Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

EvoPat: A Multi-LLM-based Patents Summarization and Analysis Agent

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read EvoPat, a multi-agent LLM system, claims to outperform GPT-4 on patent summarization, comparative analysis, and technical evaluation through five specialized roles plus retrieval and online search.

desk verdict EvoPat is a coherent multi-agent patent analysis system, but the paper's central claim that it beats GPT-4o rests on an evaluation that cannot separate output length and prompt-following from genuine analytical quality. read the letter →

arxiv 2412.18100 v1 pith:7MU2KN3L submitted 2024-12-24 cs.DL cs.AI

Suyuan Wang , Xueqian Yin , Menghao Wang , Ruofeng Guo , Kai Nan This is my paper · ORCID
classification cs.DLcs.AI
keywords multi-agentLLMpatentsummarizationretrieval-augmentedgenerationanalysisLLMLinguaAutoGenGPT-4evaluation
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

The paper presents EvoPat, a patent-analysis agent built from several large language models that split the job into five roles: identifying innovations, describing implementation methods, extracting technical details, comparing with similar patents, and surveying academic directions. The authors want to establish that this division of labor, with agents sharing intermediate results and pulling in external material through retrieval and online search, gives more detailed and useful patent reports than a single general LLM. Their evidence is that EvoPat beats GPT-4o on ROUGE and BERTScore overlap measures, and that human experts score it higher on informativeness, richness, coherence, attributability, and extensibility. If the result holds, researchers and engineers could use such an agent to triage large patent landscapes quickly.

What carries the argument

The load-bearing object is the multi-agent team: five LLM agents with distinct role profiles — innovation points, implementation method, technical details, horizontal comparison, and academic direction — communicating through AutoGen so that each later agent sees the work of earlier ones. Around this core sit retrieval-augmented generation with BGE-M3 embeddings stored in a Faiss vector database, tool calls to Google Patents and Semantic Scholar for external context, and two long-context strategies, Transform Messages and LLMLingua, with LLMLingua as the default. The machinery works by decomposing one patent into five complementary analyses and then reassembling them into a single Markdown-to-PDF report.

What would settle it

Re-run the comparison on the same 100-patent photoresist and nanoimprint sample with identical prompts, reference summaries written before any system output, and human raters blind to which system produced each report; if the ROUGE-1 gap of 0.2164 versus 0.0745 collapses or the expert scores converge, the claimed advantage of the multi-agent architecture is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a team of role-specialized LLM agents assembled by the AutoGen framework and backed by retrieval-augmented generation produces patent analyses that are both closer to gold summaries and more highly rated by domain experts than the single-model output of GPT-4o. The reported numbers are ROUGE-1 0.2164 against 0.0745, and expert ratings of 4.82 versus 4.13 for informativeness, 4.85 versus 3.95 for richness, and 4.34 versus 2.79 for extensibility on a five-point scale. The paper also claims that the LLMLingua prompt-compression strategy handles long patents better than a message-history-truncation strategy, and uses it as the system default.

Load-bearing premise

The central claim rests on the evaluation being fair: the reference summaries used for ROUGE and the human expert ratings must not have been shaped by knowledge of EvoPat's outputs, and the prompts given to GPT-4o must not have been weaker versions of EvoPat's detail-seeking instructions.

Editorial extensions

If this is right

  • Patent analysts could use EvoPat as a first-pass filter, reading the innovation and comparison sections to decide which patents deserve close manual review.
  • The same five-role decomposition can be transferred to other technical document types, such as standards, clinical trial records, or product documentation, wherever multiple expert perspectives add value.
  • Because EvoPat's largest measured gains are in informativeness and extensibility, the value of the system appears to come mostly from role specialization and external search rather than from better compression of the patent text.
  • Using LLMLingua as the default implies that heavily compressed prompts, although hard for humans to read, can preserve the information needed for detailed patent analysis while lowering token cost.

Reading between the lines

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

  • A decisive follow-up would hold prompts, gold reference summaries, and rater blinding fixed across systems, isolating whether the reported gap comes from the multi-agent architecture itself.
  • Removing single roles one at a time (for example, the academic direction scientist) and re-measuring expert ratings would reveal which specialist agent carries the most value.
  • Since BGE-M3 supports multilingual embeddings and the dataset already covers Chinese, English, Japanese, and Korean, the same scaffolding could be tested for cross-lingual patent retrieval and comparison.
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 proposes EvoPat, a multi-agent LLM system for patent summarization and analysis. The system preprocesses patent PDFs, embeds chunks in a Faiss vector database, and runs five specialized agents (innovation points, implementation method, technical details, horizontal comparison, and academic direction) coordinated through AutoGen, with GPT-4o as the underlying model and LLMLingua or Transform Messages for long-context handling. The evaluation claims that EvoPat outperforms GPT-4o on ROUGE and BERTScore (Table 1) and on human-rated informativeness, richness, coherence, attributability, and extensibility (Table 2), and that LLMLingua slightly outperforms the Transform Message strategy for long texts (Table 3). The appendix contains the prompts used for each of the five agent roles.

Significance. The application area is timely and the architecture is coherent: decomposing patent analysis into specialized roles with retrieval augmentation and external API calls is a reasonable engineering choice. However, the paper's headline contribution is empirical, and the current evidence does not establish the claimed superiority over GPT-4o. The manuscript does not release code, data, or evaluation scripts, and the evaluation protocol is underspecified in ways that directly affect the central claim. If the comparison were properly controlled, with disclosed baselines, gold references, and blinded human evaluation, the system could be a useful contribution; as it stands, the reported numbers do not justify the abstract's assertion.

major comments (4)
  1. [Section 3.1.3, Table 1] The reference summaries used for ROUGE and BERTScore are never defined. Equations (1)-(3) compute overlap against a 'reference summary,' but the manuscript does not state whether the references are original patent abstracts, expert-written gold summaries, or one of the model outputs. Without this information, the ROUGE advantage (0.2164 vs. 0.0745) cannot be interpreted; it could simply reflect which output happens to resemble the chosen reference. This is a load-bearing omission for the central outperformance claim.
  2. [Section 3.2, Table 1, and Appendix Tables 4-8] The GPT-4o baseline is not specified: no prompt, context window, temperature, maximum output length, or number of runs is reported. Meanwhile, the EvoPat prompts explicitly instruct agents to 'strive for the longest response possible,' 'output more text,' and add 'real quantitative figures' (Tables 4, 5, 6, 8). Since ROUGE-1 and ROUGE-L are recall-oriented, longer and more detailed outputs are favored by construction. The reported gap is therefore compatible with a verbosity or instruction-following confound rather than with superior analytic quality. A controlled comparison, for example matching output length or using a comparable detailed prompt for GPT-4o, is needed.
  3. [Section 3.2, Table 2] The human evaluation lacks a protocol. The manuscript reports only mean scores from four experts over 100 samples, with no rubric definitions, no rater assignment or blinding, and no inter-annotator agreement or variance. Two of the rated dimensions, Attributable and Extensible, explicitly reward the external-search and number-adding behaviors that EvoPat's prompts mandate, so the ratings may measure prompt compliance rather than output quality. A blinded, rubric-based evaluation with agreement statistics and significance testing is required to support the claim that EvoPat 'clearly outperforms GPT-4o across all dimensions.'
  4. [Section 3.3, Table 3] The long-context comparison is under-specified, and the table caption ('Evaluation comparison between EvoPat and GPT4') does not match the content, which compares two EvoPat strategies. It is unclear whether the same test set and human raters were used as in Tables 1 and 2, how many samples were evaluated, and what variability underlies the reported differences. Without this information, the conclusion that LLMLingua outperforms the Transform Message strategy is not supported.
minor comments (5)
  1. [Abstract, Section 1, Section 2.3.2] The abstract and introduction claim comparison with 'GPT-4,' while the experiments and system description use 'GPT-4o'; please align the terminology.
  2. [Throughout] There are numerous typos and copyediting issues, including 'catalogous,' 'repectively,' 'Horizaontal,' and 'EvoPatent' in Section 3.3; please proofread carefully.
  3. [Section 2.2.2] The text filtering step removes stop words from patent text before LLM analysis; this may degrade the model's comprehension, and the choice should be justified or tested.
  4. [Section 3.1.3] The paper states that BERTScore 'builds upon METEOR,' but BERTScore is based on BERT embeddings rather than METEOR; reference [30] is about METEOR and is not the appropriate basis for that claim. Please correct the description.
  5. [Figure 2] The figure presents a detailed example report without stating whether it is a verbatim system output or a curated illustration; please clarify its provenance.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: EvoPat's performance claims rest on external comparison with GPT-4o, not on self-referential definitions.

full rationale

EvoPat is an empirical systems paper, not a derivation chain. Its central claim, expressed in the abstract, is that EvoPat outperforms GPT-4 in patent summarization, comparative analysis, and technical evaluation. This claim is supported by comparisons against an external baseline, GPT-4o, in Tables 1 and 2. No parameter is fitted to the reported metrics, no uniqueness theorem from the authors is invoked, and the prior work cited for the multi-agent architecture is independent and external. The ROUGE and BERTScore definitions in Section 3.1.3 are standard and are not identified with any EvoPat input, so no equation reduces to its own assumptions. Some evaluation details are missing: gold reference summaries, the GPT-4o prompt, rater blinding, and the human-evaluation rubric are not disclosed, and the Appendix prompts instruct EvoPat agents to strive for the longest response possible and to add real quantitative figures, which could confound length-based metrics. Those are evaluation-validity concerns, not circularity: the paper's own definitions do not force the reported outcomes, and the comparison target is independent of the authors' system. The Conclusion also lists genuine limitations, indicating contingent empirical claims rather than definitional identities. The proportionate circularity score is therefore 1.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its contributions are system-design choices and evaluation claims. The key unverified inputs are the unspecified configuration of the pipeline (compression, retrieval, prompt baselines) and the validity of the evaluation protocol, both of which directly affect the headline comparison.

free parameters (3)
  • LLMLingua compression ratio = not reported
    Section 3.3 compares LLMLingua to Transform Messages; the headline ROUGE-1 (0.2164) depends on the compression settings, which are not given.
  • Number of similar patents retrieved per search = not reported
    The Horizontal Comparison Scientist (Section 2.3.2) uses Google Patents API results, and the comparison content in Table 2's human evaluation depends on how many and which patents are returned; no retrieval limit is stated.
  • Expert rating aggregation and scale = not reported
    Table 2 reports averages out of 5 for four experts, but the individual scores, aggregation method, and scale anchors are not described, making the reported 4.82 vs 4.13 difference unverifiable.
assumptions (4)
  • domain assumption ROUGE and BERTScore computed against gold summaries are valid indicators of patent analysis quality.
    Section 3.1.3 adopts these summarization metrics without explaining how gold references were created or why n-gram overlap is appropriate for an analysis that includes external web search.
  • domain assumption Human expert ratings are unbiased and consistent.
    Section 3.2 reports average scores from four experts but gives no rubric details, no inter-annotator agreement, and no statement about blinding, so the scores must be taken on faith.
  • domain assumption LLMLingua compression preserves the information needed for patent analysis.
    Section 2.3.1 sets LLMLingua as the default long-text strategy, and Section 3.3 concludes it outperforms Transform Messages, but no direct measure of information preservation is provided.
  • domain assumption BGE-M3 embedding and Faiss retrieval improve the agents' outputs.
    Section 2.2.3 introduces the vector database without an ablation or retrieval-quality evaluation, so its contribution to the reported scores is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EvoPat: A Multi-LLM-based Patents Summarization and Analysis Agent." pith.science (2026). https://pith.science/paper/7MU2KN3L

@misc{pith2026241218100,
  author       = {Pith},
  title        = {Pith review of: EvoPat: A Multi-LLM-based Patents Summarization and Analysis Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MU2KN3L}},
  note         = {Machine review of arXiv:2412.18100}
}
read the original abstract

The rapid growth of scientific techniques and knowledge is reflected in the exponential increase in new patents filed annually. While these patents drive innovation, they also present significant burden for researchers and engineers, especially newcomers. To avoid the tedious work of navigating a vast and complex landscape to identify trends and breakthroughs, researchers urgently need efficient tools to summarize, evaluate, and contextualize patents, revealing their innovative contributions and underlying scientific principles.To address this need, we present EvoPat, a multi-LLM-based patent agent designed to assist users in analyzing patents through Retrieval-Augmented Generation (RAG) and advanced search strategies. EvoPat leverages multiple Large Language Models (LLMs), each performing specialized roles such as planning, identifying innovations, and conducting comparative evaluations. The system integrates data from local databases, including patents, literature, product catalogous, and company repositories, and online searches to provide up-to-date insights. The ability to collect information not included in original database automatically is also implemented. Through extensive testing in the natural language processing (NLP) domain, we demonstrate that EvoPat outperforms GPT-4 in tasks such as patent summarization, comparative analysis, and technical evaluation. EvoPat represents a significant step toward creating AI-powered tools that empower researchers and engineers to efficiently navigate the complexities of the patent landscape.

Figures

Figures reproduced from arXiv: 2412.18100 by the authors.

Figure 1
Figure 1. System overview of EvoPat. For a given patent, the process begins with preprocessing to extract and filter useful information, which is then embedded and stored in a database for easy retrieval in the future. Next, a multi-agent system is employed to analyze the patent from five distinct perspectives. Finally, the results are integrated and outputted as a PDF document for further examination. 2 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 2
Figure 2. An example of multi-agent system in EvoPat 2.3.3 Tool invoking Due to LLMs’ limited knowledge base, accessing data outside their training corpus often leads to hallucination, which is unacceptable for patent analysis tasks. Beyond local Retrieval-Augmented Generation(RAG) [1] methods, one of the most common approaches is to retrieve relevant knowledge by calling external APIs, effectively expanding the model’s knowl… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Unstructured Communication to Intelligent RAG: Multi-Agent Automation for Supply Chain Knowledge Bases

    cs.AI 2025-06 conditional novelty 5.0 of 10

    Converting raw support tickets into a 3.4%-volume, category-structured knowledge base with three LLM agents improves RAG helpful answers from 38.60% to 48.74% on a real supply chain ticket dataset.

  2. A Comprehensive Survey of Deep Research: Systems, Methodologies, and Applications

    cs.AI 2025-06 conditional novelty 4.0 of 10

    A survey of 80+ Deep Research systems that proposes a four-layer taxonomy (foundation models, tool use, planning, synthesis) and compares commercial and open-source implementations.

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages · cited by 2 Pith papers

  1. [1]

    Retrieval-augmented generation for knowledge- intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks. Advances in Neural Information Processing Systems , 33:9459–9474, 2020

    work page 2020

  2. [2]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  3. [3]

    Researchagent: Iterative research idea generation over scientific literature with large language models

    Jinheon Baek, Sujay Kumar Jauhar, Silviu Cucerzan, and Sung Ju Hwang. Researchagent: Iterative research idea generation over scientific literature with large language models. arXiv preprint arXiv:2404.07738, 2024

    arXiv 2024

  4. [4]

    From sparse to dense: Gpt-4 summarization with chain of density prompting

    Griffin Adams, Alexander R Fabbri, Faisal Ladhak, Eric Lehman, and Noémie Elhadad. From sparse to dense: Gpt-4 summarization with chain of density prompting. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing , page 68, 2023

    work page 2023

  5. [5]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:27730–27744, 2022

    2022

  6. [6]

    Patent classification by fine-tuning bert language model

    Jieh-Sheng Lee and Jieh Hsiang. Patent classification by fine-tuning bert language model. World Patent Information, 61:101965, 2020

    work page 2020

  7. [7]

    Automatic summarization of scientific articles: A survey

    Nouf Ibrahim Altmami and Mohamed El Bachir Menai. Automatic summarization of scientific articles: A survey. Journal of King Saud University-Computer and Information Sciences , 34(4):1011–1028, 2022

    work page 2022

  8. [8]

    Bridging Research and Readers: A Multi-Modal Automated Academic Papers Interpretation System

    Feng Jiang, Kuang Wang, and Haizhou Li. Bridging research and readers: A multi-modal automated academic papers interpretation system. arXiv preprint arXiv:2401.09150, 2024

    work page Pith review arXiv 2024

Show all 32 references
  1. [9]

    Scipip: An llm-based scientific paper idea proposer

    Wenxiao Wang, Lihui Gu, Liye Zhang, Yunxiang Luo, Yi Dai, Chen Shen, Liang Xie, Binbin Lin, Xiaofei He, and Jieping Ye. Scipip: An llm-based scientific paper idea proposer. arXiv preprint arXiv:2410.23166, 2024

    2024 arXiv

  2. [10]

    Artificial intelligence exploring the patent field.arXiv preprint arXiv:2403.04105, 2024

    Lekang Jiang and Stephan Goetz. Artificial intelligence exploring the patent field.arXiv preprint arXiv:2403.04105, 2024

    2024 arXiv

  3. [11]

    Patentgpt: A large language model for intellectual property.arXiv preprint arXiv:2404.18255, 2024

    Zilong Bai, Ruiji Zhang, Linqing Chen, Qijun Cai, Yuan Zhong, Cong Wang, Yan Fang, Jie Fang, Jing Sun, Weikuan Wang, et al. Patentgpt: A large language model for intellectual property.arXiv preprint arXiv:2404.18255, 2024. 9 A PREPRINT - D ECEMBER 25, 2024

    2024 arXiv

  4. [12]

    Llm-based extraction of contradictions from patents

    Stefan Trapp and Joachim Warschat. Llm-based extraction of contradictions from patents. In International TRIZ Future Conference, pages 3–19. Springer, 2024

    2024

  5. [13]

    Markdown: Syntax documentation

    John Gruber. Markdown: Syntax documentation. https://daringfireball.net/projects/markdown/

  6. [14]

    pdfplumber: Extract text, tables, and metadata from pdfs

    Jeremy Singer-Vine. pdfplumber: Extract text, tables, and metadata from pdfs. https://github.com/jsvine/ pdfplumber/

  7. [15]

    https://python.langchain.com/docs/integrations/ document_loaders/pypdfloader/

    Pypdfloader: A pdf loading utility for langchain. https://python.langchain.com/docs/integrations/ document_loaders/pypdfloader/

  8. [16]

    Bge m3-embedding: Multi- lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multi- lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216, 2024

    2024 arXiv

  9. [17]

    Billion-scale similarity search with gpus

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535–547, 2019

    2019

  10. [18]

    https://www.meta.com/

    Meta. https://www.meta.com/

  11. [19]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023

    2023 arXiv

  12. [20]

    Llmlingua: Compressing prompts for accelerated inference of large language models

    Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Llmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736, 2023

    2023 arXiv

  13. [21]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  14. [22]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  15. [23]

    https://openai.com/

    Openai. https://openai.com/

  16. [24]

    https: //patents.google.com/patent/US20170263445A1/en?oq=US20170263445A1

    Us20170263445a1: Manufacturing method of semiconductor device and template for nanoimprint. https: //patents.google.com/patent/US20170263445A1/en?oq=US20170263445A1

  17. [25]

    https://patents.google.com/

    Google patents. https://patents.google.com/

  18. [26]

    https://www.semanticscholar.org/

    Semantic scholar. https://www.semanticscholar.org/

  19. [27]

    Sciagents: Automating scientific discovery through multi-agent intelligent graph reasoning

    Alireza Ghafarollahi and Markus J Buehler. Sciagents: Automating scientific discovery through multi-agent intelligent graph reasoning. arXiv preprint arXiv:2409.05556, 2024

    2024 arXiv

  20. [28]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. Text summarization branches out, 2004

    2004

  21. [29]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with BERT. arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  22. [30]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Banerjee Satanjeev and Lavie Alon. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005

    2005

  23. [31]

    mface: Multilingual summarization with factual consistency evaluation

    Roee Aharoni, Shashi Narayan, Joshua Maynez, Jonathan Herzig, Elizabeth Clark, and Mirella Lapata. mface: Multilingual summarization with factual consistency evaluation. arXiv preprint arXiv:2212.10622, 2022

    2022 arXiv

  24. [32]

    source pdf

    Alexander R Fabbri, Wojciech Kry ´sci´nski, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. Summeval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391–409, 2021. 10 A PREPRINT - D ECEMBER 25, 2024 5 A...

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.