Pith. sign in

REVIEW 3 major objections 5 minor 34 references

A novel language model for predicting serious adverse event results in clinical trials from their prospective registrations

T0 review · 3 major / 5 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 Pre-trial registration text encodes enough signal to predict which trial arm will have more serious adverse events, with the best model reaching 77.6% AUC.

desk verdict First shot at predicting SAEs from trial registrations shows a plausible signal, but the 'prospective' claim is not actually established by the methods. read the letter →

arxiv 2507.22919 v2 pith:XRIXUMBB submitted 2025-07-21 cs.CL cs.AI

Qixuan Hu , Xumou Zhang , Jinman Kim , Florence Bourgeois , Adam G. Dunn This is my paper
classification cs.CLcs.AI
keywords languagemodelsclinicaltrialsseriousadverseeventsTrials.govtrialregistrationtransferlearningslidingwindow
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that serious adverse event (SAE) results can be predicted from a clinical trial's prospective registration text alone, before any results exist. Using 22,107 two-arm parallel trials from ClinicalTrials.gov, the authors train a classifier that decides whether the experimental arm will have a higher proportion of participants with SAEs than the control arm, and a regressor that estimates the SAE proportion in the control arm. The best model, a frozen ClinicalT5 encoder combined with a transformer and multilayer perceptron, reaches 77.6% AUC on the classification task and 18.6% RMSE on the regression task. The authors caution that the regression error is too large for immediate deployment, but they argue the results demonstrate that registration documents contain learnable safety signals that could support monitoring, discrepancy detection, and eventually synthetic controls.

What carries the argument

The mechanism that carries the argument is a two-stage transfer-learning pipeline with a sliding-window text representation. Each registration document is split into overlapping chunks sized to the pretrained model's maximum context length (for example, a 256-token stride for BioBERT's 512-token limit), each chunk is embedded by a frozen pretrained language model such as ClinicalT5 or BioBERT, and the chunk embeddings are combined into a single document embedding that a downstream predictor (KNN, MLP, or Transformer-plus-MLP) uses for classification or regression. The sliding window matters because it gives complete coverage of long, section-structured registrations without retraining the language model, and the paper shows it improves average AUC by about 2.0 percentage points and reduces RMSE by about 1.6 percentage points across the twelve configurations tested.

What would settle it

Randomly sample 200–300 of the 22,107 trials, have human reviewers label the experimental and control arms, and rerun the classification and regression with corrected labels; if the label error rate is substantial or the AUC and RMSE change materially, the reported performance is an artifact of the regex labeling rather than a true registration signal. A second check would compare a text-only version of the best model against the structured-features-only baseline to see whether the registration text itself, rather than the structured fields, carries the predictive signal.

Watch

Extended reading notes

Core claim

On the authors' own terms, the central claim is that the serious adverse event outcomes of two-arm parallel clinical trials are partly predictable from information available at registration time, and that a transfer-learning pipeline is the way to extract that signal. The strongest quantitative result is that ClinicalT5 with sliding-window embedding and a Transformer-Encoder-plus-MLP downstream model predicts which arm will have the higher proportion of SAEs with 77.58% AUC, exceeding the structured-information random forest baseline of 69.27% AUC. The same configuration predicts the control-arm SAE proportion with 18.56% RMSE. The paper further claims that the sliding-window method, which chunks long registration documents into overlapping pieces within the language model's context limit, consistently beats the common single-pass embedding baseline across all 12 language-model/downstream-model combinations.

Load-bearing premise

The result rests on the assumption that the regular-expression procedure that labels each arm as experimental or control is correct for essentially all of the 22,107 trials, even though only five trials were manually checked.

Editorial extensions

If this is right

  • If the central claim holds, expected SAE results can be computed at registration time, letting sponsors and regulators compare predicted against reported outcomes and flag trials with unusual discrepancies.
  • The predicted control-arm SAE proportion could inform trial design, such as sample-size planning and the level of safety monitoring assigned to a protocol before enrollment.
  • The approach establishes a baseline for safety-outcome prediction from registrations, against which future context-aware models can be measured.
  • The sliding-window embedding method itself may transfer to other long, section-structured documents where local sections are self-contained, not just clinical trial registrations.
  • If control-arm estimates improve further, constructing synthetic control arms from similar completed trials becomes a more realistic alternative to placebo arms.

Reading between the lines

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

  • Because the paper does not ablate the text signal from the structured features it feeds to the baseline, an editor's testable guess is that part of the 77.6% AUC comes from easy structured clues such as phase, cancer indication, and industry funding; a text-only model would isolate the incremental contribution of the registration narrative.
  • The same frozen-encoder plus sliding-window pipeline could be pointed at other registration fields, such as primary outcome measures or recruitment targets, to predict efficacy results rather than safety outcomes.
  • A practical extension would turn the regression's prediction intervals into a monitoring flag: trials whose reported SAE proportions fall far outside the predicted range could be prioritized for audit, which is a direct but unstated use of the RMSE figure.
  • The five-trial manual validation of arm labels is thin relative to the 22,107-trial corpus, so re-estimating performance on a human-annotated random sample of arms would show how much of the reported accuracy depends on the regex labeling step.
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 / 5 minor

Summary. The paper proposes a transfer-learning pipeline for predicting serious adverse event (SAE) outcomes from ClinicalTrials.gov registration text. It frames two tasks: classifying whether the experimental arm will have a higher SAE proportion than the control arm, and regressing the SAE proportion in the control arm. Using four pretrained language models, three downstream architectures, and a sliding-window embedding strategy, the best configuration (ClinicalT5 + Transformer + MLP) achieves 77.6% AUC and 18.6% RMSE. The authors report consistent improvements from the sliding-window approach over a no-window baseline and compare against a structured-data random forest for classification.

Significance. If the central claim holds, this is the first demonstration that trial registration text contains learnable safety signals, and it would support the proposed uses in trial monitoring and synthetic-control construction. The paper has concrete strengths: a large cohort of 22,107 trials, systematic comparison across 12 model/architecture combinations, bootstrap-based Wilcoxon tests for the window comparisons, and a structured-data baseline for the classification task. The sliding-window method is a practical contribution for long registration documents. However, two load-bearing issues—temporal provenance of the registration text and inadequate validation of the arm-labeling procedure—currently prevent the prospective claim from being established, and the evaluation metrics are computed on rebalanced test sets without a regression baseline.

major comments (3)
  1. [§2.1, Abstract] The abstract claims prediction from 'registrations prior to the trial,' but Section 2.1 states only that 'Clinical trial registrations were accessed via the ClinicalTrials.gov API' and does not specify that the text was restricted to the record version available before trial enrollment. ClinicalTrials.gov records are routinely amended after results are posted, and the API returns the current version unless a versioned archive is used. Because the target labels are derived from the results section, post-trial amendments to intervention descriptions, eligibility wording, or outcome definitions could leak outcome information into the model inputs. The 77.6% AUC and 18.6% RMSE may therefore be partly retrospective, which would invalidate the proposed prospective monitoring and synthetic-control use cases. The authors should either retrieve versioned records (e.g., from the ClinicalTrials.gov archive), demonstrate that the accessed text predates trial completion for the included trials, or otherwise quantify and bound the potential leakage.
  2. [§6.1, §2.1] The regular-expression procedure that labels arms as experimental or control is validated manually on only five trials (NCT01263132, NCT01386632, NCT00059332, NCT01904032, NCT00004732) and then applied to all 22,107 trials. These labels define the classification target (which arm has a higher SAE proportion) and also appear in the rendered text used as model input. If arm labels are wrong for a non-negligible fraction of trials, both the targets and the features are corrupted, so the reported AUC and RMSE would not reflect true predictive performance. Given that the paper itself acknowledges the possibility of misclassification, the authors should provide a larger validation sample with confidence intervals, or a sensitivity analysis showing that the main results are robust to plausible arm-label error rates.
  3. [§2.1, §3] The evaluation metrics are reported on rebalanced test sets: classification uses downsampling to 11,542 trials (5,771 per class) and regression uses binned sampling up to 1,000 trials per bin to yield 5,192 trials. If the test set is rebalanced, the reported AUC, F1, and RMSE are not estimates for the original population of two-arm parallel trials, and the headline numbers may overstate real-world performance on an imbalanced distribution. The manuscript should report metrics on an untouched held-out test set (or provide reweighted/calibrated estimates) and should state explicitly whether the train/test split occurs before or after rebalancing. In addition, no structured-data baseline is reported for the regression task, so the Discussion's claim that text models outperform structured-only baselines is currently supported only for classification.
minor comments (5)
  1. [§2.1] The text says arm identification used regular expressions on 'the trial title, description of the interventions, and trial arms,' while Section 6.1 says 'based on trial title and description'; this inconsistency should be reconciled.
  2. [Tables 3 and 4] Several table entries have formatting errors, e.g., '72.0274.06' and '75.7777.58' in Table 3 and '22.5318.56' in Table 4; these should be corrected with proper spacing.
  3. [Figure 2] The caption repeats 'the sliding window approach a window size 2 and stride 1' and 'window size 2 and stride 1'; the duplication should be removed.
  4. [Throughout] The model name is inconsistently written as 'BGEm3' in tables and 'BGE-m3' in Table 1; one spelling should be used consistently.
  5. [§3] No confidence intervals are reported for the AUC or RMSE point estimates in Tables 3 and 4; given the bootstrap-based testing, bootstrap CIs would be a natural and valuable addition.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: predictions are evaluated on held-out trials; main caveats are data-provenance and labeling-noise risks, not circularity.

full rationale

The central claim (77.6% AUC for arm-level SAE comparison and 18.6% RMSE for control-arm SAE proportion) arises from a standard supervised pipeline: frozen pretrained embeddings of registration text are fed into downstream classifiers/regressors trained on labels derived from ClinicalTrials.gov summary results, with held-out test evaluation and bootstrap resampling. No equation defines the input in terms of the target, and no fitted parameter is later relabeled as a prediction. The paper contains no load-bearing self-citations: references to prior prediction work (e.g., refs. 10-13) are external studies, and no uniqueness theorem is invoked to force the modeling choice. The acknowledged limitations, such as regex-based arm identification and the unverified temporal provenance of API-retrieved registration text, are validity and leakage risks rather than circularity: to call them circular one would need to establish that the input text included the results-based target by construction, which the manuscript does not state. Against external held-out data, the derivation chain is self-contained and not equivalent to its inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the analysis relies on standard ML components and several domain assumptions about data quality and representativeness.

free parameters (6)
  • KNN k = 60 (classification), 20 (regression)
    Selected via grid search on a validation set; changes which neighbors vote.
  • Sliding window overlap = 50% stride
    Chosen by hand to balance efficiency and context; not optimized, affects all embeddings.
  • Regression binned sampling cap = 1,000 trials per bin, 10 bins
    Reshapes the outcome distribution and therefore the reported RMSE.
  • Classification downsampling size = 5,771 trials per class
    Balances the classes before evaluation, changing the test-set AUC context.
  • Downstream architecture size = 12 MLP layers; 12 Transformer layers, 8 heads
    Tuned by grid search; model capacity affects performance.
  • Training epochs = 20 (classification), 40 (regression)
    Chosen by elbow point on validation; influences convergence and overfitting.
assumptions (4)
  • domain assumption Registration text, formatted as the Researcher View, contains features predictive of future SAE outcomes.
    The whole study depends on this signal existing; the paper only tests it indirectly through model performance.
  • domain assumption The regular-expression arm identification is accurate enough at scale.
    Manually validated on five trials only (Section 6.1), yet applied to all 22,107 trials.
  • domain assumption Trials with summary results on ClinicalTrials.gov represent the population where the model will be used.
    Only 21% of interventional studies have reported results (Section 2.1); the paper does not analyze selection bias.
  • domain assumption Frozen pretrained language model embeddings transfer to trial registration text.
    Model weights are frozen (Section 2.1.1); the downstream model must work with the representations these models produce.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A novel language model for predicting serious adverse event results in clinical trials from their prospective registrations." pith.science (2026). https://pith.science/paper/XRIXUMBB

@misc{pith2026250722919,
  author       = {Pith},
  title        = {Pith review of: A novel language model for predicting serious adverse event results in clinical trials from their prospective registrations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XRIXUMBB}},
  note         = {Machine review of arXiv:2507.22919}
}
read the original abstract

Objectives: With accurate estimates of expected safety results, clinical trials could be better designed and monitored. We evaluated methods for predicting serious adverse event (SAE) results in clinical trials using information only from their registrations prior to the trial. Material and Methods: We analyzed 22,107 two-arm parallel interventional clinical trials from ClinicalTrials.gov with structured summary results. Two prediction models were developed: a classifier predicting whether a greater proportion of participants in an experimental arm would have SAEs (area under the receiver operating characteristic curve; AUC) compared to the control arm, and a regression model to predict the proportion of participants with SAEs in the control arms (root mean squared error; RMSE). A transfer learning approach using pretrained language models (e.g., ClinicalT5, BioBERT) was used for feature extraction, combined with a downstream model for prediction. To maintain semantic representation in long trial texts exceeding localized language model input limits, a sliding window method was developed for embedding extraction. Results: The best model (ClinicalT5+Transformer+MLP) had 77.6% AUC when predicting which trial arm had a higher proportion of SAEs. When predicting SAE proportion in the control arm, the same model achieved RMSE of 18.6%. The sliding window approach consistently outperformed direct comparisons. Across 12 classifiers, the average absolute AUC increase was 2.00%, and absolute RMSE reduction was 1.58% across 12 regressors. Discussion: Summary results data from ClinicalTrials.gov remains underutilized. Predicted results of publicly reported trials provides an opportunity to identify discrepancies between expected and reported safety results.

Figures

Figures reproduced from arXiv: 2507.22919 by the authors.

Figure 1
Figure 1. Distribution of 61,330 clinical trials with summary results, grouped by trial design and arm size, indicating a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. In a comparison example of the baseline approach and the sliding window approach, the sliding window [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison (Accuracy, F1, AUC) of different model design using baseline and sliding window [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance comparison (MAE, RMSE) of different model design using baseline and sliding window [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages

  1. [1]

    World Health Organi- zation, Geneva, 2018

    World Health Organization.International Standards for Clinical Trial Registries: The Registration of All Interventional Trials is a Scientific, Ethical and Moral Responsibility. World Health Organi- zation, Geneva, 2018. ISBN 9789241514743. URL https://www.who.int/publications/i/item/ international-standards-for-clinical-trial-registers

    work page 2018

  2. [2]

    Food and Drug Administration

    U.S. Food and Drug Administration. What is a serious adverse event?, May 2023. URL https://www.fda.gov/ safety/reporting-serious-problems-fda/what-serious-adverse-event. Accessed: 2025-04-24

    work page 2023

  3. [3]

    Clinical trials, n.d

    International Committee of Medical Journal Editors. Clinical trials, n.d. URL https://www.icmje.org/ recommendations/browse/publishing-and-editorial-issues/clinical-trial-registration. html. Accessed: 2025-04-24

    work page 2025

  4. [4]

    basic results

    Tony Tse, Rebecca J Williams, and Deborah A Zarin. Reporting “basic results” in clinicaltrials. gov.Chest, 136 (1):295–303, 2009

    work page 2009

  5. [5]

    Reporting summary results in clinical trial registries: updated guidance from who.The Lancet Global Health, 13(4):e759–e768, 2025

    An-Wen Chan, Ghassan Karam, Justin Pymento, Lisa M Askie, Luiza R da Silva, Ségolène Aymé, Christo- pher Marc Taylor, Lotty Hooft, Anna Laura Ross, and Vasee Moorthy. Reporting summary results in clinical trial registries: updated guidance from who.The Lancet Global Health, 13(4):e759–e768, 2025

    work page 2025

  6. [6]

    Clinical trial registration was associated with lower risk of bias compared with non-registered trials among trials included in systematic reviews

    Kristina Lindsley, Nicole Fusco, Tianjing Li, Rob Scholten, and Lotty Hooft. Clinical trial registration was associated with lower risk of bias compared with non-registered trials among trials included in systematic reviews. Journal of Clinical Epidemiology, 145:164–173, 2022

    work page 2022

  7. [7]

    Better access to information about clinical trials.Annals of Internal Medicine, 133(8):609–614, 2000

    Alexa T McCray. Better access to information about clinical trials.Annals of Internal Medicine, 133(8):609–614, 2000

    work page 2000

  8. [8]

    The clinicaltrials

    Deborah A Zarin, Tony Tse, Rebecca J Williams, Robert M Califf, and Nicholas C Ide. The clinicaltrials. gov results database—update and key issues.New England Journal of Medicine, 364(9):852–860, 2011

    work page 2011

Show all 34 references
  1. [9]

    Department of Health and Human Services

    U.S. Department of Health and Human Services. Clinical trials registration and results information submission. Fed- eral Register, 2017. URL https://www.federalregister.gov/documents/2016/09/21/2016-22129/ clinical-trials-registration-and-results-information-submission . 42 CF...

    2017

  2. [10]

    Predictive modeling of clinical trial terminations using feature engineering and embedding learning.Scientific reports, 11(1):3446, 2021

    Magdalyn E Elkin and Xingquan Zhu. Predictive modeling of clinical trial terminations using feature engineering and embedding learning.Scientific reports, 11(1):3446, 2021

    2021

  3. [11]

    Predicting publication of clinical trials using structured and unstructured data: model development and validation study.Journal of Medical Internet Research, 24(12):e38859, 2022

    Siyang Wang, Simon Šuster, Timothy Baldwin, and Karin Verspoor. Predicting publication of clinical trials using structured and unstructured data: model development and validation study.Journal of Medical Internet Research, 24(12):e38859, 2022

    2022

  4. [12]

    Key indicators of phase transition for clinical trials through machine learning.Drug discovery today, 25(2):414–421, 2020

    Felipe Feijoo, Michele Palopoli, Jen Bernstein, Sauleh Siddiqui, and Tenley E Albright. Key indicators of phase transition for clinical trials through machine learning.Drug discovery today, 25(2):414–421, 2020

    2020

  5. [13]

    Predicting phase 1 lymphoma clinical trial durations using machine learning: An in-depth analysis and broad application insights.Clinics and Practice, 14(1):69–88, 2023

    Bowen Long, Shao-Wen Lai, Jiawen Wu, and Srikar Bellur. Predicting phase 1 lymphoma clinical trial durations using machine learning: An in-depth analysis and broad application insights.Clinics and Practice, 14(1):69–88, 2023

    2023

  6. [14]

    Synthetic and external controls in clinical trials–a primer for researchers.Clinical epidemiology, pages 457–467, 2020

    Kristian Thorlund, Louis Dron, Jay JH Park, and Edward J Mills. Synthetic and external controls in clinical trials–a primer for researchers.Clinical epidemiology, pages 457–467, 2020

    2020

  7. [15]

    Table meets llm: Can large language models understand structured table data? a benchmark and empirical study

    Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. Table meets llm: Can large language models understand structured table data? a benchmark and empirical study. InProceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 645–654, 2024. 11

    2024

  8. [16]

    Turl: Table understanding through representation learning.ACM SIGMOD Record, 51(1):33–40, 2022

    Xiang Deng, Huan Sun, Alyssa Lees, You Wu, and Cong Yu. Turl: Table understanding through representation learning.ACM SIGMOD Record, 51(1):33–40, 2022

    2022

  9. [17]

    Dimitris Spathis and Fahim Kawsar. The first step is the hardest: Pitfalls of representing and tokenizing temporal data for large language models.Journal of the American Medical Informatics Association, 31(9):2151–2158, 2024

    2024

  10. [18]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining.Bioinformatics, 36 (4):1234–1240, 2020

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining.Bioinformatics, 36 (4):1234–1240, 2020

    2020

  11. [19]

    Publicly available clinical bert embeddings.arXiv preprint arXiv:1904.03323, 2019

    Emily Alsentzer, John R Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings.arXiv preprint arXiv:1904.03323, 2019

    1904 arXiv

  12. [20]

    Clinicalt5: A generative language model for clinical text

    Qiuhao Lu, Dejing Dou, and Thien Nguyen. Clinicalt5: A generative language model for clinical text. InFindings of the Association for Computational Linguistics: EMNLP 2022, pages 5436–5443, 2022

    2022

  13. [21]

    M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024

    Multi-Linguality Multi-Functionality Multi-Granularity. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024

    2024

  14. [22]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts.Transactions of the Association for Computational Linguistics, 12:157–173, 02 2024. ISSN 2307-387X. doi:10.116...

    2024 doi

  15. [23]

    Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150, 2020

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150, 2020

    2004 arXiv

  16. [24]

    Xlnet: Generalized autoregressive pretraining for language understanding.Advances in neural information processing systems, 32, 2019

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding.Advances in neural information processing systems, 32, 2019

    2019

  17. [25]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  18. [26]

    Found in the middle: How language models use long contexts better via plug-and-play positional encoding.Advances in Neural Information Processing Systems, 37:60755–60775, 2024

    Zhenyu Zhang, Runjin Chen, Shiwei Liu, Zhewei Yao, Olatunji Ruwase, Beidi Chen, Xiaoxia Wu, Zhangyang Wang, et al. Found in the middle: How language models use long contexts better via plug-and-play positional encoding.Advances in Neural Information Processing Systems, 37:6075...

    2024

  19. [27]

    Hier- archical context merging: Better long context understanding for pre-trained llms.arXiv preprint arXiv:2404.10308, 2024

    Woomin Song, Seunghyuk Oh, Sangwoo Mo, Jaehyung Kim, Sukmin Yun, Jung-Woo Ha, and Jinwoo Shin. Hier- archical context merging: Better long context understanding for pre-trained llms.arXiv preprint arXiv:2404.10308, 2024

    2024 arXiv

  20. [28]

    Transfer learning with clinical concept embeddings from large language models.arXiv preprint arXiv:2409.13893, 2024

    Yuhe Gao, Runxue Bao, Yuelyu Ji, Yiming Sun, Chenxi Song, Jeffrey P Ferraro, and Ye Ye. Transfer learning with clinical concept embeddings from large language models.arXiv preprint arXiv:2409.13893, 2024

    2024 arXiv

  21. [29]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  22. [30]

    National Academies Press Washington, DC, 2007

    Philip Aspden and Philip Aspden.Preventing medication errors, volume 8. National Academies Press Washington, DC, 2007

    2007

  23. [31]

    Postmarketing adverse drug reactions: A duty to report?Neurology: Clinical Practice, 3(4):288–294, 2013

    Eran Klein and Dennis Bourdette. Postmarketing adverse drug reactions: A duty to report?Neurology: Clinical Practice, 3(4):288–294, 2013

    2013

  24. [32]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  25. [33]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. InProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016

    2016

  26. [34]

    Large language models are zero-shot time series forecasters.Advances in Neural Information Processing Systems, 36:19622–19635, 2023

    Nate Gruver, Marc Finzi, Shikai Qiu, and Andrew G Wilson. Large language models are zero-shot time series forecasters.Advances in Neural Information Processing Systems, 36:19622–19635, 2023. 12

    2023

Pith tools

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