Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Scaling Multi-Document Event Summarization: Evaluating Compression vs. Full-Text Approaches

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read In large-scale multi-document summarization, feeding the full text or a retrieval-selected subset to a long-context model outperforms hierarchical and incremental compression pipelines in most settings.

desk verdict A careful empirical comparison that delivers one robust finding (iterative compression loses information at scale) and a weaker one (full-context vs retrieval), with a genuinely useful intermediate-recall diagnostic. read the letter →

arxiv 2502.06617 v1 pith:KCSLHKSP submitted 2025-02-10 cs.CL

Adithya Pratapa , Teruko Mitamura This is my paper · ORCID
classification cs.CL
keywords multi-documentsummarizationlong-contextlanguagemodelscompression-basedretrieval-augmentedgenerationhierarchicalincrementalcontentselectionevaluationA3CU
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

Summarizing an event from a pile of about one hundred news articles is a different task from summarizing a handful of documents, and this paper asks which system design handles it best. It compares full-text methods, which hand the entire input to a long-context model, with three compression approaches: retrieval, which selects the most relevant documents; hierarchical, which summarizes documents and then merges those summaries; and incremental, which maintains a running summary as it reads. Across three datasets, full-text and retrieval score highest on the A3CU content-selection metric in most settings, while the two iterative compression methods lag. Tracking the compression methods' intermediate outputs shows they actually find salient information early on, often beating full-context recall, but then lose it as the pipeline progresses. The paper concludes that large-scale summarization should move toward hybrid systems that compress the input selectively and then let a long-context model do the final synthesis.

What carries the argument

The load-bearing mechanism is the A3CU (Atomic Content Unit) metric, a learned reference-based score for how many atomic content units of a reference summary appear in a generated summary. Its F1 score yields the main method rankings, and its recall score, applied to the intermediate outputs of the compression pipelines, is what reveals the paper's central diagnostic: compression methods capture salient information early but bleed it away through later stages. The comparison framework, full-text, retrieval, hierarchical, and incremental, all truncated to the same 128k-token budget, ensures that the methods differ only in how much input the model sees at each step.

What would settle it

Run a best-worst human evaluation on a random sample of a few hundred examples from Background and WCEP, comparing full-context, retrieval, hierarchical, and incremental summaries against the reference summaries; if human raters pick hierarchical or incremental as best as often as they pick full-text or retrieval, the A3CU-based ranking would be overturned.

Watch

Extended reading notes

Core claim

The central claim is an empirical ranking: at the scale of roughly one hundred documents per summary, full-context and retrieval-based systems are the best choices in most settings, while hierarchical and incremental compression methods underperform. The paper establishes this with A3CU F1 scores on SummHay, Background, and WCEP across four long-context model configurations, and then shows through A3CU recall that the compression methods' best intermediate outputs retain more salient information than their final summaries, often more than full-context itself. The authors attribute this to the multi-stage pipeline and the lack of global context in iterative methods. They conclude that full-context systems are suboptimal on large-scale MDS and argue that hybrid approaches, which combine input compression with long-context reasoning, are the most promising direction.

Load-bearing premise

The rankings rest on the A3CU automatic content-coverage score being a faithful proxy for human judgment of content selection at the hundred-document scale; the only human check is a 62-example best-worst evaluation from one dataset, so a metric bias could move the main results.

Editorial extensions

If this is right

  • Large-scale MDS systems with about one hundred documents per summary should favor full-context or retrieval-augmented pipelines over hierarchical and incremental compression.
  • Iterative compression methods validated on book-length or small-scale summarization do not transfer to this scale, because they drop entities, numbers, and event specifics in higher abstraction levels.
  • The measured gap between intermediate and final recall gives a concrete target: a method that preserves or re-selects its best intermediate output could approach or beat full-context performance.
  • Because reference summaries for hundred-document inputs are expensive to obtain, further scaling of MDS will require reference-free content-selection metrics.
  • Long-context models, even with 128k-token windows, are not yet reliably better than retrieval-augmented pipelines on this task, so combining input compression with long-context reasoning is the recommended direction.

Reading between the lines

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

  • The intermediate-retention result implies a testable ceiling: re-ranking the intermediate summaries of a compression pipeline by A3CU recall and using the best one as the output would beat full-context on many examples, so an aggregator that selects among intermediate summaries is a natural next step the paper does not explore.
  • Incremental methods process documents in dataset order, and the paper's examples show later non-salient documents derailing the running summary; reordering or relevance-filtering the document stream before incremental processing could change the ranking, an experiment the paper does not run.
  • A3CU recall is sensitive to summary length, and intermediate summaries are typically longer than final ones; a length-controlled replication would clarify how much of the retention gap is genuine content selection rather than verbosity.
  • The large per-example variance and the surprising small-model-over-large-model result on SummHay suggest that long-context retrieval ability, not raw capacity, is the binding constraint, so model choice for MDS should be guided by retrieval-heavy long-context benchmarks.
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

3 major / 4 minor

Summary. The paper compares full-context and compression-based (hierarchical, incremental, retrieval) approaches to large-scale multi-document summarization on three datasets (SummHay, Background, WCEP), each with ~100 documents per summary, using four long-context transformers. The central empirical claim is that full-text and retrieval methods perform best in most settings, while iterative compression methods (hierarchical and incremental) underperform despite retaining salient information in intermediate stages. The paper also analyzes intermediate-output recall to argue that compression methods lose information through their multistage pipelines, and it advocates hybrid compression-plus-long-context systems. Evaluation is primarily based on the A3CU F1 metric, supplemented by a 62-example best-worst human evaluation on SummHay.

Significance. If the empirical ranking holds, the paper provides a valuable practical result for large-scale MDS: it suggests that full-context and retrieval-first pipelines are more reliable than iterative compression, while also identifying intermediate-stage recall as a promising direction for hybrid methods. The study is well-scoped, covers diverse models (Llama-3.1-8B/70B, Command-R, Jamba-1.5-Mini), includes three datasets with different characteristics, and provides code/data release. The intermediate-retention analysis (Figure 1) is a useful diagnostic that goes beyond end-to-end scores. The main weakness is that the central ranking rests almost entirely on the A3CU automatic metric, with only a thin human anchor on a single dataset, and the results are single-run with no variance estimation.

major comments (3)
  1. [§3.1, §3.5, Tables 7–8] The central ranking is supported only by A3CU F1, a learned metric that has not been validated on inputs of ~100 documents per summary. The only human check is a 62-example best-worst evaluation on SummHay performed by the first author (§3.5, Table 3), and it covers a single model (Llama-3.1-8B). On Background and WCEP, the differences between full-context and retrieval are within about 1–2 F1 points (e.g., Background, Table 7: Llama-3.1-8B full-context 15.6 vs. retrieval 16.9; WCEP, Table 8: Llama-3.1-8B full-context 30.7 vs. retrieval 29.7). A systematic metric bias—for example, sensitivity to summary length or phrasing—could therefore flip the full-context vs. retrieval ordering on these datasets and weaken the headline claim. The authors should either validate A3CU against human judgments on at least one non-SummHay dataset, or temper the abstract's claim to reflect that the full-text-vs-retrieval comparison is not yet firmly established outside the query-focused setting.
  2. [§3.2, Tables 5–8] All reported results come from a single inference run with no error bars, confidence intervals, or significance tests. Because summaries were sampled with temperature 0.5 (Appendix A.2), the paper could easily report variance across a few seeds. This matters because several key comparisons rest on small margins: on WCEP, full-context and retrieval differ by at most about 1 F1 point across all models (Table 8), and on Background the retrieval advantage over full-context is 1.3 points for Llama-3.1-8B and negative for the other models (Table 7). The qualitative ranking of iterative methods as clearly worse is likely robust given the large gaps, but the finer full-context-vs-retrieval comparison and some cross-model differences need uncertainty quantification.
  3. [§3.4, Table 9] The paper argues that iterative methods lose salient information because intermediate recall is high but final recall is low. However, Figure 1 uses A3CU recall, which is naturally capped by summary length, and Table 9 shows substantial length differences across methods—for example, on SummHay the hierarchical final summaries are around 106 words for Llama-3.1-8B compared with 162 for full-context and 185 for the reference. The claim in §3.4 that 'We do not find any noticeable correlation' between recall and summary length is not supported by any reported statistic. The authors should provide the actual correlation coefficient or perform a length-controlled analysis (e.g., truncating or padding summaries) to rule out the alternative explanation that iterative methods appear worse mainly because they generate shorter summaries.
minor comments (4)
  1. [§3.3] The statement that Llama-3.1-8B outperforming Llama-3.1-70B on SummHay 'aligns with their relative performance on the RULER benchmark' is vague; please cite the specific RULER results or present the relevant comparison, since the RULER ranking depends on the task and context length.
  2. [Table 2] Table 2 reports relative percentages against the full-context baseline, but the baseline absolute scores are only given in the first column. For readability, consider showing absolute scores for all methods or adding a note that the baseline values are repeated from the appendix tables.
  3. [§A.5, Table 10] The Gemini-1.5 results are described as preliminary, but the table shows that Gemini-1.5-Flash hierarchical scores are strikingly low (A3CU F1 10.2 vs. 24.6 for full-context). A brief comment on whether this reflects a prompt-format issue (as hinted later) would help readers interpret the appendix.
  4. [Limitations] The Limitations section acknowledges the lack of large-scale human evaluation, but it does not mention that the only human evaluation was conducted by the first author. Given the potential for bias, this should be stated explicitly in the limitations or in §3.5.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the main ranking is an empirical evaluation against external benchmarks, with only a minor non-load-bearing self-citation.

full rationale

The paper's central claims—full-context/retrieval outperform iterative compression, and intermediate outputs retain salient content—are empirical findings measured with A3CU F1/recall (Liu et al., 2023b), an externally developed metric, on three benchmarks. Two of the three benchmarks (SummHay, WCEP) come from independent groups, and the third (Background, Pratapa et al., 2023) is a prior published dataset with human-written reference summaries; it is used as a test collection, not fitted or constructed for the present conclusions. No parameter in the paper is fitted to the headline outcome, and no quantity is defined in terms of the claim it is said to predict. The intermediate-retention analysis tracks actual intermediate outputs; its best-of-many versus single-output comparison is a methodological limitation, not a definitional identity. The Limitations section explicitly concedes that human evaluation is infeasible at scale, which is a validity caveat about metric reliance, not evidence of circularity. The only author-overlapping artifact is the Background dataset citation, which is not load-bearing because the main findings also hold on the independent SummHay and WCEP datasets. Accordingly, the derivation chain is self-contained against external evidence and shows no circular step.

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

This is an empirical study, so the ledger lists the hand-chosen experimental settings and domain assumptions the conclusions depend on, rather than fitted parameters in a derivation. No new theoretical entities are introduced. The main risks are the validity of A3CU as the content-selection measure, the fairness of the 128k token cap, the completeness of reference summaries, and the independence of the human check.

free parameters (4)
  • Maximum input length cap = 128k tokens
    Chosen as a uniform cap for all methods (§2.2). If the cap were larger or smaller, the relative performance of full-context vs compression could shift, since full-context is truncated at this point.
  • Retrieval context limit = 32k tokens
    Selected based on the claim that all tested transformers are effective at this context length (§2.2). It determines how many documents the retrieval method sees.
  • Iterative chunk size = 4096 tokens
    Chosen for both iterative methods (§A.2); ablations show minimal improvement with larger chunks, so it is not highly load-bearing.
  • Maximum summary length percentile = 80th percentile of validation summaries
    Used to set the generation length cap per dataset (§A.2). Summary length affects A3CU recall and precision.
assumptions (4)
  • domain assumption The A3CU metric accurately measures content selection quality in multi-document summarization.
    The paper's core rankings rely on A3CU F1 (§3.1). If A3CU does not align with human judgments at this scale, the comparative conclusions may be invalid. The paper itself notes human evaluation is the gold standard and is only partially available.
  • domain assumption The 128k token input cap, with the described truncation of the longest documents and per-timestamp budgeting for Background, provides a fair comparison across methods.
    This cap is imposed on all methods (§2.2). If truncation removes documents that iterative or retrieval methods would have used, or penalizes full-context differentially, the ranking could change.
  • domain assumption The reference summaries for Background and WCEP are treated as complete ground truth for content selection.
    A3CU is reference-based (§2.1, §3.1); incomplete or biased references would bias the metric.
  • domain assumption The first author's human annotations are treated as unbiased enough to corroborate the automatic metric.
    Section 3.5 states the evaluation was done by the first author, which risks bias and limits the independence of the human check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scaling Multi-Document Event Summarization: Evaluating Compression vs. Full-Text Approaches." pith.science (2026). https://pith.science/paper/KCSLHKSP

@misc{pith2026250206617,
  author       = {Pith},
  title        = {Pith review of: Scaling Multi-Document Event Summarization: Evaluating Compression vs. Full-Text Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCSLHKSP}},
  note         = {Machine review of arXiv:2502.06617}
}
read the original abstract

Automatically summarizing large text collections is a valuable tool for document research, with applications in journalism, academic research, legal work, and many other fields. In this work, we contrast two classes of systems for large-scale multi-document summarization (MDS): compression and full-text. Compression-based methods use a multi-stage pipeline and often lead to lossy summaries. Full-text methods promise a lossless summary by relying on recent advances in long-context reasoning. To understand their utility on large-scale MDS, we evaluated them on three datasets, each containing approximately one hundred documents per summary. Our experiments cover a diverse set of long-context transformers (Llama-3.1, Command-R, Jamba-1.5-Mini) and compression methods (retrieval-augmented, hierarchical, incremental). Overall, we find that full-text and retrieval methods perform the best in most settings. With further analysis into the salient information retention patterns, we show that compression-based methods show strong promise at intermediate stages, even outperforming full-context. However, they suffer information loss due to their multi-stage pipeline and lack of global context. Our results highlight the need to develop hybrid approaches that combine compression and full-text approaches for optimal performance on large-scale multi-document summarization.

Figures

Figures reproduced from arXiv: 2502.06617 by the authors.

Figure 1
Figure 1. Salient information retention in the intermediate and final summaries (A3CU [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Salient information retention in the intermedi [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. A3CU F1 score distribution across examples. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Query-Focused Event Summarization: A Dataset and Benchmark

    cs.CL 2026-07 conditional novelty 6.0 of 10

    QFESum provides a large event-oriented QFS benchmark; RAT adaptive retrieval plus SHC hierarchical event clustering beat baselines on lexical, semantic, LLM-event-match and human metrics.

Reference graph

Works this paper leans on

41 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Carbonell, George R

    James Allan, Jaime G. Carbonell, George R. Doddington, Jonathan Yamron, and Yiming Yang. 1998. https://api.semanticscholar.org/CorpusID:9063912 Topic detection and tracking pilot study final report

    work page 1998

  2. [2]

    Javed Aslam, Fernando Diaz, Matthew Ekstrand-Abueg, Richard McCreadie, Virgil Pavlu, and Tetsuya Sakai. 2015. https://trec.nist.gov/pubs/trec24/papers/Overview-TS.pdf TREC 2015 Temporal Summarization Track Overview . In TREC

    work page 2015

  3. [3]

    Giang Binh Tran, Mohammad Alrifai, and Dat Quoc Nguyen. 2013. https://doi.org/10.1145/2487788.2487829 Predicting relevant news events for timeline summaries . In Proceedings of the 22nd International Conference on World Wide Web, WWW '13 Companion, page 91–92, New York, NY, USA. Association for Computing Machinery

    arXiv 2013

  4. [4]

    Yapei Chang, Kyle Lo, Tanya Goyal, and Mohit Iyyer. 2024. https://openreview.net/forum?id=7Ttk3RzDeu Booookscore: A systematic exploration of book-length summarization in the era of LLM s . In The Twelfth International Conference on Learning Representations

    2024

  5. [5]

    Hai Leong Chieu and Yoong Keok Lee. 2004. https://doi.org/10.1145/1008992.1009065 Query Based Event Extraction along a Timeline . In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '04, page 425–432, New York, NY, USA. Association for Computing Machinery

    work page arXiv 2004

  6. [6]

    Cohere For AI . 2024. https://huggingface.co/CohereForAI/c4ai-command-r-08-2024 c4ai-command-r-08-2024

    work page 2024

  7. [7]

    Hoa Dang and Karolina Owczarzak. 2009. https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=903465 Overview of the TAC 2008 Update Summarization Task

    work page 2009

  8. [8]

    Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. 2024. https://openreview.net/forum?id=ONOtpXLqqw Longro PE : Extending LLM context window beyond 2 million tokens . In Forty-first International Conference on Machine Learning

    work page 2024

Show all 41 references
  1. [9]

    Demian Gholipour Ghalandari, Chris Hokamp, Nghia The Pham, John Glover, and Georgiana Ifrim. 2020. https://doi.org/10.18653/v1/2020.acl-main.120 A large-scale multi-document summarization dataset from the W ikipedia current events portal . In Proceedings of the 58th Annual Mee...

    2020 doi

  2. [10]

    Tanya Goyal, Junyi Jessy Li, and Greg Durrett. 2022. https://arxiv.org/abs/2209.12356 News summarization and evaluation in the era of gpt-3 . Preprint, arXiv:2209.12356

    2022 arXiv

  3. [11]

    Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, and Boris Ginsburg. 2024. https://openreview.net/forum?id=kIoBbc76Sy RULER : What s the real context size of your long-context language models? In First Conference on Language Modeling

    2024

  4. [12]

    Kung-Hsiang Huang, Philippe Laban, Alexander Fabbri, Prafulla Kumar Choubey, Shafiq Joty, Caiming Xiong, and Chien-Sheng Wu. 2024. https://doi.org/10.18653/v1/2024.naacl-long.32 Embrace divergence for richer insights: A multi-document summarization benchmark and a case study o...

    2024 doi

  5. [13]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. https://doi.org/10.18653/v1/2024.acl-long.91 L ong LLML ingua: Accelerating and enhancing LLM s in long context scenarios via prompt compression . In Proceedings of the 62nd An...

    2024 doi

  6. [14]

    Greg Kamradt. 2023. https://github.com/gkamradt/LLMTest_NeedleInAHaystack Needle in a haystack - pressure testing llms

    2023

  7. [15]

    Marzena Karpinska, Katherine Thai, Kyle Lo, Tanya Goyal, and Mohit Iyyer. 2024. https://arxiv.org/abs/2406.16264 One thousand and one pairs: A "novel" challenge for long-context language models . Preprint, arXiv:2406.16264

    2024 arXiv

  8. [16]

    Yekyung Kim, Yapei Chang, Marzena Karpinska, Aparna Garimella, Varun Manjunatha, Kyle Lo, Tanya Goyal, and Mohit Iyyer. 2024. https://openreview.net/forum?id=YfHxQSoaWU FABLES : Evaluating faithfulness and content selection in book-length summarization . In First Conference on...

    2024

  9. [17]

    Svetlana Kiritchenko and Saif Mohammad. 2017. https://doi.org/10.18653/v1/P17-2074 Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (...

    2017 doi

  10. [18]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. https://dl.acm.org/doi/10.1145/3600006.3613165 Efficient memory management for large language model serving with pagedattention . In Proceedings...

    2023

  11. [19]

    Fabbri, Caiming Xiong, and Chien-Sheng Wu

    Philippe Laban, Alexander R. Fabbri, Caiming Xiong, and Chien-Sheng Wu. 2024. https://arxiv.org/abs/2407.01370 Summary of a haystack: A challenge to long-context llms and rag systems . Preprint, arXiv:2407.01370

    2024 arXiv

  12. [20]

    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

  13. [21]

    Yixin Liu, Alex Fabbri, Pengfei Liu, Yilun Zhao, Linyong Nan, Ruilin Han, Simeng Han, Shafiq Joty, Chien-Sheng Wu, Caiming Xiong, and Dragomir Radev. 2023 a . https://doi.org/10.18653/v1/2023.acl-long.228 Revisiting the gold standard: Grounding summarization evaluation with ro...

    2023 doi

  14. [22]

    Yixin Liu, Alexander Fabbri, Yilun Zhao, Pengfei Liu, Shafiq Joty, Chien-Sheng Wu, Caiming Xiong, and Dragomir Radev. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.1018 Towards interpretable and efficient automatic reference-based summarization evaluation . In Proceedin...

    2023 doi

  15. [23]

    Chiu, Siyu Ren, Fei Yuan, Wenting Zhao, Zhiyong Wu, and Alexander M

    Yi Lu, Jing Nathan Yan, Songlin Yang, Justin T. Chiu, Siyu Ren, Fei Yuan, Wenting Zhao, Zhiyong Wu, and Alexander M. Rush. 2024. https://arxiv.org/abs/2409.12181 A controlled study on long context extension and generalization in llms . Preprint, arXiv:2409.12181

    2024 arXiv

  16. [24]

    Rui Meng*, Ye Liu*, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2024. https://huggingface.co/Salesforce/SFR-Embedding-2_R Sfr-embedding-2: Advanced text embedding with multi-stage training

    2024

  17. [25]

    Meta. 2024. https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md Llama 3.1 model card

    2024

  18. [26]

    Ani Nenkova and Rebecca Passonneau. 2004. https://aclanthology.org/N04-1019/ Evaluating content selection in summarization: The pyramid method . In Proceedings of the Human Language Technology Conference of the North A merican Chapter of the Association for Computational Lingu...

    2004

  19. [27]

    Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. 2023. https://arxiv.org/abs/2309.00071 Yarn: Efficient context window extension of large language models . Preprint, arXiv:2309.00071

    2023 arXiv

  20. [28]

    Adithya Pratapa, Kevin Small, and Markus Dreyer. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.505 Background summarization of event timelines . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8111--8136, Singapore. Associat...

    2023 doi

  21. [29]

    Mathieu Ravaut, Aixin Sun, Nancy Chen, and Shafiq Joty. 2024. https://doi.org/10.18653/v1/2024.acl-long.153 On context utilization in summarization with large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1...

    2024 doi

  22. [30]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. 2024. https://openreview.net/forum?id=GN921JHCRw RAPTOR : Recursive abstractive processing for tree-organized retrieval . In The Twelfth International Conference on Learning Repres...

    2024

  23. [31]

    Ori Shapira, David Gabay, Yang Gao, Hadar Ronen, Ramakanth Pasunuru, Mohit Bansal, Yael Amsterdamer, and Ido Dagan. 2019. https://doi.org/10.18653/v1/N19-1072 Crowdsourcing lightweight pyramids for manual summary evaluation . In Proceedings of the 2019 Conference of the North ...

    2019 doi

  24. [32]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2021. https://arxiv.org/abs/2104.09864 Roformer: Enhanced transformer with rotary position embedding . Preprint, arXiv:2104.09864

    2021 arXiv

  25. [33]

    Jamba Team, Barak Lenz, Alan Arazi, Amir Bergman, Avshalom Manevich, Barak Peleg, Ben Aviram, Chen Almagor, Clara Fridman, Dan Padnos, Daniel Gissin, Daniel Jannai, Dor Muhlgay, Dor Zimberg, Edden M Gerber, Elad Dolev, Eran Krakovsky, Erez Safahi, Erez Schwartz, Gal Cohen, Gal...

    2024 arXiv

  26. [34]

    Giang Tran, Mohammad Alrifai, and Eelco Herder. 2015. https://link.springer.com/chapter/10.1007/978-3-319-16354-3_26 Timeline summarization from relevant headlines . In Advances in Information Retrieval, pages 245--256, Cham. Springer International Publishing

    2015 doi

  27. [35]

    Jesse Vig, Alexander Fabbri, Wojciech Kryscinski, Chien-Sheng Wu, and Wenhao Liu. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.109 Exploring neural models for query-focused summarization . In Findings of the Association for Computational Linguistics: NAACL 2022, pages...

    2022 doi

  28. [36]

    Lu Wang, Claire Cardie, and Galen Marchetti. 2015. https://doi.org/10.3115/v1/N15-1112 Socially-informed timeline generation for complex events . In Proceedings of the 2015 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Langua...

    2015 doi

  29. [37]

    Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2024. https://openreview.net/forum?id=mlJLVigNHp RECOMP : Improving retrieval-augmented LM s with context compression and selective augmentation . In The Twelfth International Conference on Learning Representations

    2024

  30. [38]

    Tan Yu, Anbang Xu, and Rama Akkiraju. 2024. https://arxiv.org/abs/2409.01666 In defense of rag in the era of long-context language models . Preprint, arXiv:2409.01666

    2024 arXiv

  31. [39]

    Dawei Zhu, Liang Wang, Nan Yang, Yifan Song, Wenhao Wu, Furu Wei, and Sujian Li. 2024. https://arxiv.org/abs/2404.12096 Longembed: Extending embedding models for long context retrieval . Preprint, arXiv:2404.12096

    2024 arXiv

  32. [40]

    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...

  33. [41]

    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 8, 2026 · model on record in the stance chip above.