Pith. sign in

REVIEW 4 major objections 5 minor 47 references

A Hybrid Framework for Subject Analysis: Integrating Embedding-Based Regression Models with Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a hybrid framework—a regression model setting the subject-heading count, constrained LLM generation, and embedding-based post-editing to the LCSH vocabulary—raises recall from 43% to 63% and precision from 8% to 26%…

desk verdict A credible hybrid framework for LCSH prediction, but the unvalidated nearest-neighbor post-processing makes the headline recall gains hard to trust. read the letter →

arxiv 2507.22913 v1 pith:KCBB774Q submitted 2025-07-19 cs.CL cs.AI

classification cs.CLcs.AI
keywords subjectanalysisLCSHlibrarycataloginglargelanguagemodelschain-of-thoughtpromptingfine-tuningmulti-labelclassificationcontrolledvocabulary
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 sets out to show that large language models alone are not trustworthy enough for assigning Library of Congress Subject Headings (LCSH), but pairing them with small embedding-based machine learning models turns them into a controllable cataloging assistant. The proposed hybrid framework first predicts how many subject headings a book should receive, then prompts the LLM to generate that many candidate terms, and finally replaces any term not in the official LCSH vocabulary with the semantically nearest valid heading. On the paper's test set this raises recall from 43% for an unconstrained zero-shot LLM to 63% for chain-of-thought generation with post-processing, and precision from 8% to 26% when the predicted count is used as a hard limit. These results are offered as evidence that both major LLM failure modes in subject analysis, over-generation and hallucination, can be corrected outside the model itself.

What carries the argument

The load-bearing mechanism is a three-phase pipeline. First, an embedding-based count controller encodes the title and abstract with a sentence-embedding model and uses linear regression to predict the number of subject headings. Second, count-constrained generation asks the LLM to produce exactly that many terms, optionally through multi-round chain-of-thought prompts that build up coverage while avoiding repeats. Third, a vocabulary post-editor takes every generated term absent from LCSH and replaces it with its nearest neighbor among about 318,500 valid headings, using PCA-reduced embedding similarity and a fast nearest-neighbor search. The count controller makes output size controllable, and the post-editor makes vocabulary membership guaranteed, so recall and precision can be traded off by relaxing the count constraint.

What would settle it

Take a random sample of test records, run the full framework, and give professional catalogers the list of post-processed substitutions—each out-of-vocabulary generated term paired with its nearest valid LCSH replacement—and ask whether the replacement is appropriate for the book. If catalogers reject most substitutions as irrelevant or wrong, the reported 9–12% recall gains from post-processing would be artifacts of exact-match scoring rather than real improvements.

Watch

Extended reading notes

Core claim

The paper's central claim is that LLMs' two failure modes in subject analysis, uncertainty about how many headings to produce and inventing terms outside the controlled vocabulary, can be corrected externally without retraining the architecture. A lightweight regression model trained on title and abstract embeddings predicts the number N of LCSH terms; the LLM is then prompted to generate N (or 2N or 3N) candidate headings. A post-processing module encodes roughly 318,500 valid LCSH terms, reduces the embedding dimension with PCA, and uses nearest-neighbor search to snap every generated term that is not in the vocabulary to its closest valid heading. On one open 8-billion-parameter model, recall on catalog records rises from 43% to 63% when chain-of-thought prompting is combined with this post-editing, while precision rises from 8% to 26% when the predicted count is used as a limit. Gains are measured as exact-match agreement with headings originally assigned by human catalogers.

Load-bearing premise

The framework assumes that the nearest valid LCSH term to a generated out-of-vocabulary term, found by embedding similarity, is the subject heading a cataloger would have chosen; the paper reports no human check of the substitutions.

Editorial extensions

If this is right

  • Constraining generation to the predicted count cuts the average number of output terms from 14.89 to 3.14 (about 79%) and raises F1 from 0.135 to 0.244.
  • Chain-of-thought prompting combined with post-processing delivers the largest recall gain on the studied model, from 0.43 to 0.63.
  • The post-editor's benefit is uneven: fine-tuned models gain only about 5% recall and no precision, while prompt-only zero-shot and CoT methods gain 9% and 12%, suggesting fine-tuning already teaches vocabulary alignment.
  • Relaxing the count constraint from N to 2N to 3N gives a tunable recall–precision trade-off, useful for choosing between broad coverage and clean catalog records.
  • Because post-processing is applied after generation, it can improve vocabulary alignment for any LLM without modifying or retraining it.

Reading between the lines

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

  • If human evaluation later validates the nearest-neighbor substitutions, the same generate-then-snap pattern should transfer to other controlled vocabularies, such as medical subject headings or corporate taxonomies, where exact-match scoring is standard.
  • The count-prediction module is a generic control knob: any generative system that over- or under-produces structured outputs could be steered by a small regressor on embeddings, not just subject headings.
  • A risk the paper leaves open is that embedding-similarity replacement can turn a hallucinated term into a plausible but still wrong LCSH heading; if that happens often, the reported post-processing gains overstate true cataloging quality.
  • The paper's baseline results show stronger closed models already outperform the locally run model, so the framework's gains plausibly compound on stronger generators, but that direction was not tested here.
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 / 5 minor

Summary. The paper proposes a hybrid framework for automatic subject analysis using Library of Congress Subject Headings (LCSH). It combines a regression model that predicts the number of subject headings from title/abstract embeddings with LLMs (Llama-3.1-8B, GPT-3.5/4, DeepSeek) that generate candidate headings, followed by a FAISS-based post-processing step that maps non-LCSH outputs to the nearest LCSH term. Experiments on about 78,000 UNT catalog records report that the framework raises Llama-3.1-8B recall from 43% (zero-shot) to 63% (CoT with post-processing) and precision from 8% to 26% (limit-n with post-processing). The main contribution is a low-cost integration of ML count prediction, LLM generation, and vocabulary-constrained post-processing, with an explicit trade-off between recall and precision.

Significance. If the reported gains are genuine, the framework offers a practical and scalable approach for semi-automatic cataloging, with a clear trade-off between recall and precision. The paper is strong in its use of a large real-world dataset, systematic comparisons of zero-shot, CoT, and fine-tuning (SFT vs LoRA), and explicit acknowledgment of limitations. The central weakness is that the post-processing gains rest on unvalidated nearest-neighbor substitutions, and the automatic exact-match metrics may not reflect cataloging subjectivity. The framework itself is sensible and the paper is honest about its limitations, but the magnitude and meaning of the reported gains require additional validation.

major comments (4)
  1. [4.6, Table 7] The post-processing step (Section 4.6) replaces every generated term not in the LCSH vocabulary with its nearest LCSH heading, and the 'after' metrics in Table 7 are computed by exact match against the cataloger-assigned gold headings. No human review or substitution-accuracy analysis is reported, so the +9 to +12 percentage-point recall gains attributed to post-processing may be inflated by nearest-neighbor replacements that coincide with gold headings without being semantically appropriate. This is load-bearing for the headline claim that recall rises from 43% to 63%. Please report the fraction of terms replaced, a human evaluation of a sample of substitutions (e.g., whether the replacement is a genuine paraphrase of the original term or equals the gold heading), and at a minimum an analysis that separates genuine paraphrases from coincidental embedding-space matches.
  2. [6] The paper's own limitation statement (Section 6) acknowledges that automatic exact-match metrics 'may be skewed' because subject assignment is subjective. The test set is labeled by catalogers, but no inter-cataloger agreement is reported, so it is unclear what ceiling a system can reach and how much of the reported gains are semantic improvements versus artifacts of a single labeler. Please add human evaluation or agreement statistics, and report the expected upper bound on recall given labeler variability.
  3. [5.2.1, Table 4] The count predictor is the proposed 'guide' for the hybrid framework, but the best model (linear regression on All-mpnet-base-v2 embeddings) has a Pearson correlation of only 0.37 and an average absolute difference of about 1.09 labels. The paper does not isolate whether the constraint's benefit comes from the predicted N or simply from imposing any output limit. Please include an ablation comparing N against a fixed constant (e.g., the average gold count or a randomly drawn count in the same range) to demonstrate that the regression model adds value beyond length control.
  4. [5, Tables 2, 3, 7] No confidence intervals or significance tests are reported for the differences across methods. Several comparisons in Tables 2, 3, and 7 are close (e.g., recall 0.51 vs 0.52, precision 0.04 vs 0.06), and the test set has only 2,100 samples. Please report bootstrap or paired tests for the central comparisons, or otherwise quantify the uncertainty of the reported gains.
minor comments (5)
  1. [Abstract] The phrase 'providing initial predictions to guide LLM generations' is imprecise; the ML model predicts a label count, not initial subject predictions. Clarify to 'predicting the number of labels.'
  2. [4.6] The text states the embedding dimension is 628, but all-mpnet-base-v2 outputs 768-dimensional embeddings; please verify the dimension and the PCA reduction.
  3. [4.6] 'NSS was implemented' should read 'NNS was implemented.'
  4. [4.3 and 5.3] The relationship between CoT prompt versions V1-V3 and the CoT row in Table 7 is unclear; please specify which version is used in the main CoT results.
  5. [Table 4] The selection of linear regression over XGBoost is based on average difference, but XGBoost has lower RMSE and higher PCC; please justify the selection criterion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the count predictor and post-processing are fitted or constructed independently of the test labels, and reported gains are empirical rather than forced by construction.

full rationale

The derivation chain is self-contained. The label-count predictor (Section 4.5) is trained on a separate training split and used only to set the prompt constraint N; it does not by construction determine which LCSH terms are evaluated. The post-processor (Section 4.6) maps non-LCSH LLM outputs to their nearest LCSH vocabulary entries using a pretrained embedding model, PCA, and FAISS, with no reference to the gold test labels; whether that nearest neighbor coincides with a cataloger-assigned heading is an empirical question, not a definitional identity. Final recall, precision, and F1 are computed by exact match against the gold headings, so improvements such as 0.43 to 0.63 are not encoded in the framework itself. There are no load-bearing self-citations or imported uniqueness theorems. The paper's own limitation that automatic metrics 'may be skewed' because subject assignment is subjective is a validity concern, not circularity. The phrase 'optimal number of LCSH labels' overstates a weak regressor (PCC 0.37), but that is an overclaim, not a circular step.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a fitted count predictor, several hand-set hyperparameters, and three domain assumptions about the evaluation ground truth, the sufficiency of title-abstract features, and the reliability of embedding-based vocabulary mapping. No new entities are postulated beyond the framework components themselves.

free parameters (5)
  • Number-of-terms prediction model (linear regression on All-mpnet-base-v2 embedding) = N/A (model); average difference 1.093, PCC 0.370, RMSE 1.528
    Selected as the best count predictor from six combinations; its output N is used to constrain LLM generation. The word 'optimal' in the paper overstates this predictor's accuracy.
  • max-new-tokens = 150 (best for recall; 50 and 100 also tested)
    Controls output length and is tuned on the test set; the choice of 150 produces high recall but low precision for fine-tuned models, and lower values improve F1.
  • Constraint multiplier (N, 2N, 3N) = N, 2N, 3N all reported; N used for the headline precision result
    A hand-set design choice explored to trade off recall and precision; the multiplier is not derived from data.
  • CoT rounds and per-round label counts = 3 rounds with variants 2/3/5, n/n/2n, n/2n/AMAP; 4-round variant tested
    The prompt schedule and label budgets per round are chosen by hand; the 4-round variant did not improve recall.
  • PCA dimensionality = 50 (reduced from 628)
    Chosen to make FAISS search tractable; nearest-neighbor substitution quality depends on this reduction.
assumptions (3)
  • domain assumption LCSH terms assigned by UNT catalogers are treated as the ground truth for automatic evaluation.
    The paper's recall and precision are exact matches against these labels; Section 6 acknowledges subject assignment is subjective and automatic metrics may be skewed.
  • domain assumption Titles and abstracts contain enough topical signal to predict both label count and labels.
    The framework uses only MARC 245 and 520 fields as input; if these are insufficient, both the count regression and the LLM generation are limited.
  • domain assumption All-mpnet-base-v2 embeddings preserve semantic similarity between hallucinated phrases and correct LCSH terms after PCA reduction to 50 dimensions.
    Post-processing relies on FAISS nearest neighbor over these reduced embeddings to substitute valid LCSH terms; no manual validation of the substitutions is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hybrid Framework for Subject Analysis: Integrating Embedding-Based Regression Models with Large Language Models." pith.science (2026). https://pith.science/paper/KCBB774Q

@misc{pith2026250722913,
  author       = {Pith},
  title        = {Pith review of: A Hybrid Framework for Subject Analysis: Integrating Embedding-Based Regression Models with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCBB774Q}},
  note         = {Machine review of arXiv:2507.22913}
}
read the original abstract

Providing subject access to information resources is an essential function of any library management system. Large language models (LLMs) have been widely used in classification and summarization tasks, but their capability to perform subject analysis is underexplored. Multi-label classification with traditional machine learning (ML) models has been used for subject analysis but struggles with unseen cases. LLMs offer an alternative but often over-generate and hallucinate. Therefore, we propose a hybrid framework that integrates embedding-based ML models with LLMs. This approach uses ML models to (1) predict the optimal number of LCSH labels to guide LLM predictions and (2) post-edit the predicted terms with actual LCSH terms to mitigate hallucinations. We experimented with LLMs and the hybrid framework to predict the subject terms of books using the Library of Congress Subject Headings (LCSH). Experiment results show that providing initial predictions to guide LLM generations and imposing post-edits result in more controlled and vocabulary-aligned outputs.

Figures

Figures reproduced from arXiv: 2507.22913 by the authors.

Figure 1
Figure 1. Method Overview 3 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The Workflow of Post-processing 5 EXPERIMENT RESULTS AND ANALYSIS To evaluate LCSH term generation, we primarily use recall, precision, F1-score, and the average number of generations as key metrics. These metrics help assess the accuracy of predictions and the trade-off between label counts and output redundancy. 5.1 Evaluation of LLM-Only Methods in Subject Analysis 5.1.1 Prompt Engineering Methods We initially ex… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 29 canonical work pages

  1. [1]

    Cataloging and classification: an introduction

    Athena Salaba and Lois Mai Chan. Cataloging and classification: an introduction. Rowman & Littlefield, 2023

  2. [2]

    Machine learning in automated text categorization

    Fabrizio Sebastiani. Machine learning in automated text categorization. ACM computing surveys (CSUR) , 34(1):1–47, 2002

  3. [3]

    Automated knowledge organisation: Ai/ml-based subject indexing system for libraries

    Mustak Ahmed, Mondrita Mukhopadhyay, and Parthasarathi Mukhopadhyay. Automated knowledge organisation: Ai/ml-based subject indexing system for libraries. DESIDOC Journal of Library & Information Technology, 43(1), 2023

  4. [4]

    Automatically cataloging scholarly articles using library of congress subject headings

    Nazmul Kazi, Nathaniel Lane, and Indika Kahanda. Automatically cataloging scholarly articles using library of congress subject headings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 43–49, 2021

  5. [5]

    Hierarchical multi-label classification networks

    Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. Hierarchical multi-label classification networks. In International conference on machine learning, pages 5075–5084. PMLR, 2018

  6. [6]

    A review of methods for imbalanced multi-label classification

    Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification. Pattern Recognition, 118:107965, 2021

  7. [7]

    An experiment with the use of chatgpt for lcsh subject assignment on electronic theses and dissertations

    Eric HC Chow, TJ Kao, and Xiaoli Li. An experiment with the use of chatgpt for lcsh subject assignment on electronic theses and dissertations. Cataloging & Classification Quarterly, 62(5):574–588, 2024

  8. [8]

    Few-shot prompting for subject indexing of german medical book titles

    Lisa Kluge and Maximilian Kähler. Few-shot prompting for subject indexing of german medical book titles. In Proceedings of the 20th Conference on Natural Language Processing (KONVENS 2024), pages 141–148, 2024

Show all 47 references
  1. [9]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

  2. [10]

    Fine-tuning large language models for scientific text classification: A comparative study

    Zhyar Rzgar K Rostam and Gábor Kertész. Fine-tuning large language models for scientific text classification: A comparative study. In 2024 IEEE 6th International Symposium on Logistics and Industrial Informatics (LINDI), pages 000233–000238. IEEE, 2024

  3. [11]

    A performance model of the length and number of subject headings and index phrases

    Robert Losee. A performance model of the length and number of subject headings and index phrases. Knowledge organization, 31(4):245–251, 2004

  4. [12]

    Artificial intelligence (ai) library services innovative conceptual framework for the digital transformation of university education

    Rifqah Olufunmilayo Okunlaya, Norris Syed Abdullah, and Rose Alinda Alias. Artificial intelligence (ai) library services innovative conceptual framework for the digital transformation of university education. Library Hi Tech, 40(6):1869–1892, 2022

  5. [13]

    Defining artificial intelligence for librarians

    Andrew M Cox and Suvodeep Mazumdar. Defining artificial intelligence for librarians. Journal of librarianship and information science, 56(2):330–340, 2024

  6. [14]

    A survey of knowledge-enhanced text generation

    Wenhao Yu, Chenguang Zhu, Zaitang Li, Zhiting Hu, Qingyun Wang, Heng Ji, and Meng Jiang. A survey of knowledge-enhanced text generation. ACM Computing Surveys, 54(11s):1–38, 2022

  7. [15]

    Machine learning and ontology-based novel semantic document indexing for information retrieval

    Anil Sharma and Suresh Kumar. Machine learning and ontology-based novel semantic document indexing for information retrieval. Computers & Industrial Engineering, 176:108940, 2023

  8. [16]

    Hdltex: Hierarchical deep learning for text classification

    Kamran Kowsari, Donald E Brown, Mojtaba Heidarysafa, Kiana Jafari Meimandi, Matthew S Gerber, and Laura E Barnes. Hdltex: Hierarchical deep learning for text classification. In 2017 16th IEEE international conference on machine learning and applications (ICMLA), pages 364–371....

  9. [17]

    Large scale subject category classification of scholarly papers with deep attentive neural networks

    Bharath Kandimalla, Shaurya Rohatgi, Jian Wu, and C Lee Giles. Large scale subject category classification of scholarly papers with deep attentive neural networks. Frontiers in research metrics and analytics, 5:600382, 2021

  10. [18]

    A survey of controllable text generation using transformer-based pre-trained language models

    Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. A survey of controllable text generation using transformer-based pre-trained language models. ACM Computing Surveys, 56(3):1–37, 2023. 11

  11. [19]

    An integrated clustering and bert framework for improved topic modeling

    Lijimol George and P Sumathy. An integrated clustering and bert framework for improved topic modeling. International Journal of Information Technology, 15(4):2187–2195, 2023

  12. [20]

    A survey of multi-label classification based on supervised and semi-supervised learning

    Meng Han, Hongxin Wu, Zhiqiang Chen, Muhang Li, and Xilong Zhang. A survey of multi-label classification based on supervised and semi-supervised learning. International Journal of Machine Learning and Cybernetics, 14(3):697–724, 2023

  13. [21]

    Advancing automatic subject indexing: Combining weak supervision with extreme multi-label classification

    Lakshmi Rajendram Bashyam and Ralf Krestel. Advancing automatic subject indexing: Combining weak supervision with extreme multi-label classification. In International Workshop on Natural Scientific Language Processing and Research Knowledge Graphs, pages 214–223. Springer, 2024

  14. [22]

    Artificial intelligence chatbots in academic libraries: the rise of chatgpt

    Adebowale Jeremy Adetayo. Artificial intelligence chatbots in academic libraries: the rise of chatgpt. Library Hi Tech News, 40(3):18–21, 2023

  15. [23]

    Data-efficient fine-tuning for llm-based recommendation

    Xinyu Lin, Wenjie Wang, Yongqi Li, Shuo Yang, Fuli Feng, Yinwei Wei, and Tat-Seng Chua. Data-efficient fine-tuning for llm-based recommendation. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pages 365–374, 2024

  16. [24]

    Comparative study of domain driven terms extraction using large language models

    Sandeep Chataut, Tuyen Do, Bichar Dip Shrestha Gurung, Shiva Aryal, Anup Khanal, Carol Lushbough, and Etienne Gnimpieba. Comparative study of domain driven terms extraction using large language models. arXiv preprint arXiv:2404.02330, 2024

  17. [25]

    Llm-take: Theme-aware keyword extraction using large language models

    Reza Yousefi Maragheh, Chenhao Fang, Charan Chand Irugu, Parth Parikh, Jason Cho, Jianpeng Xu, Saranyan Sukumar, Malay Patel, Evren Korpeoglu, Sushant Kumar, et al. Llm-take: Theme-aware keyword extraction using large language models. In 2023 IEEE International Conference on B...

  18. [26]

    Utilising a large language model to annotate subject metadata: A case study in an australian national research data catalogue

    Shiwei Zhang, Mingfang Wu, and Xiuzhen Zhang. Utilising a large language model to annotate subject metadata: A case study in an australian national research data catalogue. arXiv preprint arXiv:2310.11318, 2023

  19. [27]

    Structured information extraction from complex scientific text with fine-tuned large language models

    Alexander Dunn, John Dagdelen, Nicholas Walker, Sanghoon Lee, Andrew S Rosen, Gerbrand Ceder, Kristin Persson, and Anubhav Jain. Structured information extraction from complex scientific text with fine-tuned large language models. arXiv preprint arXiv:2212.05238, 2022

  20. [28]

    Gpt4roi: Instruction tuning large language model on region-of-interest

    Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. In European conference on computer vision, pages 52–70. Springer, 2024

  21. [29]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:2773...

  22. [30]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

  23. [31]

    Discover: The unt library catalog, n.d

    University of North Texas Libraries. Discover: The unt library catalog, n.d

  24. [32]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  25. [33]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

  26. [34]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

  27. [35]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

  28. [36]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

  29. [37]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 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 langu...

  30. [38]

    Scibert: A pretrained language model for scientific text

    Iz Beltagy, Kyle Lo, and Arman Cohan. Scibert: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676, 2019. 12

  31. [39]

    Transformer in transformer

    Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in neural information processing systems, 34:15908–15919, 2021

  32. [40]

    Transformer models: an introduction and catalog

    Xavier Amatriain, Ananth Sankar, Jie Bing, Praveen Kumar Bodigutla, Timothy J Hazen, and Michaeel Kazi. Transformer models: an introduction and catalog. arXiv preprint arXiv:2302.07730, 2023

  33. [41]

    Ammus: A survey of transformer- based pretrained models in natural language processing

    Katikapalli Subramanyam Kalyan, Ajit Rajasekharan, and Sivanesan Sangeetha. Ammus: A survey of transformer- based pretrained models in natural language processing. arXiv preprint arXiv:2108.05542, 2021

  34. [42]

    Random forests

    Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001

  35. [43]

    Linear regression

    Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani, and Jonathan Taylor. Linear regression. In An introduction to statistical learning: With applications in python, pages 69–134. Springer, 2023

  36. [44]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016

  37. [45]

    Analysis of a complex of statistical variables into principal components

    Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933

  38. [46]

    Approximate nearest neighbor search in high dimensions

    Alexandr Andoni, Piotr Indyk, and Ilya Razenshteyn. Approximate nearest neighbor search in high dimensions. In Proceedings of the International Congress of Mathematicians: Rio de Janeiro 2018, pages 3287–3318. World Scientific, 2018

  39. [47]

    The faiss library

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The faiss library. arXiv preprint arXiv:2401.08281, 2024. 13

Pith tools

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