Pith. sign in

REVIEW 4 major objections 5 minor 32 references

LAQuer: Localized Attribution Queries in Content-grounded Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LAQuer reframes attribution as a user-initiated query: highlight any span in a generated answer and the system returns the concise source spans that support the highlighted claim's decontextualized meaning.

desk verdict A worthwhile new task and benchmark for span-level attribution, with a strong length-reduction result, but the synthetic query pipeline needs validation before I'd trust the accuracy numbers. read the letter →

arxiv 2506.01187 v1 pith:SXX6R2EY submitted 2025-06-01 cs.CL

Eran Hirsch , Aviv Slobodkin , David Wan , Elias Stengel-Eskin , Mohit Bansal , Ido Dagan This is my paper · ORCID
classification cs.CL
keywords LAQuerlocalizedattributionuser-initiateddecontextualizationattributedtextgenerationmulti-documentsummarizationlong-formquestionansweringspanalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces Localized Attribution Queries (LAQuer), a task in which a user highlights a specific span of a machine-generated answer and the system returns the few source-document spans that support that exact claim, rather than the whole sentence or whole document. The authors argue that fixed sentence-level or document-level citations force readers to check far more text than they need, and that attribution should be user-initiated and span-level on both the output and source sides. They build a two-stage framework that first decontextualizes the highlighted span into a stand-alone claim and then finds concise source spans entailing it, using either an LLM prompt or hidden-state similarities. Across multi-document summarization and long-form QA, LAQuer reduces attributed text length by two orders of magnitude for plain generation (from 1,681.6 to 32.0 content words in the MDS setting) while keeping roughly two-thirds of attributions correct, though decontextualized facts remain hard.

What carries the argument

The central object is the LAQuer task itself, formalized as the requirement that the concatenated source spans entail the decontextualized meaning of the highlighted output spans, written $\mathrm{concat}(s_1,\ldots,s_m) \models I(o_1,\ldots,o_n|o)$. The framework carries out the task in two steps: (A) decontextualization, which reformulates the highlight into a stand-alone claim using the MolecularFacts approach, and (B) query-focused attribution, which finds verbatim source spans supporting that claim, implemented either by prompting an LLM or by cosine-similarity matching over hidden states of an open model. The evaluation machinery is a synthetic pipeline that decomposes outputs into atomic facts with FActScore and aligns them back to spans lexically, simulating what a user would highlight.

What would settle it

Run a user study in which people highlight spans they want fact-checked on the same generated outputs, feed those real highlights into the LAQuer pipeline, and compare AutoAIS and attributed length against the synthetic-highlight numbers in Table 4; a decisive variant would replace AutoAIS with human-annotated gold source spans on a random sample to check exact-match coverage.

Watch

Extended reading notes

Core claim

The central claim is that attribution can be made user-directed and localized on both the output side and the source side without a large loss in correctness, and that the right formulation is to require the concatenated source spans to entail the decontextualized meaning of the highlighted output spans. The paper proposes the LAQuer task, a two-stage modeling framework, and a synthetic evaluation protocol; its headline finding is that LAQuer methods cut the length of attributed text from thousands of characters to tens of content words, and that prompting an LLM beats using hidden-state internal similarities for the query-focused attribution step. The authors also find that when the generation method already produces localized source spans (Attr.First), LAQuer leverages that to give shorter attributions, but decontextualized facts are often not supported by the originally selected spans, which they attribute to the generation method's failure to decontextualize its own attributions.

Load-bearing premise

The entire benchmark assumes that the synthetic highlights produced by decomposing sentences into atomic facts and aligning them lexically behave like real users' highlights, in granularity, wording, and the kinds of claims they would fact-check.

Editorial extensions

If this is right

  • Interactive fact-checking becomes practical: a user who clicks or highlights a phrase reads tens of content words instead of scanning whole retrieved documents or full sentences.
  • LAQuer reduces attributed text length by two orders of magnitude for plain (Vanilla) and ALCE-generated outputs, and by about 59% on average for Attr.First outputs.
  • Prompting an LLM for the query-focused attribution step outperforms using hidden-state internals across all generation methods and both tasks.
  • When the underlying generator already localizes source spans (Attr.First), LAQuer yields the shortest attributions, but decontextualized facts get lower support scores, indicating that generation methods should decontextualize the information in their own attributions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Real-user highlights may be coarser or paraphrase around the highlighted words, so the synthetic pipeline in this paper could overstate both accuracy and the length reduction; a user study is the natural stress test.
  • The two-stage design separates decontextualization from attribution, so a better decontextualizer could improve the decontextualized AutoAIS numbers without changing the attribution step, making the reported decontextualization failure a target for modular improvement.
  • Because the internals method fails mainly when the source is a full document, combining hidden-state matching with a prior retrieval-narrowing step (such as Attr.First metadata) may close the gap to the prompt method at lower cost than LLM prompting.
  • The framework is not limited to grounded generation: post-hoc retrieval after generation could extend LAQuer to parametric-knowledge outputs, turning any model answer into a verifiable document.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces LAQuer, a task in which a user highlights an output span of a grounded generation and the system returns the concise source span(s) that support the decontextualized meaning of that highlight. The authors propose a two-stage framework: (A) decontextualize the highlighted spans into a standalone fact, and (B) perform query-focused attribution via an LLM prompt or via hidden-state internals. They evaluate three generation baselines (Vanilla, ALCE, Attr. First) on Multi-Document Summarization (SPARK) and Long-Form QA (Liu et al.), using a synthetic pipeline (FActScore decomposition with GPT-4o plus lexical alignment) to simulate user highlights. The headline finding is that LAQuer reduces attributed-text length from 1681.6 to 32.0 content words in the Vanilla MDS setting (Table 4), and that the LLM-prompt method outperforms the internals method across all settings.

Significance. If validated, LAQuer addresses a genuine usability problem in attributed generation: sentence- or document-level citations are too coarse for fact-checking specific claims. The framework is simple, the code is released, and the paper provides a reproducible evaluation pipeline. The length-reduction result is large and consistent, and the observation that decontextualization is a major failure point for current attribution methods is useful for the community. However, the benchmark's validity rests on the assumption that GPT-4o-generated atomic facts and lexical alignments faithfully mimic real user highlights, and the only human evaluation is a small author-run study. The significance of the accuracy claims is therefore currently bounded by this external-validity gap.

major comments (4)
  1. [Section 5.1, Appendix E] The synthetic LAQuer pipeline, which uses FActScore decomposition with GPT-4o and the Levenshtein edit-script aligner, is never validated against real user highlight behavior. Since all Table 4 accuracy and length numbers are computed on these synthetic queries, the claim that LAQuer is 'attractive' for actual users is not yet established. I would like to see either a small human-collected highlight study or a distributional analysis comparing synthetic facts/highlights to human span choices; without this, the benchmark's external validity is a load-bearing gap.
  2. [Section 6.2, Table 5] The human evaluation is described as 'conducted by the authors,' covers only 20 examples per condition, reports no inter-annotator agreement, and does not collect human highlights. This is too thin to serve as validation of the AutoAIS metric or to support the paper's claims about decontextualized-fact difficulty. Independent annotators, agreement statistics, and ideally human-collected highlights would be needed.
  3. [Section 4.2, Table 4] The comparison between LLM Prompt and LLM Internals is confounded: the prompt method uses GPT-4o, while the internals method uses Llama-3.1-8B-Instruct. The observed superiority of the prompt method may reflect model scale, instruction-following ability, or access to better in-context reasoning rather than the alignment strategy itself. A matched-model comparison (e.g., prompting Llama-3.1-8B as well, or running internals on a model of comparable capability) is needed to support the paper's 'best-performing LAQuer method' claim.
  4. [Section 5.3, Table 9] The evaluation uses 65 MDS and 45 LFQA test instances, with ten facts sampled per instance; the fact-level observations are not independent, and the paper reports no significance tests for the AutoAIS differences in Table 4. For the headline length-reduction result this is less concerning because the effect is large and consistent, but the accuracy comparisons (e.g., 62.5 vs. 18.0 for Vanilla MDS, or 71.5 vs. 28.6 for Attr. First MDS) would benefit from bootstrap confidence intervals or per-instance significance testing.
minor comments (5)
  1. [Section 5.1] The text contains a duplicated word: 'long-form QA setting setting (LFQA).'
  2. [Section 6.1] 'degredation' should be 'degradation.'
  3. [Table 4 caption] 'contextualizd' should be 'contextualized.'
  4. [Section 6.4] Attribution lengths in Table 6 are reported in characters, while Table 4 reports content words; please make the units consistent or state explicitly why characters are appropriate for the SPARK-based estimate.
  5. [Figure 3] The span-type categories are assigned via SpaCy heuristics, but no examples of misclassification or a reliability check are provided; a small validation of the categorization would strengthen the claim that the synthesized facts are diverse.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the LAQuer benchmark compares methods empirically, and the length-reduction and accuracy claims are not forced by construction or by self-citation.

full rationale

The paper proposes a new task and benchmarks existing or reimplemented methods, so there is no derivation chain whose output is equivalent to its own input. The central quantitative claim that LAQuer reduces attributed length (e.g., from 1681.6 to 32.0 content words for Vanilla MDS in Table 4) is measured from system outputs; although the prompt tells the model to keep spans short, the method could still fail to find short supporting spans, and the trade-off with accuracy is an empirical result rather than a construction. The decontextualized-fact evaluation in Section 5.3 and Appendix B uses GPT-4o with the MolecularFacts prompt to create decontextualized targets, and GPT-4o also powers the LLM Prompt attribution method; this is a same-model external-validity concern, not a reduction that forces the reported scores, especially since the human analysis in Section 6.2 reports similar trends. The use of the co-authored Attr. First method as one of three generation baselines is a self-citation, but it is not load-bearing: the main findings also hold for the Vanilla and ALCE settings, which do not depend on the authors' prior work. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and the FActScore and MolecularFacts prompts are external prior work. The Limitations section explicitly acknowledges that AutoAIS is an LLM-based metric and that the human evaluation is small-scale; these are stated limitations, not evidence of circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters or invented entities are introduced. The central claim rests on the synthetic query-generation pipeline, the AutoAIS metric, and the decontextualization step, each of which is a domain assumption rather than a fitted parameter. Hyperparameters for the internals method are adopted from Phukan et al. (2024), not re-fitted here.

assumptions (4)
  • domain assumption The FActScore-style LLM decomposition of output sentences into atomic facts, followed by lexical alignment, produces query highlights that faithfully simulate real user fact-checking interests.
    Invoked in Section 5.1 and Appendix E; the entire benchmark's query distribution depends on this proxy, but real users may highlight at different granularities or with different intents.
  • domain assumption AutoAIS using T5-XXL TRUE NLI is a valid automatic proxy for human attribution judgments in this setting.
    Used as the primary citation-quality metric in Section 5.3; the authors acknowledge this in the Limitations section and provide only a small human check.
  • domain assumption The decontextualization prompt from MolecularFacts applied with GPT-4o produces the correct decontextualized meaning of the user's highlights.
    Step (A) of the framework, Section 4.1; if decontextualization is wrong, the subsequent attribution may be correct for the wrong claim.
  • domain assumption The source-grounded generation outputs are reliable enough to evaluate attribution on, and factual errors in the generated output are not separated from attribution errors.
    Implicit in Sections 5 and 6; the benchmark attributes generated text that may itself contain unsupported claims, so attribution quality conflates generation errors with alignment errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LAQuer: Localized Attribution Queries in Content-grounded Generation." pith.science (2026). https://pith.science/paper/SXX6R2EY

@misc{pith2026250601187,
  author       = {Pith},
  title        = {Pith review of: LAQuer: Localized Attribution Queries in Content-grounded Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SXX6R2EY}},
  note         = {Machine review of arXiv:2506.01187}
}
read the original abstract

Grounded text generation models often produce content that deviates from their source material, requiring user verification to ensure accuracy. Existing attribution methods associate entire sentences with source documents, which can be overwhelming for users seeking to fact-check specific claims. In contrast, existing sub-sentence attribution methods may be more precise but fail to align with users' interests. In light of these limitations, we introduce Localized Attribution Queries (LAQuer), a new task that localizes selected spans of generated output to their corresponding source spans, allowing fine-grained and user-directed attribution. We compare two approaches for the LAQuer task, including prompting large language models (LLMs) and leveraging LLM internal representations. We then explore a modeling framework that extends existing attributed text generation methods to LAQuer. We evaluate this framework across two grounded text generation tasks: Multi-document Summarization (MDS) and Long-form Question Answering (LFQA). Our findings show that LAQuer methods significantly reduce the length of the attributed text. Our contributions include: (1) proposing the LAQuer task to enhance attribution usability, (2) suggesting a modeling framework and benchmarking multiple baselines, and (3) proposing a new evaluation setting to promote future research on localized attribution in content-grounded generation.

Figures

Figures reproduced from arXiv: 2506.01187 by the authors.

Figure 1
Figure 1. Top: example RAG scenario. Bottom: our Localized Attribution Queries (LAQuer), where the at￾tribution is constructed per user query, highlighted in yellow. Existing sentence-level attribution methods, un￾derlined in green, can often be disorienting and lengthy. Yet, model outputs frequently diverge from these sources, resulting in factual inaccuracies, or ‘hal￾lucinations’ (Mishra et al., 2024). To address this, use… view at source ↗
Figure 2
Figure 2. Overview of our LAQuer framework. The top section illustrates the generation of an output based on [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of span types based on syntactic [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Example prompt for LLM-based post-hoc alignment. The instructions are depicted in green, input to the model in black, and model’s output in red. This example is one of three few-shot examples. The source texts of the few-shot examples are adapated based on the generati…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 7 canonical work pages

  1. [1]

    Satanjeev Banerjee and Alon Lavie. 2005. METEOR : An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization , pages 65--72, Ann Arbor, Michigan. Association for Computational Linguistics

    work page 2005

  2. [2]

    Bernd Bohnet, Vinh Q. Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, Kai Hui, Tom Kwiatkowski, Ji Ma, Jianmo Ni, Lierni Sestorain Saralegui, Tal Schuster, William W. Cohen, Michael Collins, Dipanjan Das, Donald Metzler, Slav Petrov, and Kellie Webster. 2023. htt...

    arXiv 2023

  3. [3]

    Jan Buchmann, Xiao Liu, and Iryna Gurevych. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.463 Attribute or abstain: Large language models as long document assistants . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8113--8140, Miami, Florida, USA. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.emnlp-main.463 2024

  4. [4]

    Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. https://doi.org/10.1162/tacl_a_00377 Decontextualization: Making sentences stand-alone . Transactions of the Association for Computational Linguistics, 9:447--461

    doi:10.1162/tacl_a_00377 2021

  5. [5]

    Benjamin Cohen-Wang, Harshay Shah, Kristian Georgiev, and Aleksander Madry. 2024. https://openreview.net/forum?id=7CMNSqsZJt Contextcite: Attributing model generation to context . In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page 2024

  6. [6]

    Qiang Ding, Lvzhou Luo, Yixuan Cao, and Ping Luo. 2024. https://arxiv.org/abs/2412.11404 Attention with dependency parsing augmentation for fine-grained attribution . Preprint, arXiv:2412.11404

    arXiv 2024

  7. [7]

    Zi-Yi Dou and Graham Neubig. 2021. Word alignment by fine-tuning embeddings on parallel corpora. In Conference of the European Chapter of the Association for Computational Linguistics (EACL)

    work page 2021

  8. [8]

    Ori Ernst, Ori Shapira, Aviv Slobodkin, Sharon Adar, Mohit Bansal, Jacob Goldberger, Ran Levy, and Ido Dagan. 2024. https://doi.org/10.18653/v1/2024.findings-acl.389 The power of summary-source alignments . In Findings of the Association for Computational Linguistics: ACL 2024, pages 6527--6548, Bangkok, Thailand. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.findings-acl.389 2024

Show all 32 references
  1. [9]

    Alexander Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir Radev. 2019. https://doi.org/10.18653/v1/P19-1102 Multi-news: A large-scale multi-document summarization dataset and abstractive hierarchical model . In Proceedings of the 57th Annual Meeting of the Association for...

    2019 doi

  2. [10]

    Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023 a . https://doi.org/10.18653/v1/2023.acl-long.910 RARR : Researching and revising what language models say, using la...

    2023 doi

  3. [11]

    Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.398 Enabling large language models to generate text with citations . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6465--...

    2023 doi

  4. [12]

    Anisha Gunjal and Greg Durrett. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.215 Molecular facts: Desiderata for decontextualization in LLM fact verification . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3751--3768, Miami, Florida, ...

    2024 doi

  5. [13]

    Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. https://doi.org/10.18653/v1/2022.naacl-main.287 TRUE : Re-evaluating factual consistency evaluation . In Procee...

    2022 doi

  6. [14]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f78...

    2020

  7. [15]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  8. [16]

    Nelson Liu, Tianyi Zhang, and Percy Liang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.467 Evaluating verifiability in generative search engines . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7001--7025, Singapore. Association for C...

    2023 doi

  9. [17]

    Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, and Nat McAleese. 2022. https://api.semanticscholar.org/CorpusID:247594830 Teaching language models to support an...

    2022 arXiv

  10. [18]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.741 FA ct S core: Fine-grained atomic evaluation of factual precision in long form text generatio...

    2023 doi

  11. [19]

    Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Graham Neubig, Yulia Tsvetkov, and Hannaneh Hajishirzi. 2024. https://api.semanticscholar.org/CorpusID:266999558 Fine-grained hallucination detection and editing for language models . ArXiv, abs/2401.06855

    2024 arXiv

  12. [20]

    OpenAI. 2024. https://arxiv.org/abs/2410.21276 Gpt-4o system card . Preprint, arXiv:2410.21276

    2024 arXiv

  13. [21]

    Anirudh Phukan, Shwetha Somasundaram, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. 2024. https://doi.org/10.18653/v1/2024.findings-acl.682 Peering into the mind of language models: An approach for attribution in contextual question answering . In Findings of t...

    2024 doi

  14. [22]

    Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui. 2021. Mauve: Measuring the gap between neural text and human text using divergence frontiers. In NeurIPS

    2021

  15. [23]

    Jirui Qi, Gabriele Sarti, Raquel Fern \'a ndez, and Arianna Bisazza. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.347 Model internals-based answer attribution for trustworthy retrieval-augmented generation . In Proceedings of the 2024 Conference on Empirical Methods in Na...

    2024 doi

  16. [24]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. https://doi.org/10.1162/tacl_a_00605 In-context retrieval-augmented language models . Transactions of the Association for Computational Linguistics, 11:1316--1331

    2023 doi

  17. [25]

    Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2023. https://doi.org/10.1162/coli_a_00486 Measuring attribution in natural language generation models . Computational Ling...

    2023 doi

  18. [26]

    Tal Schuster, Adam Lelkes, Haitian Sun, Jai Gupta, Jonathan Berant, William Cohen, and Donald Metzler. 2024. https://doi.org/10.18653/v1/2024.naacl-long.74 SEMQA : Semi-extractive multi-source question answering . In Proceedings of the 2024 Conference of the North American Cha...

    2024 doi

  19. [27]

    Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. Bleurt: Learning robust metrics for text generation. In Proceedings of ACL

    2020

  20. [28]

    Aviv Slobodkin, Eran Hirsch, Arie Cattan, Tal Schuster, and Ido Dagan. 2024. https://doi.org/10.18653/v1/2024.acl-long.182 Attribute first, then generate: Locally-attributable grounded text generation . In Proceedings of the 62nd Annual Meeting of the Association for Computati...

    2024 doi

  21. [29]

    Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam M. Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, Yaguang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, Jame...

    2022 arXiv

  22. [30]

    Weinberger, and Yoav Artzi

    Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with bert . In International Conference on Learning Representations

    2020

  23. [31]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  24. [32]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.