Pith. sign in

REVIEW 4 major objections 8 minor 47 references

MedReadCtrl: Personalizing medical text generation with readability-controlled instruction learning

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MedReadCtrl claims that instruction tuning on explicit grade-level targets gives an 8-billion-parameter open model finer, meaning-preserving control over medical text readability than GPT-4, with lower readability instruction-following…

desk verdict Plausible and useful framework, but the paper's own text and tables disagree about what the model actually trained on, so the generalization claims cannot be verified as written. read the letter →

arxiv 2507.07419 v1 pith:TK3IBJDU submitted 2025-07-10 cs.CL cs.AI

Hieu Tran , Zonghai Yao , Won Seok Jang , Sharmin Sultana , Allen Chang , Yuan Zhang , Hong Yu This is my paper
classification cs.CLcs.AI
keywords readabilitycontrolinstructiontuningmedicaltextsimplificationhealthliteracypatienteducationlargelanguagemodelsgenerationsemanticentailment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a large language model can be taught to write medical text at any requested reading grade level, from first grade through twelfth, while keeping the clinical meaning intact. The authors build MedReadCtrl by instruction-tuning an open 8-billion-parameter model on simplification, paraphrase, and entailment examples tagged with a target readability score, then compare it with GPT-4, GPT-3.5, and Claude-3. Across nine datasets, the tuned model follows grade-level instructions more closely (readability error 1.39 vs. 1.59 on one medical dataset, p<0.001) and improves automatic quality scores on unseen clinical text, with physicians preferring its output 71.7% of the time. If the claim holds, patient education materials, discharge instructions, and consent forms could be generated at each reader's literacy level, directly addressing health-literacy disparities.

What carries the argument

The load-bearing mechanism is readability-controlled instruction learning: the paper fine-tunes a base LLM on instructions of the form "output an entailment with a readability score around target readability score," where each training pair is built from six source datasets and assigned a target grade from 1 to 12. Readability is measured by a single scalar, the average Reading Grade Level (RGL) of four surface-level formulas: Flesch-Kincaid Grade Level, Gunning Fog Index, Automated Readability Index, and Coleman-Liau Index. The same RGL average is used both to construct the 12-grade training buckets and to score instruction-following errors at evaluation, so the target and the measurement stick are the same four-formula average. The whole claim of personalized comprehension rests on that scalar proxy.

What would settle it

Give adults with low health literacy the same clinical note simplified by MedReadCtrl at Grades 2, 5, 8, and 11 and measure their understanding with multiple-choice or cloze questions; the central claim would be refuted if comprehension does not improve as the target grade falls, or if the Grade-2 version is understood no better than the original note.

Watch

Extended reading notes

Core claim

The paper's central claim is that readability can be made a controllable output dimension of a language model without sacrificing medical meaning. The authors assert that LLaMA-3-MedReadCtrl, an open 8-billion-parameter model instruction-tuned on simplification, paraphrase, and entailment examples labeled with target grade levels, follows explicit readability instructions more precisely than GPT-4 across medical datasets (mean absolute readability error 1.39 vs. 1.59 on ReadMe, p<0.001) and keeps that advantage on general-domain text. They further claim large automatic-metric gains on unseen clinical tasks, notably +14.7 ROUGE-L and +6.18 SARI on MTSamples, and that two physician evaluators preferred its outputs over GPT-4's by 71.7% to 23.3%, with the gap largest at the lowest requested grade levels. The intended upshot is that patient education, discharge instructions, and consent materials can be generated at a reader's literacy level while preserving clinical intent.

Load-bearing premise

The load-bearing premise is that an unweighted average of four surface-level readability formulas (word length, sentence length, syllable count, character count) is a valid proxy for a patient's actual comprehension, because that same proxy is used to build the 12-grade training data and to measure whether the model followed the instruction; if the proxy does not track comprehension of clinical language, the model hits the numeric target without delivering true personalization.

Editorial extensions

If this is right

  • An open, 8-billion-parameter model can follow readability instructions more precisely than much larger proprietary models, so fine-grained patient-level language control does not require access to a frontier API.
  • The control transfers to datasets never used in training, including MTSamples, WikiSmall, MultiNLI, and MRPC, which argues the effect is a learned general capability rather than memorized outputs.
  • Simplification for very low literacy levels need not trade away clinical fidelity: accuracy and consistency scores stayed high at Grades 2 and 5, and expert clarity ratings improved sharply over GPT-4.
  • Because physicians preferred the tuned model 71.7% versus 23.3% with strong inter-rater agreement (0.807), the readability difference is visible to expert human judgment, not only to automatic metrics.
  • Deployment would still need safeguards: the same evaluation shows occasional hallucinations, overly technical verbosity at Grade 11, and misread instructions, which the authors propose to address with retrieval grounding and preference-based training.

Reading between the lines

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

  • The paper does not test whether hitting a target RGL actually changes comprehension for the end users; its evaluators were physicians and LLM judges, not patients or caregivers. A direct comprehension study with low-literacy readers would be the natural next experiment.
  • The same instruction loop could, in principle, control any text property that can be scored automatically, such as length, tone, jargon density, or formality, since readability is just one scalar training signal; this is an extension the authors gesture at but do not develop.
  • Because the readability proxy is an average of word- and sentence-length formulas, a model could learn to game the number by producing short sentences and one-syllable words without making the underlying clinical concepts easier; checking MedReadCtrl's grades against cloze or recall scores would reveal how much of the control is substantive.
  • A practical deployment path the authors leave implicit is coupling a validated health-literacy screening score to the grade dial, so the requested readability level is chosen by the patient's measured literacy rather than by a generic default.
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 / 8 minor

Summary. The paper introduces MedReadCtrl, a readability-controlled instruction-tuning framework built on LLaMA3-8B, and evaluates it against GPT-4, GPT-3.5, and Claude-3 on nine datasets spanning text simplification, paraphrase generation, and semantic entailment generation in both general and medical domains. The authors report that MedReadCtrl achieves significantly lower readability instruction-following errors than GPT-4 across ReadMe, MTSamples, and MedNLI, and that it improves content-preservation metrics (ROUGE, BLEU, SARI) on both seen and unseen datasets, with human experts preferring its outputs 71.7% of the time overall. The central claim is that instruction tuning on readability targets gives an LLM fine-grained, meaning-preserving control over text complexity, which could enable personalized patient education materials.

Significance. If the central claim holds, the work addresses a real need in patient-facing medical communication: producing text at a specified literacy level while preserving meaning. The paper's strengths include the breadth of evaluation across multiple tasks and domains, the inclusion of human expert judgments with reported inter-annotator agreement (Cohen's Kappa 0.807), and the public release of code and dataset links. However, the significance is contingent on the experimental record being internally consistent and on the readability proxy being a valid stand-in for patient comprehension. The paper ships neither machine-checked proofs nor a fully reproducible training pipeline; the reported results cannot currently be independently verified because the training-data construction is not described. The contradiction between Section 4.3 and Table 3 on which datasets were used for instruction tuning is a serious obstacle to assessing the generalization claims.

major comments (4)
  1. [§4.3 vs. Table 3 and Figure 1] Section 4.3 states that 'instruction tuning is performed on the training sets of ASSET, PAWS, and SNLI' and defines the seen setting as their test sets, with WikiSmall, MRPC, and MultiNLI as unseen. However, Table 3, captioned 'Distribution of examples readability scores from instruction tuning datasets,' lists train/dev/test splits for ReadMe and MedNLI, and Table 1 and Figure 1 label ReadMe and MedNLI as 'seen.' These statements are mutually incompatible. If ReadMe and MedNLI were part of instruction tuning, then the reported superiority over GPT-4 on those datasets is in-domain fine-tuning rather than evidence of a general readability-control capability, and the 'unseen' claim rests solely on MTSamples. If they were not part of tuning, the caption of Table 3 and the 'seen' labels are incorrect. Either way, the method as described is not reproducible and the main results cannot be verified from the text.
  2. [§4.3 and Figure 3] The construction of the grade-specific instruction-tuning corpus is not described. Figure 3 is labeled 'Overview of MedReadCtrl data construction,' and Table 3 gives per-grade example counts, but the text does not specify how source pairs from ASSET, PAWS, SNLI (and, if applicable, ReadMe and MedNLI) were transformed into 12 grade-specific training examples. The paper does not state how target readability grades were assigned, whether examples were filtered or regenerated using the RGL formula, which instruction templates were used for the simplification and paraphrase tasks, or how the train/dev/test splits were derived. Without this information, the method cannot be reproduced, and the grade distributions in Table 3 cannot be checked for systematic bias or contamination.
  3. [Eq. (1) in §4.2] Equation (1) labels the metric as 'Flesch-Kincaid Grade Level (FKGL)' but gives the formula for the Flesch Reading Ease score: 206.835 - 1.015*(totalWords/totalSentences) - 84.6*(totalSyllables/totalWords). The standard Flesch-Kincaid Grade Level formula is 0.39*(totalWords/totalSentences) + 11.8*(totalSyllables/totalWords) - 15.59. Because the average Reading Grade Level (RGL) is computed from these four formulas and is used both to construct training targets and to measure instruction-following error in Figure 1, the numerical results depend directly on which formula was actually implemented. The paper must correct this equation or state that the Flesch Reading Ease score was used under the wrong name, and it must re-run the affected evaluations if the implementation matched the equation as written.
  4. [§4.2 and Figure 1] The claimed 'readability control' is measured with the same unweighted average of four surface-level formulas (Eqs. 1-4) that define the training targets. The instruction-following errors in Figure 1 therefore largely measure how well the model fits its training objective, not how comprehensible the output is to a patient at a given literacy level. The Discussion partially acknowledges this ('our use of traditional readability metrics ... may be insufficient to assess fluency and understandability in fragmented or telegraphic clinical texts'), but the abstract and introduction make the stronger claim that the system 'personalizes' text by literacy level without compromising meaning. To support that claim, the evaluation should include at least one comprehension-oriented measure not used in training (for example, expert ratings of grade-level appropriateness, which are already collected in Table 2 but are reported separately and not used in Figure 1).
minor comments (8)
  1. [§5] In Data Availability, 'Night datasets' should read 'Nine datasets.'
  2. [Throughout] The model name is written inconsistently: 'LLaMA3-MedReadCtrl,' 'LLAMA3-MedReadCtrl,' and 'Llama3-MedReadCtrl' all appear (e.g., in Table 1, Figure 1 captions, and Section 4.5). Please standardize.
  3. [§4.2] The instruction template quoted in Section 4.2, 'Given an input text, please output an entailment with a readability score around target readability score,' is specific to the entailment task; the paper should provide the corresponding instruction templates used for text simplification and paraphrase generation, since these are separate tasks with different target relationships between input and output readability.
  4. [Table 3] The caption of Table 3 says 'Distribution of examples readability scores from instruction tuning datasets,' but the table also includes rows for MTSamples, WikiSmall, MultiNLI, and MRPC test sets. The caption should be revised to distinguish tuning datasets from held-out evaluation datasets.
  5. [Table 2] Table 2 uses green/red/underline formatting to indicate 'Ours better,' 'GPT-4 better,' and 'Same.' This formatting will not be visible in grayscale print or for color-blind readers; please add explicit textual markers such as plus/minus signs or boldface.
  6. [§4.8 and Figure 2] Section 4.8 says AI evaluation was conducted using GPT-3.5, GPT-4, and Claude 3, but Figure 2 reports a single 'AI judge' preference percentage (e.g., 58.3%, 60.0%). Please specify which model or ensemble produced the reported numbers, and report per-judge results if multiple models were used.
  7. [§4.7] The wording 'we randomly sampled 10 data from the test datasets of 3 data sets' and '6 data' should be revised to '10 examples' and '6 examples' for clarity.
  8. [Appendix] Tables 7-9 are referenced in the Discussion but are not labeled as tables in the main text; please include explicit cross-references and consistent numbering.

Circularity Check

0 steps flagged · score 2.0 of 10

No by-construction circularity; the readability metric doubles as the training target, and a Section 4.3/Table 3 training-set contradiction blocks verification without constituting a circular reduction.

full rationale

The paper's contribution is an empirical instruction-tuning pipeline, not a formal derivation, so there is no theorem whose conclusion is identical to its premises. The closest concern is that Section 4.2 derives an average Reading Grade Level (RGL) from FKGL, GFI, ARI, and CLI, and the same RGL is used both to define the target readability in the instruction data and to compute the Figure 1 'readability instruction-following errors.' This makes the low absolute error on that axis unsurprising, but the paper never defines a training loss that is literally the RGL error, and the headline comparison against GPT-4 on held-out test splits, the external ROUGE/BLEU/SARI references, and the MTSamples generalization results are independent of the training objective. The ReadMe and BioInstruct self-citations are dataset and method resources, not load-bearing uniqueness theorems, so the self-citation pattern does not apply. A more serious flaw is internal inconsistency: Section 4.3 states that 'instruction tuning is performed on the training sets of ASSET, PAWS, and SNLI,' while Table 3 lists ReadMe and MedNLI train/dev splits under 'instruction tuning datasets' and Table 1 labels ReadMe and MedNLI as 'seen.' This prevents verification of which results are in-domain, but it is a reproducibility and attribution failure rather than a by-construction circularity. Overall, no circular step can be established from the paper's own equations; the score of 2 reflects the metric overlap and unresolved training-set ambiguity without claiming a circular reduction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two unverified assumptions: that the average of four surface-level readability formulas is a valid comprehension proxy for clinical text, and that the undisclosed grade-specific training-data construction is unbiased. The human evaluation of meaning preservation also rests on ratings by two co-authors. No invented entities are introduced.

free parameters (1)
  • readability metric weights = 0.25 each, unweighted average of FKGL, GFI, ARI, CLI
    Section 4.2 defines RGL as the average of four readability formulas. This hand-chosen aggregation is used both to construct the training data and to evaluate readability adherence, so the central claim of readability control depends on it.
assumptions (3)
  • domain assumption The average of FKGL, GFI, ARI, and CLI is a valid measure of text readability/comprehension for clinical content.
    Section 4.2 defines RGL as this average and uses it for both training data construction and evaluation. These formulas ignore medical jargon, semantics, and discourse structure, so the validity of this proxy is load-bearing for the claim of personalized readability.
  • ad hoc to paper A grade-specific instruction-tuning corpus can be generated from the source datasets without systematic bias or contamination.
    Section 4.3 and Figure 3 present the data construction overview, but the actual generation method (which LLM, prompts, filtering criteria) is not described. The central results depend on this unseen pipeline.
  • domain assumption The two clinician authors' ratings are a reliable measure of medical accuracy and meaning preservation.
    Section 4.7 and Table 2: human evaluation of accuracy and consistency is performed by co-authors A.C. and Y.Z. The claim 'without compromising meaning' rests on their judgment. No independent rater pool or inter-rater reliability beyond the same two authors is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedReadCtrl: Personalizing medical text generation with readability-controlled instruction learning." pith.science (2026). https://pith.science/paper/TK3IBJDU

@misc{pith2026250707419,
  author       = {Pith},
  title        = {Pith review of: MedReadCtrl: Personalizing medical text generation with readability-controlled instruction learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TK3IBJDU}},
  note         = {Machine review of arXiv:2507.07419}
}
read the original abstract

Generative AI has demonstrated strong potential in healthcare, from clinical decision support to patient-facing chatbots that improve outcomes. A critical challenge for deployment is effective human-AI communication, where content must be both personalized and understandable. We introduce MedReadCtrl, a readability-controlled instruction tuning framework that enables LLMs to adjust output complexity without compromising meaning. Evaluations of nine datasets and three tasks across medical and general domains show that MedReadCtrl achieves significantly lower readability instruction-following errors than GPT-4 (e.g., 1.39 vs. 1.59 on ReadMe, p<0.001) and delivers substantial gains on unseen clinical tasks (e.g., +14.7 ROUGE-L, +6.18 SARI on MTSamples). Experts consistently preferred MedReadCtrl (71.7% vs. 23.3%), especially at low literacy levels. These gains reflect MedReadCtrl's ability to restructure clinical content into accessible, readability-aligned language while preserving medical intent, offering a scalable solution to support patient education and expand equitable access to AI-enabled care.

Figures

Figures reproduced from arXiv: 2507.07419 by the authors.

Figure 1
Figure 1. Comparison of readability control precision between our LLAMA3-MedReadCtrl model and GPT-4. (a) LLAMA3-MedReadCtrl demonstrates significantly lower errors in readability control across medical datasets. (b) presents the ability of different LLMs to follow readability-control instructions in general-domain texts. The y-axis represents the model’s actual readability output, while the x-axis denotes the target readabil… view at source ↗
Figure 2
Figure 2. Human preference. Cohen’s Kappa for two medical experts: 0.807. with an average score of 4.56 compared to GPT-4’s 4.17. Notably, LLAMA3-MedReadCtrl demonstrates particularly strong gains at lower readability levels, where simplifying biomedical content is most challenging. For Grade 2 outputs, our model significantly outperforms GPT-4 in Clarity across all datasets—ReadMe (4.05 vs. 2.60), MTSamples (4.75 vs. 3.10), … view at source ↗
Figure 3
Figure 3. Overview of MedReadCtrl data construction. We employ the instruction tuning approach conditional on readability for all these tasks. This method provides explicit instructions to the model to control the output text’s readability score, ensuring that the generated content aligns with the intended complexity level for the target audience. This approach underlines our belief that these tasks can all contribute to read… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 32 canonical work pages

  1. [1]

    & Ramasamy, R

    Golan, R., Reddy, R. & Ramasamy, R. The rise of artificial intelligence-driven health communication.Transl. Androl. Urol.13, 356 (2024). 2.Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. Ai in health and medicine.Nat. medicine28, 31–38 (2022)

    work page 2024

  2. [3]

    Medicine7, 88 (2024)

    Silcox, C.et al.The potential for artificial intelligence to transform healthcare: perspectives from international health leaders.NPJ Digit. Medicine7, 88 (2024). 4.Yao, Z. & Yu, H. A survey on llm-based multi-agent ai hospital.OSF(2025)

    work page 2024

  3. [5]

    & Hallowell, N

    Sauerbrei, A., Kerasidou, A., Lucivero, F. & Hallowell, N. The impact of artificial intelligence on the person-centred, doctor-patient relationship: some problems and solutions.BMC Med. Informatics Decis. Mak.23, 73 (2023)

    work page 2023

  4. [6]

    L.et al.Reviewing the potential role of artificial intelligence in delivering personalized and interactive pain medicine education for chronic pain patients.J

    Robinson, C. L.et al.Reviewing the potential role of artificial intelligence in delivering personalized and interactive pain medicine education for chronic pain patients.J. Pain Res.923–929 (2024). 7.Delbanco, T.et al.Open notes: doctors and patients signing on (2010)

    work page 2024

  5. [8]

    About the blue button movement

    HealthIT.gov. About the blue button movement. https://www.healthit.gov/patients-families/about-blue-button-movement (2024). [accessed 2024-10-29]

    work page 2024

  6. [9]

    Artificial intelligence and health literacy—proceed with caution.Heal

    Nutbeam, D. Artificial intelligence and health literacy—proceed with caution.Heal. Lit. Commun. Open1, 2263355 (2023)

    work page 2023

  7. [10]

    & Mazur, L

    Khasawneh, A., Kratzke, I., Adapa, K., Marks, L. & Mazur, L. Effect of notes’ access and complexity on opennotes’ utility. Appl. Clin. Informatics13, 1015–1023 (2022). 11.Meng, X.et al.The application of large language models in medicine: A scoping review.Iscience27(2024). 12.Achiam, J.et al.Gpt-4 technical report.arXiv preprint arXiv:2303.08774(2023)

    arXiv 2022

  8. [13]

    H.et al.Performance of chatgpt on usmle: potential for ai-assisted medical education using large language models.PLoS digital health2, e0000198 (2023)

    Kung, T. H.et al.Performance of chatgpt on usmle: potential for ai-assisted medical education using large language models.PLoS digital health2, e0000198 (2023)

    work page 2023

Show all 47 references
  1. [14]

    Yang, Z.et al.Unveiling gpt-4v’s hidden challenges behind high accuracy on usmle questions: Observational study.J. Med. Internet Res.27, e65146 (2025)

    2025

  2. [15]

    In Findings of the Association for Computational Linguistics: EMNLP 2024, 12609–12629 (2024)

    Yao, Z.et al.Readme: Bridging medical jargon and lay understanding for patient education through data-centric nlp. In Findings of the Association for Computational Linguistics: EMNLP 2024, 12609–12629 (2024). 15/19

    2024

  3. [16]

    & Cohen, T

    Guo, Y ., Qiu, W., Wang, Y . & Cohen, T. Automated lay language summarization of biomedical scientific reviews. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, 160–168 (2021)

    2021

  4. [17]

    & Ananiadou, S

    Luo, Z., Xie, Q. & Ananiadou, S. Readability controllable biomedical document summarization.arXiv preprint arXiv:2210.04705(2022)

    2022 arXiv

  5. [18]

    Yao, Z. & Yu, H. Improving formality style transfer with context-aware rule injection. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), ...

    2021

  6. [19]

    & Wang, L

    Guo, Y ., August, T., Leroy, G., Cohen, T. & Wang, L. L. Appls: A meta-evaluation testbed for plain language summarization. arXiv preprint arXiv:2305.14341(2023)

    2023 arXiv

  7. [20]

    Liang, X.et al.Controllable text generation for large language models: A survey.arXiv preprint arXiv:2408.12599(2024)

    2024 arXiv

  8. [21]

    & Albarrak, A

    V o, A., Tao, Y ., Li, Y . & Albarrak, A. The association between social determinants of health and population health outcomes: ecological analysis.JMIR Public Heal. Surveillance9, e44070 (2023)

    2023

  9. [22]

    Shi, R.et al.From general to specific: Tailoring large language models for personalized healthcare.arXiv preprint arXiv:2412.15957(2024)

    2024 arXiv

  10. [23]

    Medicine7, 329 (2024)

    Stanceski, K.et al.The quality and safety of using generative ai to produce patient-centred discharge instructions.npj Digit. Medicine7, 329 (2024)

    2024

  11. [24]

    K., Glick, A

    Morrison, A. K., Glick, A. & Yin, H. S. Health literacy: implications for child health.Pediatr. review40, 263–277 (2019)

    2019

  12. [25]

    & Paudyal, V

    Suen, K., Shrestha, S., Osman, S. & Paudyal, V . Association between patient race/ethnicity, health literacy, socio-economic status, and incidence of medication errors: A systematic review.J. Racial Ethn. Heal. Disparities1–12 (2025)

    2025

  13. [26]

    (ed.)AHRQ Health Literacy Universal Precautions Toolkit(Agency for Healthcare Research and Quality, Rockville, MD, 2024), 3rd edn

    Brach, C. (ed.)AHRQ Health Literacy Universal Precautions Toolkit(Agency for Healthcare Research and Quality, Rockville, MD, 2024), 3rd edn. AHRQ Publication No. 15-0023-EF

    2024

  14. [27]

    Schooley, B.et al.Integrated digital patient education at the bedside for patients with chronic conditions: observational study.JMIR mHealth uHealth8, e22947 (2020)

    2020

  15. [28]

    & Unger, J

    Stephen, A., Connelly, D., Hung, L. & Unger, J. Staff-family communication methods in long-term care homes: An integrative review.J. Long Term Care220 (2024)

    2024

  16. [29]

    neural information processing systems33, 9459–9474 (2020)

    Lewis, P.et al.Retrieval-augmented generation for knowledge-intensive nlp tasks.Adv. neural information processing systems33, 9459–9474 (2020)

    2020

  17. [30]

    31.Yuan, W.et al.Following length constraints in instructions.arXiv preprint arXiv:2406.17744(2024)

    Chung, P.et al.Verifact: Verifying facts in llm-generated clinical text with electronic health records.arXiv preprint arXiv:2501.16672(2025). 31.Yuan, W.et al.Following length constraints in instructions.arXiv preprint arXiv:2406.17744(2024)

    2025 arXiv

  18. [32]

    & Liu, Q

    Jie, R., Meng, X., Shang, L., Jiang, X. & Liu, Q. Prompt-based length controlled generation with reinforcement learning. arXiv preprint arXiv:2308.12030(2023). 33.Zhang, S.et al.Instruction tuning for large language models: A survey.arXiv preprint arXiv:2308.10792(2023)

    2023 arXiv

  19. [34]

    35.Gu, J.et al.A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594(2024)

    Qin, Y .et al.Infobench: Evaluating instruction following ability in large language models.arXiv preprint arXiv:2401.03601 (2024). 35.Gu, J.et al.A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594(2024)

    2024 arXiv

  20. [36]

    arXiv preprint arXiv:2410.20266(2024)

    Szymanski, A.et al.Limitations of the llm-as-a-judge approach for evaluating llm outputs in expert knowledge tasks. arXiv preprint arXiv:2410.20266(2024). 37.Li, D.et al.Preference leakage: A contamination problem in llm-as-a-judge.arXiv preprint arXiv:2502.01534(2025)

    2024 arXiv

  21. [38]

    & Margetis, K

    Aydin, S., Karabacak, M., Vlachos, V . & Margetis, K. Large language models in patient education: a scoping review of applications in medicine.Front. Medicine11, 1477898 (2024)

    2024

  22. [39]

    & Atun, R

    Panch, T., Pearson-Stuttard, J., Greaves, F. & Atun, R. Artificial intelligence: opportunities and risks for public health.The Lancet Digit. Heal.1, e13–e14 (2019)

    2019

  23. [40]

    M., Schmerder, M

    Dennstädt, F., Hastings, J., Putora, P. M., Schmerder, M. & Cihoric, N. Implementing large language models in healthcare while balancing control, collaboration, costs and security.npj Digit. Medicine8, 143 (2025)

    2025

  24. [41]

    Shaping the future of ai in healthcare through ethics and governance.Humanit

    Bouderhem, R. Shaping the future of ai in healthcare through ethics and governance.Humanit. social sciences communica- tions11, 1–12 (2024). 16/19

    2024

  25. [42]

    for Comput

    Cai, P.et al.Paniniqa: Enhancing patient education through interactive question answering.Transactions Assoc. for Comput. Linguist.11, 1518–1536 (2023). 43.Wei, J.et al.Finetuned language models are zero-shot learners.arXiv preprint arXiv:2109.01652(2021)

    2023 arXiv

  26. [44]

    Wang, Y .et al.Self-instruct: Aligning language models with self-generated instructions.arXiv preprint arXiv:2212.10560 (2022)

    2022 arXiv

  27. [45]

    Tran, H., Yang, Z., Yao, Z. & Yu, H. Bioinstruct: instruction tuning of large language models for biomedical natural language processing.J. Am. Med. Informatics Assoc.31, 1821–1832 (2024)

    2024

  28. [46]

    Alva-Manchego, F.et al.Asset: A dataset for tuning and evaluation of sentence simplification models with multiple rewriting transformations.arXiv preprint arXiv:2005.00481(2020)

    2020 arXiv

  29. [47]

    & Gurevych, I

    Zhu, Z., Bernhard, D. & Gurevych, I. A monolingual tree-based translation model for sentence simplification. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), 1353–1361 (2010)

    2010

  30. [48]

    InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 11733–11751 (2022)

    Kwon, S.et al.Medjex: A medical jargon extraction model with wiki’s hyperlink span and contextualized masked language model score. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 11733–11751 (2022)

    2022

  31. [49]

    InAMIA Annual Symposium Proceedings, vol

    Moramarco, F.et al.Towards more patient friendly clinical notes through language models and ontologies. InAMIA Annual Symposium Proceedings, vol. 2021, 881 (2022)

    2022

  32. [50]

    Zhang, Y ., Baldridge, J. & He, L. Paws: Paraphrase adversaries from word scrambling.arXiv preprint arXiv:1904.01130 (2019)

    2019 arXiv

  33. [51]

    & Brockett, C

    Dolan, B. & Brockett, C. Automatically constructing a corpus of sentential paraphrases. InThird international workshop on paraphrasing (IWP2005)(2005)

    2005

  34. [52]

    R., Angeli, G., Potts, C

    Bowman, S. R., Angeli, G., Potts, C. & Manning, C. D. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326(2015)

    2015 arXiv

  35. [53]

    & Bowman, S

    Williams, A., Nangia, N. & Bowman, S. R. A broad-coverage challenge corpus for sentence understanding through inference.arXiv preprint arXiv:1704.05426(2017)

    2017 arXiv

  36. [54]

    & Shivade, C

    Romanov, A. & Shivade, C. Lessons from natural language inference in the clinical domain.arXiv preprint arXiv:1808.06752(2018)

    2018 arXiv

  37. [55]

    Rouge: A package for automatic evaluation of summaries

    Lin, C.-Y . Rouge: A package for automatic evaluation of summaries. InText summarization branches out, 74–81 (2004)

    2004

  38. [56]

    & Zhu, W.-J

    Papineni, K., Roukos, S., Ward, T. & Zhu, W.-J. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, 311–318 (2002)

    2002

  39. [57]

    x-ray" or even

    Xu, W., Napoles, C., Pavlick, E., Chen, Q. & Callison-Burch, C. Optimizing statistical machine translation for text simplification.Transactions Assoc. for Comput. Linguist.4, 401–415 (2016). 17/19 Appendix Table 7.Qualitative Analysis for lower readability (e.g., Grade 2 & 5),...

    2016

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.