source is trustworthy, or the related topic of misinformation.

    “This work does not address the question of whether a source is trustworthy, or the related topic of misinformation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The standard deviations of Attr auto, Pres Lev, and F1 AP across 3 random runs of RARR on PaLM outputs on NQ are 1.2, 0.5, and 1.0 respectively.

    “The standard deviations of Attr auto , Pres Lev , and F1 AP are 1.2, 0.5, and 1.0 respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Occasionally ignoring the retrievals can make the models more robust to bad retrievals, illustrating that end-task performance and attribution are not always aligned.

    “In fact, occasionally ignoring the retrievals can make the models more robust to bad retrievals (Khandelwal et al., 2020), illustrating that end-task performance and attribution are not always aligned.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Khandelwal et al., 2020

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The implementation of RARR requires only a handful of training examples, a large language model, and standard web search.

    “Furthermore, the implementation of RARR requires only a handful of training examples, a large language model, and standard web search.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The main experiments in Section 5 are based on a single run.

    “The main experiments in Section 5 are based on a single run.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Asking the edit model to first identify a particular span in y that needs to be edited before generating the revised y helps reduce the editor's deviation from the current y.

    “This helps reduce the editor's deviation from the current y.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    EFEC's summarizing behavior generally increases the attribution score because it is usually easier to make one sentence fully attributable than many sentences.

    “This behavior generally increases the attribution score, because it is usually easier to make one sentence fully attributable than many sentences.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    A user cannot easily determine whether their outputs are trustworthy or not, because most LMs do not have any built-in mechanism for attribution to external evidence.

    “A user cannot easily determine whether their outputs are trustworthy or not, because most LMs do not have any built-in mechanism for attribution to external evidence.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Most existing LMs lack a built-in mechanism for attribution.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The slight cost in SQA downstream task performance may be due to noisy retrievals, which sometimes provide misleading evidence.

    “We suspect that this may be due to noisy retrievals, which sometimes provide misleading evidence (exemplified in Figure 7d).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    To more fairly compare the editing capabilities of EFEC, the authors used evidence retrieved by their research stages (CQGen and web search).

    “To more fairly compare the editing capabilities of EFEC, we instead use the evidence retrieved by our research stages (CQGen and web search).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The authors have manually looked through most of the data and found no personal identifiers.

    “The authors have manually looked through most of the data and found no personal identifiers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    In the main paper, two automatic metrics are presented to gauge attribution and preservation, but human evaluation is relied on as the gold standard.

    “In the main paper, we present two automatic metrics to conveniently gauge these properties, but rely on a human evaluation as the gold standard.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    We reject an editing if the edit distance is more than 50 characters or more than half of the original text length.

    “We reject an editing if the edit distance is more than 50 characters or more than half of the original text length.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Three annotators rated each example in the pilot phase to measure inter-annotator agreement, and a single rater annotated each example afterwards.

    “We had 3 annotators rate each example in the pilot phase to measure inter-annotator agreement, and had a single rater annotate each example afterwards.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    We tune our hyperparameters on the validation set as well.

    “We tune our hyperparameters on the validation set as well.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    News article summaries are often about less widely known people and events, which is challenging for retrievers, leading to low attribution.

    “These news article summaries are often about less widely known people and events, which is challenging for retrievers, leading to low attribution.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Using web search with standard question generation methods may fail to capture important context from the input, and is not sufficient for the attribution task.

    “This result suggests that using web search with standard question generation methods may fail to capture important context from the input, and is not sufficient for the attribution task.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    We manually wrote our prompts by eye-balling quality on a dozen of examples from a separate validation set.

    “We manually wrote our prompts by eye-balling quality on a dozen of examples from a separate validation set.”
    Quote in the paperNot a substantial claim
59 of 59

  • The paper relies on this without support.

    Most existing LMs lack a built-in mechanism for attribution.

    “Most existing LMs, such as those based on sequence-to-sequence architectures, lack a builtin mechanism for attribution.”

  • The paper relies on this without support.

    Pres intent requires human annotation.

    “Since Pres intent requires human annotation, we use Pres Lev as an automated metric for model development.”

  • The paper relies on this without support.

    Gold or reference edits are often only available for specialized domains.

    “which are often only available for specialized domains.”

  • The paper relies on this without support.

    After retrieving evidence, some parts of the passage may be properly attributed, but other parts remain unattributed and should be revised.

    “After retrieving evidence, certain parts of x may now be properly attributed, but other parts remain unattributed and should be revised.”

  • The paper relies on this without support.

    A prompt with six human demonstrations is sufficient for PaLM to adequately learn the comprehensive question generation task.

    “A prompt with six human demonstrations was sufficient for PaLM to adequately learn the task.”

  • The paper relies on this without support.

    LaMDA was not optimized for preservation.

    “since LaMDA was not optimized for preservation.”

  • The paper relies on this without support.

    EFEC's training data was limited to single sentences.

    “which was limited to single sentences.”

  • The paper relies on this without support.

    The 40 manually inspected examples are representative of the question generation model's coverage.

    “a manual inspection of 40 examples shows > 80% with questions that fully cover all aspects of the input text.”

  • The authors state this as an assumption.

    Manually editing examples until judged 100% attributable approximates the minimal amount of editing needed to make a text passage fully attributed.

    “As an approximation, we manually edited 30 examples in our NQ benchmark until we judged them to be 100% attributable.”

  • The paper relies on this without support.

    Attribution and preservation are both subjective properties that may change with even small edits.

    “Attribution and preservation are both subjective properties that may change with even small edits.”

  • The paper relies on this without support.

    There is currently no linguistic consensus on what constitutes a claim.

    “While each sentence may contain multiple claims that could be attributed independently, there is currently no linguistic consensus on what constitutes a claim.”

  • The paper relies on this without support.

    Utterances can make assertions with varying levels of directness.

    “In addition to varying scope, utterances can also make assertions with varying levels of directness.”

  • The paper relies on this without support.

    Some content may be simultaneously supported by certain sources, while contradicted by others.

    “There is also the possibility that some content may be simultaneously supported by certain sources, while contradicted by others.”

  • The paper relies on this without support.

    Few-shot prompting has proven to be effective for many recent large language models.

    “Few-shot prompting has proven to be effective for many recent large language models.”
14 of 14

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Decomposition of sentences into atomic facts is not necessary to achieve MiniCheck's high performance.

    “Furthermore, we can do this without a separate claim decomposition step.”

    From MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents (2024) · cites this paper as Gao et al., 2023

    In “B Details on automated evaluation” (page not available), RARR says it uses sentences as claims rather than independently decomposing each sentence’s claims. That supports the feasibility of omitting a separate claim-decomposition step in RARR’s evaluation, but the paper does not establish that decomposition is unnecessary for MiniCheck’s high performance. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • Gao et al. (2023), Min et al. (2023), and Wang et al. (2023) decompose each generated sentence into atomic facts and then search for relevant documents to support each atomic fact.

    “Gao et al. (2023); Min et al. (2023); Wang et al. (2023) decompose each generated sentence into atomic facts and then search for relevant documents to support each atomic fact.”

    From MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents (2024) · cites this paper as Gao et al. (2023)

    The Query generation section says RARR generates questions covering aspects of a passage and retrieves web evidence for each question. It does not say RARR decomposes each sentence into atomic facts; Appendix B instead says it uses sentences as claims for evaluation. (LLM, conf=0.88)

    Experimental — accuracy not yet measured

  • Recent studies highlight the potential of LLMs to detect factual content in texts.

    “Recent studies (Gao et al. 2023; Chern et al. 2023) highlight the potential of LLMs to detect factual content in texts.”

    From A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions (2025) · cites this paper as Gao et al. 2023

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Reported maximum attribution gain is smaller than a table result

    Minor

    The paper says RARR increases attribution by up to 13 percentage points, but Table 1 reports an increase from 13.2 to 28.3 on QReCC-15.1 percentage points.

  • "Over 90%" intent preservation includes an exactly 90% result

    Minor

    The claim applies to RARR across the reported tasks, but Table 1 gives its NQ intent-preservation result as 90.0%, not over 90%.

  • Query-generation model size is given as both 64B and 62B

    Minor

    The appendix describes shrinking query generation to 64B, whereas the corresponding Table 4 configuration labels it 62B.

  • Human AIS metric is called automated

    Minor

    Attr AIS is defined using annotator judgments; the paper separately names Attr auto as its automated approximation.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Jingyi Zhang, Masao Utiyama, Eiichiro Sumita, Gra- ham Neubig, and Satoshi Nakamura. 2018. Guiding neural machine translation with retrieved translation pieces. In NAACL. · DOI

Cited by

Sorted by citations

Related

Authors
Gao, L., Dai, Z., Pasupat, P., Chen, A., Chaganty, A. T., Fan, Y., Zhao, V., Lao, N., Lee, H., Juan, D.-C., & Guu, K.
Affiliations
  • Carnegie Mellon University (US)
  • Google (United States) (US)
  • University of California, Irvine (US)
ORCID
Venue
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Publisher
Association for Computational Linguistics
Published
2023
Pages
16477-16508
ISSN
0736-587X (ISSN-L)
DOI
10.18653/v1/2023.acl-long.910
OpenAlex
W4385570481
Cited by
141
References
57
Citations by year
2026: 61 · 2025: 43 · 2024: 21 · 2023: 16
FWCI
25.00
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
Proceedings of the conference - Association for Computational Linguistics. Meeting (publishedVersion, cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
Keywords
language models, large language models, research question formulation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-29.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-29.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review