Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Trained adapters let an 8B model beat a 70B model on low-resource summarization.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

PEFT adapters trained on high-resource summarization domains can improve Llama-3-8B's summaries on unseen domains, but the reported gains are weakened by test-set selection and missing significance tests.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful PEFT summarization benchmark, but the headline claim is undercut by holdout-based adapter selection. the 3 major comments →

arxiv 2509.01314 v1 pith:GDIEZUCQ submitted 2025-09-01 cs.CL

Can Smaller LLMs do better? Unlocking Cross-Domain Potential through Parameter-Efficient Fine-Tuning for Text Summarization

Anum Afzal , Mehul Kumawat , Florian Matthes This is my paper
classification cs.CL MSC 68T5068T07
keywords parameter-efficient fine-tuningtext summarizationdomain adaptationlow-resource domainsadaptersLlama-3cross-domain generalizationfactual consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that a relatively small 8B-parameter language model, fitted with lightweight adapter modules trained on high-resource summarization data, can outperform both few-shot prompting of the same model and a much larger 70B model on summarization in four unseen low-resource domains: scientific, medical, legal, and news. The authors benchmark six parameter-efficient fine-tuning methods across 14 training datasets and then test the trained adapters on fresh holdout corpora in each domain. Their central result is that within-domain adapters, built without any target-domain labels, beat the baselines on most automatic metrics, and that combining more than two adapters usually hurts. The paper also finds that when no within-domain adapter exists, cross-domain adapters chosen by linguistic and statistical similarity offer a practical fallback. If the result holds, specialized summarization capability can be delivered by a small modular model rather than by a much larger one, with no labeled target data.

Core claim

The core claim is that parameter-efficient fine-tuning on high-resource datasets transfers to low-resource summarization domains through reusable adapters: given a target domain with no labeled summaries, one can attach an adapter trained on a similar in-domain dataset (Within-Domain) and get better summaries than the base 8B model with few-shot examples or the 70B model in zero-shot. The paper proposes a two-stage selection pipeline: for each of 14 training datasets, train six PEFT variants and pick the best by rank aggregation across ROUGE, BERTScore, BLEU, METEOR, and FActScore; then designate the top-3 dataset adapters per domain as domain representatives for inference on the holdout set

What carries the argument

The carrying mechanism is the domain adapter: a small set of trainable parameters attached to the frozen self-attention blocks of Llama-3-8B-Instruct, trained on 1,000 samples from a high-resource dataset using one of six PEFT methods (AdaLoRA, (IA)3, LoHA, LoKr, LoRA, OFT). The selection pipeline, which trains all six on every dataset, ranks them with a multi-metric rank aggregation, and chooses the top-3 dataset adapters per domain as domain representatives, is what turns individual adapter accuracy into a claim about domain adaptation. The modularity of adapters is what allows Within-Domain, Cross-Domain, and combined-adapter inference without retraining the base model.

Load-bearing premise

The load-bearing premise is that the adapters reported as within-domain winners were fixed before evaluation on the holdout corpora; if the selection was influenced by those holdout results, the claimed advantage over the 70B baseline would not be a fair comparison.

What would settle it

Re-run the holdout evaluation with exactly the top-3 adapters declared in Table 12 (e.g., for Medical: CORD19-AdaLoRA, PubMed-LoRA, SciLay-OFT) and check whether any of them beats Llama-3-70B-Instruct on ROUGE; if none does, the headline result only holds for adapters selected with access to the holdout. An independent check would replace the four holdout corpora with fresh post-2024 articles in the same domains and verify that the same adapter lists keep their advantage.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A practitioner facing a new low-resource domain can, in principle, skip data collection and label annotation: train or fetch a high-resource adapter from a similar domain and attach it at inference time.
  • An 8B model with the right adapter can replace a 70B model for domain-specific summarization, cutting inference cost and memory requirements.
  • Cross-domain transfer works best when the target corpus is lexically and distributionally close to a training dataset, so adapter choice should be guided by similarity metrics.
  • Combining more than two adapters degrades performance, so adapter merging is not a free lunch; single well-chosen adapters usually suffice.
  • The same adapter bank can be reused across domains additively, since adapters can be swapped without touching the base model weights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's similarity-based account suggests a testable selection rule: before training any adapter, compute vocabulary and TF-IDF overlap between a candidate source and the unlabeled target, and the source with highest overlap should transfer best; this rule could be validated on new domains beyond the four studied.
  • The same transfer logic may extend to other long-input generation tasks such as question answering or report drafting, where in-domain labels are scarce and context windows are a bottleneck.
  • If the result generalizes across model families, the practical conclusion is not 'small beats large' but 'small plus modular expertise beats large generic'; the paper leaves open how adapter quality degrades as source-target similarity decreases.
  • A direct extension would be to pre-register the top-3 adapter lists before holdout evaluation and release the final selection logs, making the selection protocol auditable and easier to replicate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper benchmarks six PEFT methods with Llama-3-8B-Instruct on 14 training datasets from Scientific, Medical, Legal, and News domains, and then evaluates Within-Domain (WID) and Cross-Domain (CD) adapter configurations on 100-sample holdout validation sets per domain. The headline claim is that WID adapters can outperform few-shot Llama-3-8B and zero-shot Llama-3-70B on low-resource summarization domains. The paper also reports a human evaluation and a linguistic-similarity analysis intended to explain when WID/CD adapters help.

Significance. If the central claim held, the practical significance would be high: a single 8B model with reusable adapters could replace a much larger 70B model in low-resource summarization niches without target-domain labels. The paper also provides a broad PEFT benchmark across 14 datasets and four domains, and it makes training scripts and trained models available, which are useful community resources. However, as reported, the evaluation protocol does not support the headline comparison, and the secondary similarity-based explanation is internally inconsistent. The core empirical claim therefore needs to be re-established with a sound protocol before the paper can be accepted.

major comments (3)
  1. [§4.2, Tables 4 and 12] The WID configurations in Table 4 do not match the declared top-3 domain representatives in Table 12. For Medical, Table 12 lists CORD19-AdaLoRA, PubMed-LoRA, and SciLay-OFT, but Table 4's Medical WID rows include CORD19-OFT (not in Table 12) and a combination CORD19-AdaLoRA+MSLR-IA3. For Law, Table 4 lists MultiLex-LoKr, BillSum-AdaLoRA, and MultiLex-LoRA+BillSum-LoKr, none of which appear in the Law top-3 of Table 12. The News WID rows likewise do not match the declared top-3. Since Section 4.2 states that the top-3 are selected before holdout evaluation, Table 4 can only be explained by selecting the best adapter/combination using the holdout itself. The comparison against baselines in Table 6 is therefore circular, and the abstract's claim about WID adapters beating 70B is not supported by the stated protocol.
  2. [§5.3, Table 6; §3.2] All headline comparisons are on 100-sample holdouts per domain, yet no confidence intervals, significance tests, or multiple-comparison corrections are reported. The differences are often small—e.g., Medical ROUGE 0.2856 vs. 0.2575 for 8B and 0.2527 for 70B; Law ROUGE 0.2411 vs. 0.2360 and 0.2158—and may be within sampling noise. Because the paper searches over six PEFTs and many combinations across domains, the probability of finding a holdout-best configuration that beats a baseline by chance is nontrivial. Even if the selection issue in the first major comment were resolved, these results need uncertainty quantification before the superiority claim is established.
  3. [§5.2, Table 5] The similarity-based explanation is not supported by the paper's own table. The text states that SciTLDR-AdaLoRA achieves the best Scientific WID performance and that this 'aligns with high dataset similarity,' but Table 5 shows SciTLDR has the lowest vocabulary and TF-IDF overlap among the three Scientific WID training datasets (16.75 and 18.31, respectively), with only KL divergence and contextual overlap favoring it. The paper should state which similarity measure is being used and reconcile these conflicting signals; without that, the claimed correlation between similarity and adapter performance is not established.
minor comments (5)
  1. [Table 3] The footnote says '*' implies that the combination is among top-3 adapters, but the table rows are single dataset/PEFT pairs, not combinations. The wording should be clarified.
  2. [§4.3] The definition of ROUGE as the geometric mean of R1, R2, and RL appears only in the note under Table 4; it should be defined in the metrics subsection.
  3. [Table 5] The KL divergence column should state whether lower or higher values indicate greater similarity. Currently the reader must infer the direction.
  4. [Ethical Statement] The statement says all training scripts and trained models are made available, but no repository URL or access link is provided. The camera-ready version should include one.
  5. [§5.2 Human Evaluation] The human evaluation uses only 25 articles per domain with two annotators. Reporting per-domain top-1 agreement and possibly a chance-level baseline would help interpret the 74.5% aggregate agreement.

Circularity Check

1 steps flagged

Table 4's Within-Domain adapters are not the declared top-3 of Table 12, indicating holdout-based adapter selection and making the headline WID-vs-baseline claim partially circular.

specific steps
  1. fitted input called prediction [Section 4.2 (Domain Adapters), Table 4 (WID rows), Table 12 (top-3 domain adapters)]
    "Section 4.2: "We select the top-3 dataset-specific adapters from each domain as domain representatives, which are subsequently used in Within-Domain and Cross-Domain experiments on the Holdout Datasets." Table 12 Medical top-3: "CORD19-AdaLoRA ... PubMed-LoRA ... SciLay-OFT". Table 4 Medical WID: "CORD19-OFT ... PubMed-LoRA ... CORD19-AdaLoRA + MSLR-IA3"."

    The stated protocol fixes the WID configurations before looking at the holdout: they must be the Table 12 top-3. Table 4 violates this for Medical (CORD19-OFT is not in the top-3 and is not even the best PEFT for CORD19 on source data; Table 3 lists CORD19-AdaLoRA), and similar mismatches occur for Scientific, Law, and News. The paper provides no alternative pre-registered criterion that would select CORD19-OFT. The only evident way to obtain the Table 4 WID rows is to evaluate a broader set of adapters on the target holdout and keep the best ones. The reported 'Within-Domain Adapters achieve better performance than Few-Shot as well as a much larger Llama-3-70B-Instruct' is then not an independent prediction but the result of selecting the adapter on the same data used to measure the win,

full rationale

The paper's central derivation chain is: train six PEFTs on 14 source datasets, rank them on source test sets, pick top-3 per domain as representatives, then evaluate these on four holdout domains and compare to few-shot/70B. The final comparison is the load-bearing evidence for the abstract claim. That step is compromised because Table 4's WID rows do not match the declared top-3 representatives in Table 12. If the Table 4 configurations were selected by holdout peeking, the headline result reduces to an optimistic selection on the evaluation set, which is a form of fitted-input-called-prediction circularity. The mismatch is concrete and quotable, so this is not a vague concern. I do not count the self-citation of Afzal et al. 2024 (Section 4.5) as load-bearing: it only motivates the choice of Llama-3-8B, and the comparison to 70B would stand or fall independently. There is also no imported uniqueness theorem or ansatz-by-citation here. The score is 6 rather than higher because the trained adapters, baseline numbers, and source-domain rankings are real and independently checkable; the circularity is in the configuration-selection step that produces the headline comparison, not in the training or metric definitions.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces no new theoretical entities, but it relies on several domain assumptions: that the holdout is uncontaminated, that adapters can be summed/merged, that Borda Count is an appropriate aggregation, and that lexical similarity predicts transfer. The main unstated parameter is the choice of which adapter configurations to report, which appears to be made after seeing the holdout results.

free parameters (3)
  • PEFT hyperparameters (rank 64, alpha 8, lr 0.0005, 5 epochs) = rank=64, alpha=8, lr=0.0005, epochs=5
    Chosen by hand and held constant across all PEFTs; these values affect adapter quality and therefore the central transfer claim.
  • Training sample count and selection = 1000 samples per dataset, shortest articles first
    Training is limited to 1000 samples due to compute, and the shortest articles are selected to fit the 4096 context window; this biases the domain representation.
  • Number of top adapters per domain = 3
    The top-k threshold for selecting domain representatives is arbitrary and affects which adapters are later evaluated on the holdout.
axioms (4)
  • domain assumption Holdout validation articles are dated after Llama 3 release and thus untainted
    Stated in Section 3.2, but the legal holdout from Shukla et al. 2022 may predate Llama 3, so the contamination control is not consistently enforced.
  • domain assumption Adapter weights can be merged and combined without loss of specialization
    Implicit in the combination experiments (Section 4.2, Table 4); the merging method is not specified, yet the central claims depend on it.
  • standard math Borda Count ranking across multiple metrics is a valid aggregate
    Used to rank PEFTs and adapters (Section 4.3); assumes the metrics are commensurable and that aggregation by Borda Count preserves meaningful differences.
  • domain assumption Dataset similarity metrics (vocab overlap, TF-IDF, KL divergence) correlate with transferability
    Used in Section 5.2 to explain why some adapters transfer better; the correlation is observed, not causally validated.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Smaller LLMs do better? Unlocking Cross-Domain Potential through Parameter-Efficient Fine-Tuning for Text Summarization." pith.science (2026). https://pith.science/paper/GDIEZUCQ

@misc{pith2026250901314,
  author       = {Pith},
  title        = {Pith review of: Can Smaller LLMs do better? Unlocking Cross-Domain Potential through Parameter-Efficient Fine-Tuning for Text Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GDIEZUCQ}},
  note         = {Machine review of arXiv:2509.01314}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs), being generic task solvers, are versatile. However, despite the vast amount of data they are trained on, there are speculations about their adaptation capabilities to a new domain. Additionally, the simple fine-tuning of the model to incorporate knowledge of a new domain is computationally expensive and time-consuming. This becomes more challenging when the domain in question is also low-resource, and labeled data is unavailable. We leverage parameter-efficient fine-tuning techniques (PEFTs) on high-resource datasets to address these challenges to improve performance on unseen low-resource domains. Throughout our experiments, we evaluate whether intrinsic linguistic commonalities between datasets can be leveraged for efficient domain adaptation. We benchmark six PEFTs with \texttt{Llama-3-8B-Instruct} on 14 training datasets from the Scientific, Medical, Legal, and News domains for a Text Summarization task. Our experiments show that for low-resource domains, inference using Within-Domain Adapters can achieve better performance than Few-Shot as well as a much larger \texttt{Llama-3-70B-Instruct}. Lastly, in the absence of Within-Domain Adapters, we explore the concept of using Cross-Domain Adapters as well as the strategic combinations of adapters to leverage intrinsic language similarities across domains, facilitating better adaptability and performance in low-resource settings.

Figures

Figures reproduced from arXiv: 2509.01314 by Anum Afzal, Florian Matthes, Mehul Kumawat.

Figure 1
Figure 1. Figure 1: Our methodology focuses on 1) benchmarking state-of-the-art adapters on all datasets from a domain, to [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training and Evaluation pipeline for PEFT selection for each dataset. A similar approach is used for [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Average Perplexity Scores per domain computed during fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 17 canonical work pages · 6 internal anchors

  1. [1]

    Anum Afzal, Ribin Chalumattu, Florian Matthes, and Laura Mascarell. 2024. https://doi.org/10.18653/v1/2024.customnlp4u-1.8 A dapt E val: Evaluating large language models on domain adaptation for text summarization . In Proceedings of the 1st Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Indiv...

    work page doi:10.18653/v1/2024.customnlp4u-1.8 2024

  2. [2]

    Dennis Aumiller, Ashish Chouhan, and Michael Gertz. 2022. https://arxiv.org/abs/2210.13448 EUR-Lex-Sum: A Multi- and Cross-lingual Dataset for Long-form Summarization in the Legal Domain . CoRR, abs/2210.13448

    Pith/arXiv arXiv 2022

  3. [3]

    Charith Chandra Sai Balne, Sreyoshi Bhaduri, Tamoghna Roy, Vinija Jain, and Aman Chadha. 2024. https://arxiv.org/abs/2404.13506 Parameter efficient fine tuning: A comprehensive analysis across applications . Preprint, arXiv:2404.13506

    Pith/arXiv arXiv 2024

  4. [4]

    Satanjeev Banerjee and Alon Lavie. 2005. https://aclanthology.org/W05-0909/ 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 Compu...

    2005

  5. [5]

    Do, Yan Xu, and Pascale Fung

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. https://arxiv.org/abs/2302.04023 A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity . Preprint, arXiv:2302.04023

    Pith/arXiv arXiv 2023

  6. [6]

    Isabel Cachola, Kyle Lo, Arman Cohan, and Daniel S. Weld. 2020. TLDR : Extreme summarization of scientific documents. arXiv:2004.15011

    Pith/arXiv arXiv 2020

  7. [7]

    Yu, and Lichao Sun

    Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S. Yu, and Lichao Sun. 2023. https://arxiv.org/abs/2303.04226 A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt . Preprint, arXiv:2303.04226

    Pith/arXiv arXiv 2023

  8. [8]

    Efficient In-Domain Question Answering for Resource-Constrained Environments

    Isaac Chung, Phat Vo, Arman C. Kizilkale, and Aaron Reite. 2024. https://arxiv.org/abs/2409.17648 Efficient in-domain question answering for resource-constrained environments . Preprint, arXiv:2409.17648

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. https://doi.org/10.18653/v1/N18-2097 A discourse-aware attention model for abstractive summarization of long documents . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human...

    doi:10.18653/v1/n18-2097 2018

  10. [10]

    Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R

    Alexander R. Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R. Radev. 2019. https://arxiv.org/abs/1906.01749 Multi-news: a large-scale multi-document summarization dataset and abstractive hierarchical model . Preprint, arXiv:1906.01749

    Pith/arXiv arXiv 2019

  11. [11]

    Max Grusky, Mor Naaman, and Yoav Artzi. 2018. http://aclweb.org/anthology/N18-1065 Newsroom: A dataset of 1.3 million summaries with diverse extractive strategies . In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 708--719, New Orleans, Louisiana. Assoc...

    work page 2018

  12. [12]

    Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. https://proceedings.neurips.cc/paper_files/paper/2015/file/afdec7005cc9f14302cd0474fd0f3c96-Paper.pdf Teaching machines to read and comprehend . In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc

    work page 2015

  13. [13]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://proceedings.mlr.press/v97/houlsby19a.html Parameter-efficient transfer learning for NLP . In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine ...

    work page 2019

  14. [14]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . Preprint, arXiv:2106.09685

    Pith/arXiv arXiv 2021

  15. [15]

    Yukun Huang, Yanda Chen, Zhou Yu, and Kathleen McKeown. 2022. https://arxiv.org/abs/2212.10670 In-context learning distillation: Transferring few-shot learning ability of pre-trained language models . Preprint, arXiv:2212.10670

    Pith/arXiv arXiv 2022

  16. [16]

    Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. 2023. https://arxiv.org/abs/2108.06098 Fedpara: Low-rank hadamard product for communication-efficient federated learning . Preprint, arXiv:2108.06098

    Pith/arXiv arXiv 2023

  17. [17]

    Daniel Kershaw and Rob Koeling. 2020. https://doi.org/10.17632/zm33cdndxs.2 Elsevier oa cc-by corpus . Mendeley

    work page doi:10.17632/zm33cdndxs.2 2020

  18. [18]

    Anastassia Kornilova and Vladimir Eidelman. 2019. https://doi.org/10.18653/v1/D19-5406 B ill S um: A corpus for automatic summarization of US legislation . In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 48--56, Hong Kong, China. Association for Computational Linguistics

    doi:10.18653/v1/d19-5406 2019

  19. [19]

    Chin-Yew Lin. 2004. https://www.aclweb.org/anthology/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

  20. [20]

    Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022. https://arxiv.org/abs/2205.05638 Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning . Preprint, arXiv:2205.05638

    Pith/arXiv arXiv 2022

  21. [21]

    Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2025. https://arxiv.org/abs/2308.08747 An empirical study of catastrophic forgetting in large language models during continual fine-tuning . Preprint, arXiv:2308.08747

    Pith/arXiv arXiv 2025

  22. [22]

    Iain McLean. 1990. https://doi.org/10.1007/BF01560577 The borda and condorcet principles: Three medieval applications . Social Choice and Welfare, 7:99--108

    doi:10.1007/bf01560577 1990

  23. [23]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://arxiv.org/abs/2305.14251 Factscore: Fine-grained atomic evaluation of factual precision in long form text generation . Preprint, arXiv:2305.14251

    Pith/arXiv arXiv 2023

  24. [24]

    Cohen, and Mirella Lapata

    Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. https://arxiv.org/abs/1808.08745 Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization . Preprint, arXiv:1808.08745

    Pith/arXiv arXiv 2018

  25. [25]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 Bleu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL '02, page 311–318, USA. Association for Computational Linguistics

    arXiv 2002

  26. [26]

    Karmvir Singh Phogat, Sai Akhil Puranam, Sridhar Dasaratha, Chetan Harsha, and Shashishekar Ramakrishna. 2024. https://arxiv.org/abs/2408.12337 Fine-tuning smaller language models for question answering over financial documents . Preprint, arXiv:2408.12337

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. 2024. https://arxiv.org/abs/2306.07280 Controlling text-to-image diffusion by orthogonal finetuning . Preprint, arXiv:2306.07280

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    Luca Ragazzi, Paolo Italiani, Gianluca Moro, and Mattia Panni. 2024. https://doi.org/10.18653/v1/2024.findings-acl.561 What are you token about? differentiable perturbed top- k token selection for scientific document summarization . In Findings of the Association for Computational Linguistics: ACL 2024, pages 9427--9440, Bangkok, Thailand. Association for...

    work page doi:10.18653/v1/2024.findings-acl.561 2024

  29. [29]

    Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. 2017. https://arxiv.org/abs/1705.08045 Learning multiple visual domains with residual adapters . Preprint, arXiv:1705.08045

    Pith/arXiv arXiv 2017

  30. [30]

    Joshua Robinson, Christopher Michael Rytting, and David Wingate. 2023. https://arxiv.org/abs/2210.12353 Leveraging large language models for multiple choice question answering . Preprint, arXiv:2210.12353

    Pith/arXiv arXiv 2023

  31. [31]

    Zejiang Shen, Kyle Lo, Lauren Yu, Nathan Dahlberg, Margo Schlanger, and Doug Downey. 2022. https://doi.org/10.48550/arXiv.2206.10883 Multi-lexsum: Real-world summaries of civil rights lawsuits at multiple granularities . CoRR, abs/2206.10883

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2206.10883 2022

  32. [32]

    Ken Shi and Gerald Penn. 2025. https://aclanthology.org/2025.wraicogs-1.2/ Semantic masking in a needle-in-a-haystack test for evaluating large language model long-text capabilities . In Proceedings of the First Workshop on Writing Aids at the Crossroads of AI, Cognitive Science and NLP (WRAICOGS 2025), pages 16--23, Abu Dhabi, UAE. International Committe...

    work page 2025

  33. [33]

    Abhay Shukla, Paheli Bhattacharya, Soham Poddar, Rajdeep Mukherjee, Kripabandhu Ghosh, Pawan Goyal, and Saptarshi Ghosh. 2022. https://arxiv.org/abs/2210.07544 Legal case document summarization: Extractive and abstractive methods and their evaluation . Preprint, arXiv:2210.07544

    Pith/arXiv arXiv 2022

  34. [34]

    Sotaro Takeshita, Tommaso Green, Ines Reinig, Kai Eckert, and Simone Paolo Ponzetto. 2024. https://arxiv.org/abs/2403.05303 Aclsum: A new dataset for aspect-based summarization of scientific publications . Preprint, arXiv:2403.05303

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Generating (Factual?) Narrative Summaries of RCTs: Experiments with Neural Multi-Document Summarization

    Byron C. Wallace, Sayantani Saha, Frank Soboczenski, and Iain James Marshall. 2020. Generating (factual?) narrative summaries of rcts: Experiments with neural multi-document summarization. AMIA Annual Symposium, abs/2008.11293

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  36. [36]

    Funk, Rodney Michael Kinney, Ziyang Liu, W

    Lucy Lu Wang, Kyle Lo, Yoganand Chandrasekhar, Russell Reas, Jiangjiang Yang, Darrin Eide, K. Funk, Rodney Michael Kinney, Ziyang Liu, W. Merrill, P. Mooney, D. Murdick, Devvret Rishi, Jerry Sheehan, Zhihong Shen, B. Stilson, A. Wade, K. Wang, Christopher Wilhelm, Boya Xie, D. Raymond, Daniel S. Weld, Oren Etzioni, and Sebastian Kohlmeier. 2020. Cord-19: ...

    work page 2020

  37. [37]

    Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. https://arxiv.org/abs/2304.13712 Harnessing the power of llms in practice: A survey on chatgpt and beyond . Preprint, arXiv:2304.13712

    Pith/arXiv arXiv 2023

  38. [38]

    Zhuoyi Yang, Ming Ding, Yanhui Guo, Qingsong Lv, and Jie Tang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.514 Parameter-efficient tuning makes a good classification head . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7576--7586, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics

    work page doi:10.18653/v1/2022.emnlp-main.514 2022

  39. [39]

    Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard B W Yang, Giyeong Oh, and Yanmin Gong. 2024. https://arxiv.org/abs/2309.14859 Navigating text-to-image customization: From lycoris fine-tuning to model evaluation . Preprint, arXiv:2309.14859

    Pith/arXiv arXiv 2024

  40. [40]

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. https://arxiv.org/abs/2303.10512 Adalora: Adaptive budget allocation for parameter-efficient fine-tuning . Preprint, arXiv:2303.10512

    Pith/arXiv arXiv 2023

  41. [41]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://arxiv.org/abs/1904.09675 Bertscore: Evaluating text generation with bert . Preprint, arXiv:1904.09675

    Pith/arXiv arXiv 2020

  42. [42]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023. Lima: less is more for alignment. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA. Curran Ass...

    work page 2023

  43. [43]

    Xin Zhou, Ruotian Ma, Yicheng Zou, Xuanting Chen, Tao Gui, Qi Zhang, Xuanjing Huang, Rui Xie, and Wei Wu. 2022. https://aclanthology.org/2022.coling-1.615/ Making parameter-efficient tuning more efficient: A unified framework for classification tasks . In Proceedings of the 29th International Conference on Computational Linguistics, pages 7053--7064, Gyeo...

    work page 2022

  44. [44]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  45. [45]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.