Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Reconstructing Biological Pathways by Applying Selective Incremental Learning to (Very) Small Language Models

T0 review · 3 major / 6 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 The paper claims that a 110M-parameter BERT model fine-tuned incrementally on only its most confident mistakes reaches 93% validation accuracy on regulatory-relationship prediction, using 117 of 517 curated statements.

desk verdict A promising observation about overconfident errors in active learning for biomedical relation classification, but the headline 93% vs 75.7% result is undersupported without a random-selection baseline and matched training-set-size control. read the letter →

arxiv 2507.04432 v1 pith:WGLUAUTW submitted 2025-07-06 q-bio.MN cs.CLcs.ITcs.LGcs.PFmath.IT

Pranta Saha , Joyce Reimer , Brook Byrns , Connor Burbridge , Neeraj Dhar , Jeffrey Chen , Steven Rayan , Gordon Broderick This is my paper
classification q-bio.MNcs.CLcs.ITcs.LGcs.PFmath.IT MSC 92C4268T07
keywords smalllanguagemodelsBERTactivelearninginformationentropyregulatoryinteractionspathwayreconstructiontuberculosissampleselection
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 argues that for narrow biomedical tasks, very small language models can outperform far larger ones, provided the few fine-tuning examples are chosen by an informative policy. It attempts to reconstruct host-pathogen regulatory relationships around tuberculosis from a curated reference set of 517 statements, using active learning to select new training examples. The central claim is that the most valuable examples are the model's incorrect predictions made with highest certainty (lowest entropy), not its uncertain but correct guesses. On this basis a 110M-parameter BERT reaches 93% balanced accuracy with less than 25% of the reference data, whereas a policy that also consumes high-entropy correct examples reaches only 75.7%. The broader thesis is that selective incremental learning can make small, domain-specific models a rational alternative to general-purpose LLMs in accuracy-critical biomedical settings.

What carries the argument

The central object is the Shannon entropy over the softmax output distribution of the BERT classifier, averaged over ten replicate queries, used as a confidence estimate for ranking validation predictions. Low entropy means the model is certain; the policy selects the lowest-entropy predictions that are wrong, forcing correction of overconfident errors. The active learning loop starts with a baseline of 20 curated statements, adds four newly selected examples per iteration (two per class), retrains for three epochs with the AdamW optimizer, and repeats for 25 iterations with full memory retention.

What would settle it

A direct test would run the same 25-iteration active learning protocol on a second set of curated regulatory statements from a different pathway system, comparing low-entropy-incorrect selection against random sampling and pure highest-entropy uncertainty sampling; if the low-entropy policy does not beat the random baseline in that setting, the claim that overconfident errors are the most informative examples would be refuted.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a very small BERT model, guided by active learning, can reconstruct intracellular regulatory interactions with high accuracy while using only a fraction of the available curated data. The authors define confidence as the Shannon entropy of the softmax output distribution, averaged over ten replicate queries, and compare two selection policies: one that adds both low-entropy incorrect and high-entropy correct predictions, and one that adds only low-entropy incorrect predictions. The latter policy achieves 93.0% overall accuracy versus 75.7% for the former, with the difference concentrated in the negative regulation class (96.7% versus 68.7%). The authors interpret this as evidence that overconfident mistakes are more informative than uncertain correct guesses, and they show that the low-entropy-incorrect policy preferentially recruits poorly connected target species that are regulated by hub mediators such as TNF, a pattern consistent with false confidence inherited from frequently occurring source terms.

Load-bearing premise

The central result depends on the assumption that the model's softmax entropy, averaged over ten replicate queries, reliably identifies which incorrect examples are hardest for the model and therefore most informative for retraining; if this entropy estimate is noisy or poorly calibrated, the observed accuracy gains may not generalize.

Editorial extensions

If this is right

  • A 110M-parameter BERT model, tuned this way, achieves over 80% accuracy using roughly 21 to 23 percent of the reference set, supporting the use of very small domain-specific models for pathway gap-filling.
  • The low-entropy-incorrect selection policy significantly outperforms the joint policy that also includes high-entropy correct examples (93.0% versus 75.7%, p=0.03), implying that overconfident errors carry more learning signal than uncertain correct answers.
  • The zero-shot comparison shows that Llama 3.2 1B, despite being about ten times larger, delivers roughly random accuracy (50.4%) with high confidence, whereas BERT achieves similar accuracy (54.4%) with more balanced class assignments and higher uncertainty.
  • The preferred policy recruits more target species that are connected to network hub mediators, which suggests that low-entropy errors often involve rare targets of frequent sources and may indicate where the model's confidence is misplaced.
  • The pattern implies that label-aware sample selection can be steered toward infrequent target terms regulated by well-connected mediators, potentially guiding the design of new experiments that would maximally enrich the reference set.

Reading between the lines

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

  • The entropy estimate may be confounded by source-term frequency: a frequently seen source like TNF can produce low-entropy but wrong predictions for rare targets; a direct test would separate source frequency from target rarity to identify the true driver of informativeness.
  • The paper compares against a policy that includes high-entropy correct examples, but does not report comparisons to standard active learning baselines such as random sampling or pure highest-entropy uncertainty sampling; without those, the size of the advantage over conventional strategies remains open.
  • Because the accuracy gain is concentrated in the negative class (96.7% versus 68.7%), the policy may be particularly effective for imbalanced classes; testing on balanced relation sets would clarify whether the benefit generalizes.
  • If the result transfers to other relation-extraction domains, the same low-entropy-error selection could reduce annotation costs for tasks like drug-target or gene-disease curation, where labeled training data is expensive to obtain.
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 / 6 minor

Summary. The manuscript reports an active-learning study in which a BERT-base (110M) model is incrementally fine-tuned on small curated sets of regulatory relationships (up/downregulation) from a tuberculosis pathway reference set of 517 statements. The authors compare two selection policies: (Policy A) adding per class the correct example with highest entropy and the incorrect example with lowest entropy, and (Policy B) adding per class the two incorrect examples with lowest entropy. Over 25 iterations and three trials, Policy B yields 93.0% average balanced validation accuracy using ~84 training samples, versus 75.7% with ~93 samples for Policy A (Table 3). The authors interpret this as evidence that selectively adding low-entropy (overconfident) incorrect predictions is the key driver of data-efficient learning. The paper also includes a zero-shot comparison of BERT-base and Llama 3.2 1B showing both near chance.

Significance. If the causal attribution were supported, the paper would provide a valuable empirical demonstration that very small LMs can be tuned efficiently with active selection on a biomedical relation-classification task, with potential benefits for resource-limited settings. The study is commendably transparent in reporting per-trial results, class-level accuracies, and detailed supplementary tables on the selected examples and network statistics. However, the absence of a random-selection baseline and the small number of replicates currently leave the central efficiency claim unsupported; the observed advantage of Policy B over Policy A could be due to differences in the number of selected samples or to the inclusion of correct examples, rather than to the entropy criterion per se. The authors themselves appropriately note the preliminary nature and the need for broader testing.

major comments (3)
  1. [Section 2.5, Table 3] The manuscript states in Section 2.5 that models were fine-tuned on both unique and random datasets under identical conditions, yet no random-selection results appear anywhere in the paper. Without a random-selection baseline run under the same 25-iteration, 4-new-examples-per-iteration protocol, the headline claim that a 110M BERT model reaches >80% accuracy using <25% of the reference set cannot be causally attributed to selective incremental learning; cumulative fine-tuning on ~110 random examples with the same validation set might produce similar results.
  2. [Table 3] The comparison between the two policies is not a clean ablation of the entropy criterion. Policy A adds one high-entropy correct and one low-entropy incorrect example per class (two examples per class), while Policy B adds two low-entropy incorrect examples per class (also two per class). Thus the policies differ simultaneously in the target selection (correct vs incorrect) and in the number of qualifying errors; the observed accuracy difference (93.0% vs 75.7%) cannot be uniquely attributed to the low-entropy incorrect criterion.
  3. [Table 3] With only three trials and a 20-example balanced validation set (10 per class), the overall p-value of 0.03 is fragile. The class-level comparisons are not significant (positive class p=0.424, negative class p=0.112), and the average total training samples differ between policies (93 vs 84, p=0.326), so the comparison is not matched for training-set size. Additional replicates with matched sample sizes, or an analysis that controls for sample count, are required to support the claimed difference.
minor comments (6)
  1. [Section 2.5 vs Table 2] The text states that fine-tuning was performed for 3 epochs, while Table 2 lists the best epoch value as 4; please reconcile.
  2. [Section 3.2 vs Table 3] The first experiment is described as 'two replicate numerical experiments' but Table 3 reports three trials for each policy; please clarify which trials are intended.
  3. [Abstract/Introduction vs Results] The abstract and Introduction mention 3B, 1B, and 100M models, but only Llama 3.2 1B and BERT-base 110M results are actually reported; please either include the 3B results or adjust the text.
  4. [Table 1] Llama 3.2 is labeled '1B parameters' in the text but the table reports 1.3 billion parameters; please make the parameter counts consistent.
  5. [Section 2.4] Equations (2) and (3) are displayed with garbled notation; please ensure the Softmax and Shannon-entropy formulas are rendered correctly and define all variables.
  6. [Table S3] The table title mentions 'Llama 3B' while the header says 'Llama 3.2 1B'; please correct the title to match the reported model.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: held-out evaluation is independent of the oracle-based selection procedure.

full rationale

This paper reports an empirical active-learning study. The central accuracy claim (93% vs 75.7% in Table 3) is evaluated on a validation set that the paper explicitly describes as a "consistent held-out validation set" (Section 2.5). Training examples are selected from an unlabeled pool using true labels in an oracle setting, which is standard practice in active learning and does not make the reported accuracy circular; the labels are used for selection, not for constructing the validation scores. The entropy-based selection criterion is not defined in terms of the reported accuracy, and no equation in the paper reduces the claimed result to its inputs by construction. The paper does not rely on load-bearing self-citations or imported uniqueness theorems. The main weaknesses are experimental-design gaps: the absence of a random-selection baseline, the unmatched number of training samples between policies, and the small number of trials and validation samples. These are correctness or generalizability concerns, not circularity. The paper's own limitation statement ("these observations are derived from a very limited and specific use case") is a caveat rather than evidence of a circular step. Accordingly, no specific circular step can be exhibited, and the score reflects only minor ambiguities in reporting, not circularity.

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

The paper does not introduce any new theoretical entities. It depends on the standard assumption that curated pathway databases provide reliable labels, and on the validity of softmax entropy as an uncertainty measure for active learning.

free parameters (4)
  • Number of initial baseline examples per class = 10
    Chosen by hand; the active learning trajectory starts from this balanced seed set.
  • Number of selected examples per class per iteration = 2 (or 1+1 in the joint policy)
    Chosen by hand; determines the growth of the training set.
  • Number of fine-tuning epochs = 4
    Selected by Ray Tune; a training hyperparameter, not a scientific parameter, but affects the learning curve.
  • Learning rate = 3e-5
    Selected by Ray Tune; training hyperparameter.
assumptions (3)
  • domain assumption Pathway Commons and BEL Selventa provide accurate ground-truth regulatory relationships.
    The entire evaluation treats these curated statements as the correct labels for the protein pairs.
  • domain assumption The query 'Does protein X regulate protein Y positively or negatively?' captures the relation to be predicted.
    The model's predictions are evaluated against the curated label for the same pair.
  • domain assumption Softmax output entropy reflects the model's uncertainty in a way that is meaningful for sample selection.
    The active learning policy ranks examples by entropy; this assumption is load-bearing but not validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reconstructing Biological Pathways by Applying Selective Incremental Learning to (Very) Small Language Models." pith.science (2026). https://pith.science/paper/WGLUAUTW

@misc{pith2026250704432,
  author       = {Pith},
  title        = {Pith review of: Reconstructing Biological Pathways by Applying Selective Incremental Learning to (Very) Small Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGLUAUTW}},
  note         = {Machine review of arXiv:2507.04432}
}
read the original abstract

The use of generative artificial intelligence (AI) models is becoming ubiquitous in many fields. Though progress continues to be made, general purpose large language AI models (LLM) show a tendency to deliver creative answers, often called "hallucinations", which have slowed their application in the medical and biomedical fields where accuracy is paramount. We propose that the design and use of much smaller, domain and even task-specific LM may be a more rational and appropriate use of this technology in biomedical research. In this work we apply a very small LM by today's standards to the specialized task of predicting regulatory interactions between molecular components to fill gaps in our current understanding of intracellular pathways. Toward this we attempt to correctly posit known pathway-informed interactions recovered from manually curated pathway databases by selecting and using only the most informative examples as part of an active learning scheme. With this example we show that a small (~110 million parameters) LM based on a Bidirectional Encoder Representations from Transformers (BERT) architecture can propose molecular interactions relevant to tuberculosis persistence and transmission with over 80% accuracy using less than 25% of the ~520 regulatory relationships in question. Using information entropy as a metric for the iterative selection of new tuning examples, we also find that increased accuracy is driven by favoring the use of the incorrectly assigned statements with the highest certainty (lowest entropy). In contrast, the concurrent use of correct but least certain examples contributed little and may have even been detrimental to the learning rate.

Figures

Figures reproduced from arXiv: 2507.04432 by the authors.

Figure 2
Figure 2. Sequential tuning and data selection. Repeated sequential cycles of active learning with 100% memory using AdamW optimization algorithm using new examples selected as highly confident (low information entropy) but incorrect assignments. 2.4 Entropy-based sample selection To assess the level of certainty with which each validation querie is assigned to weak or strong affinity class, we first use the Softmax activatio… view at source ↗
Figure 3
Figure 3. David and Goliath. A zero-shot inference using 80% of the overall reference set for tuning and 20% for predictive validation sampled in 5 repeat queries consisting of 23 questions for each class [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. A joint reward and penalty strategy. Predictive accuracy achieved with incremental active learning using both underconfident and overconfident examples [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Contributing policy elements. Selection of overconfident examples is distinguishing feature in better performing active learning trial [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: A reward focused approach. Predictive accuracy using a penalty-based approach selecting only high confidence but incorrect learning examples [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 17 canonical work pages

  1. [1]

    based on the same initial baseline tuning set with the only difference being the random selection of subsequent validation sets. Initial results suggest that the results can be quite variable with repeat experiments achieving an overall accuracy of 69% to 83% (Figure 4; Table 3), using roughly 21% (109 and 110 of 517 relationship statements) of the overal...

    work page 2025

  2. [3]

    ACM BCB’25, October 2025, Philadelphia, Pennsylvania, USA P

    Summary of active learning results. ACM BCB’25, October 2025, Philadelphia, Pennsylvania, USA P. Saha et al. molecules currently missing from our understanding of biological pathways. Using a set of regulatory relationships extracted from manually curated databases as a reference truth, we show that even when tuned on 80% of the available data, the 100M p...

    work page 2025

  3. [4]

    Predictive accuracy achieved with incremental active learning using both underconfident and overconfident examples

    A joint reward and penalty strategy. Predictive accuracy achieved with incremental active learning using both underconfident and overconfident examples. ACM BCB’25, October 2025, Philadelphia, Pennsylvania, USA P. Saha et al. only 23% (117 out of 517 relationship statements) of the overall data available over the same 25 tuning cycles (Figure 6; Table 3)....

    work page 2025

  4. [5]

    Selection of overconfident examples is distinguishing feature in better performing active learning trial

    Contributing policy elements. Selection of overconfident examples is distinguishing feature in better performing active learning trial. ACM BCB’25, October 2025, Philadelphia, Pennsylvania, USA P. Saha et al. certainty, but correct predictions may be less informative than incorrect predictions made with high certainty. Focusing exclusively on the latter a...

    work page 2025

  5. [6]

    Zorik Gekhman, Gal Yona, Roee Aharoni, Matan Eyal, Amir Feder, Roi Reichart, Jonathan Herzig. (2024). Does fine-tuning LLMs on new knowledge encourage hallucinations?. arXiv preprint arXiv:2405.05904 (May 2024). DOI: https://doi.org/10.48550/arXiv.2405.05904

    arXiv doi:10.48550/arxiv.2405.05904 2024

  6. [7]

    In 2024 47th MIPRO ICT and Electronics Convention (MIPRO)

    Hallucinations in llms: Understanding and addressing challenges. In 2024 47th MIPRO ICT and Electronics Convention (MIPRO). IEEE, 2084-2088. DOI: https://doi.org/10.1109/MIPRO60963.2024.10569238

    arXiv 2024

  7. [8]

    arXiv preprint arXiv:2410.02707 (Oct 2024)

    Llms know more than they show: On the intrinsic representation of llm hallucinations. arXiv preprint arXiv:2410.02707 (Oct 2024). DOI: https://doi.org/10.48550/arXiv.2410.02707

    arXiv doi:10.48550/arxiv.2410.02707 2024

  8. [12]

    Active Learning for Robust and Representative LLM Generation in Safety-Critical Scenarios

    Active Learning for Robust and Representative LLM Generation in Safety-Critical Scenarios. arXiv preprint arXiv:2410.11114 (Oct. 2024). DOI: https://doi.org/10.48550/arXiv.2410.11114

    work page Pith review arXiv doi:10.48550/arxiv.2410.11114 2024

Show all 38 references
  1. [13]

    arXiv preprint arXiv:2501.12948 (Jan 2025)

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (Jan 2025). DOI: https://doi.org/10.48550/arXiv.2501.12948

    2025 doi arXiv

  2. [15]

    Neurocomputing 503, 7 (Sep 2022), 92-108

    Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing 503, 7 (Sep 2022), 92-108. DOI: https://doi.org/10.1016/j.neucom.2022.06.111

    2022 doi

  3. [20]

    Annu Rev Physiol 10, 85 (Feb 2023), 71-91

    Infectious and inflammatory pathways to cough. Annu Rev Physiol 10, 85 (Feb 2023), 71-91. DOI: https://doi.org/10.1146/annurev-physiol-031422-092315

    2023 doi

  4. [21]

    J Inform Electr Electron Eng 4, 3 (Nov 2023), 1-3

    Optical character recognition development using python. J Inform Electr Electron Eng 4, 3 (Nov 2023), 1-3. DOI: http://dx.doi.org/10.54060/jieee.2023.75

    2023 doi

  5. [22]

    In Proceedings of the 18th BioNLP Workshop and Shared Task

    ScispaCy: fast and robust models for biomedical natural language processing. In Proceedings of the 18th BioNLP Workshop and Shared Task. Association for ACM BCB’25, October 2025, Philadelphia, Pennsylvania, USA P. Saha et al. Computational Linguistics, 319-327. DOI: https://do...

    2025 doi

  6. [24]

    Nucleic Acids Res, 43, D1 (Jan 2015), D204-12

    UniProt: a hub for protein information. Nucleic Acids Res, 43, D1 (Jan 2015), D204-12. DOI: https://doi.org/10.1093/nar/gku989

    2015 doi

  7. [25]

    Nucleic Acids Res, 48, D1 (Jan 2020), D489-97

    Pathway Commons 2019 Update: integration, analysis and exploration of pathway data. Nucleic Acids Res, 48, D1 (Jan 2020), D489-97. DOI: https://doi.org/10.1093/nar/gkz946

    2019 doi

  8. [26]

    Database, 2019, (June 2019), baz068

    Re-curation and rational enrichment of knowledge graphs in Biological Expression Language. Database, 2019, (June 2019), baz068. DOI: https://doi.org/10.1093/database/baz068

    2019 doi

  9. [27]

    Mol Syst Biol, 19, 5 (May 2023), e11325

    Automated assembly of molecular mechanisms at scale from text mining and curated databases. Mol Syst Biol, 19, 5 (May 2023), e11325. DOI: https://doi.org/10.15252/msb.202211325

    2023 doi

  10. [28]

    arXiv preprint arXiv:2405.00208 (Apr 2024)

    A primer on the inner workings of transformer-based language models. arXiv preprint arXiv:2405.00208 (Apr 2024). DOI: https://doi.org/10.48550/arXiv.2405.00208

    2024 doi arXiv

  11. [29]

    arXiv preprint arXiv:1711.05101 (Nov 2017)

    Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (Nov 2017). DOI: https://doi.org/10.48550/arXiv.1711.05101

    2017 doi arXiv

  12. [31]

    arXiv preprint arXiv:2305.14264 (May 2023)

    Active learning principles for in-context learning with large language models. arXiv preprint arXiv:2305.14264 (May 2023). DOI: https://doi.org/10.48550/arXiv.2305.14264

    2023 doi arXiv

  13. [32]

    arXiv preprint arXiv:2410.15326 (Oct 2024)

    A survey of uncertainty estimation in llms: Theory meets practice. arXiv preprint arXiv:2410.15326 (Oct 2024). DOI: https://doi.org/10.48550/arXiv.2410.15326

    2024 doi arXiv

  14. [33]

    arXiv preprint arXiv:2408.09172 (Aug 2024)

    Unlocking the power of llm uncertainty for active in-context example selection. arXiv preprint arXiv:2408.09172 (Aug 2024). DOI: https://doi.org/10.48550/arXiv.2408.09172

    2024 doi arXiv

  15. [34]

    DOI: https://doi.org/10.48550/arXiv.2412.07255

    arXiv preprint arXiv:2412.07255 (Dec 2024). DOI: https://doi.org/10.48550/arXiv.2412.07255

    2024 doi arXiv

  16. [36]

    arXiv preprint arXiv:2312.12148 (Dec 2023)

    Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 (Dec 2023). DOI: https://doi.org/10.48550/arXiv.2312.12148

    2023 doi arXiv

  17. [37]

    Appl Sci 15, 6 (Mar 2025),

    Analyzing LLAMA3 performance on classification task using LoRA and QLoRA techniques. Appl Sci 15, 6 (Mar 2025),

    2025

  18. [2006]

    Nucleic Acids Res, 34, suppl_1 (Jan 2006), D319-21

    The HUGO gene nomenclature database, 2006 updates. Nucleic Acids Res, 34, suppl_1 (Jan 2006), D319-21. DOI: https://doi.org/10.1093/nar/gkj147

    2006 doi

  19. [2014]

    Semin Immunol 26, 6 (Dec 2024), 486-496

    Crosstalk between Mycobacterium tuberculosis and the host cell. Semin Immunol 26, 6 (Dec 2024), 486-496. DOI: https://doi.org/10.1016/j.smim.2014.09.002

    2024 doi

  20. [2015]

    In 2015 Proceedings of the IEEE international conference on computer vision (ICCV)

    Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In 2015 Proceedings of the IEEE international conference on computer vision (ICCV). IEEE, 19-27. DOI: https://doi.ieeecomputersociety.org/10.1109/ICCV.2015.11

    2015 doi

  21. [2017]

    PloS one 12, 12 (Dec 2017), e0190046

    Wikipedia as a gateway to biomedical research: The relative distribution and use of citations in the English Wikipedia. PloS one 12, 12 (Dec 2017), e0190046. DOI: https://doi.org/10.1371/journal.pone.0190046

    2017 doi

  22. [2018]

    arXiv preprint arXiv:1807.05118 (Jul 2018)

    Tune: A research platform for distributed model selection and training. arXiv preprint arXiv:1807.05118 (Jul 2018). DOI: https://doi.org/10.48550/arXiv.1807.05118

    2018 doi arXiv

  23. [2019]

    In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

    Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). Association fo...

    2019

  24. [2020]

    In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics

    Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 2699-2712. DOI: https://doi.org/10.18653/v1/2020.acl-main.240

    2020 doi

  25. [2021]

    JAIR 70 (Apr 2021), 1373-411

    Confident learning: Estimating uncertainty in dataset labels. JAIR 70 (Apr 2021), 1373-411. DOI: https://doi.org/10.1613/jair.1.12125

    2021 doi

  26. [2022]

    arXiv preprint arXiv:2211.04325 (Oct 2022)

    Will we run out of data? Limits of LLM scaling based on human-generated data. arXiv preprint arXiv:2211.04325 (Oct 2022). DOI: https://doi.org/10.48550/arXiv.2211.04325

    2022 doi arXiv

  27. [2023]

    In Architecture and System Support for Transformer Models (ASSYST@ ISCA 2023)

    Scaling infrastructure to support multi-trillion parameter LLM training. In Architecture and System Support for Transformer Models (ASSYST@ ISCA 2023)

    2023

  28. [2024]

    In 2024 IEEE 21st International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET)

    Large Language Models (LLM) in Industry: A Survey of Applications, Challenges, and Trends. In 2024 IEEE 21st International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET). IEEE, 229-234. DOI: https://doi.org/10.1109/HONET63146.2024...

    2024

  29. [2025]

    Curr Rev Musculoskelet Med 18, 7 (Jul 2025), 1-21

    A Current Review of Generative AI in Medicine: Core Concepts, Applications, and Current Limitations. Curr Rev Musculoskelet Med 18, 7 (Jul 2025), 1-21. DOI: https://doi.org/10.1007/s12178-025-09961-y

    2025 doi

  30. [3087]

    DOI: https://doi.org/10.3390/app15063087 Table S1. Reference truth statemenst extracted from Pathway Commons and BEL Large Corpus Source Target Relationship Action IL6 A2M Activation IL6 ABCB1 Activation TNF ABCB1 Activation IL6 ABCC1 Activation IL6 ABCG2 Activation IL6 ADAMTS...

Pith tools

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