Pith. sign in

REVIEW 4 major objections 5 minor 57 references

Cost-efficient generative AI summarization for scalable automated essay scoring in educational assessment

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A mid-tier GPT summarizer scores essays more reliably than the flagship model, while the best essays suffer the most compression loss.

desk verdict An honest, narrowly scoped within-pipeline comparison of GPT summarizers for essay scoring; the mini-model result is plausible but the missing baselines and unstated embedding context limit keep it from being more than a conditional pass. read the letter →

arxiv 2607.15829 v1 pith:7H5JQMKI submitted 2026-07-17 cs.CL cs.LG

Haowei Hua This is my paper · ORCID
classification cs.CLcs.LG
keywords automatedessayscoringGPTsummarizationlong-formtextcompressionhybridfeaturerepresentationquadraticweightedkappasummaryfidelitymodelcosttrade-offconstructpreservation
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

Long student essays often exceed the token limits of standard embedding models, forcing automated scoring pipelines to truncate them and risk losing argumentative content. This paper tries to establish that generative summarization can serve as an information-preserving alternative: compress each essay to a 512-token summary, embed that summary, and combine it with 22 handcrafted features from the original text before feeding a gradient-boosted scorer. On a standard 1–6 essay benchmark, the mid-size GPT-5 mini variant achieved the best human agreement (QWK 0.8435), slightly beating the full-size GPT-5 (0.8350) and the lightweight nano (0.8332), despite much lower API cost. The authors also find that summary fidelity—measured by lexical overlap, semantic similarity, entity coverage, and keyphrase overlap—declines as essay scores rise, meaning the most complex writing is the hardest to compress without loss. If correct, the result lets a cost-conscious practitioner pick a smaller summarizer without sacrificing scoring quality, while flagging a systematic risk for high-achieving essays.

What carries the argument

The load-bearing piece is the hybrid representation formed by concatenating an embedding of the 512-token GPT summary with 22 handcrafted linguistic features—surface statistics, lexical richness, discourse connectives, readability indices, and error counts—extracted from the original unsummarized essay. The summary length is enforced by an adaptive iterative resummarization rule: if output exceeds 512 tokens, the prompt requests 80% of the previous target repeatedly until compliant. This mechanism lets one pipeline process essays of arbitrary length through fixed-length classifiers; a gradient-boosted decision tree turns the concatenated vector into a 1–6 score, evaluated with quadratic weig

What would settle it

Run the same pipeline with the first 512 tokens of each essay replacing the GPT summary, keeping embeddings, handcrafted features, classifiers, and data splits identical. If truncation attains QWK near or above 0.8435, the central claim—that summarization is a loss-reducing preprocessing step—would be disproved. A second check: if score-band fidelity metrics show entity and keyphrase loss in the highest bands alongside a systematic underestimate of scores in those bands, the fairness concern would be confirmed experimentally.

Watch

Extended reading notes

Core claim

The central claim is not that one LLM is a better scorer, but that the choice of GPT summarizer changes what a downstream scoring model sees, and therefore changes quadratic weighted kappa. Under an otherwise identical hybrid pipeline—adaptive GPT summarization capped at 512 tokens, embeddings of the summary from a 4-billion-parameter instruction-tuned embedding model, 22 handcrafted features from the original essay, and gradient-boosted classifiers—GPT-5 mini yields 0.8435 agreement, GPT-5 0.8350, and GPT-5 nano 0.8332. The authors interpret this as a model-selection result: a smaller optimized variant can beat its larger and smaller siblings on downstream scoring, so cost and accuracy are

Load-bearing premise

The whole case rests on the assumption that a 512-token GPT summary plus 22 handcrafted features preserves more of the construct-relevant writing signal than simply truncating a long essay—yet the paper reports no truncation baseline, and its own score-band plots show the summaries of high-scoring essays are the least faithful.

Editorial extensions

If this is right

  • Choosing a mid-size summarizer can deliver the same or better human agreement as the full model at a fraction of the per-token cost, so large-scale deployments need not default to the largest model.
  • Summarization quality is not a reliable proxy for scoring quality: GPT-5 scored best on ROUGE, semantic similarity, entity coverage, and keyphrase preservation, yet GPT-5 mini scored best on QWK.
  • Higher-scoring essays are systematically harder to compress; any deployment should measure summary fidelity by score band, not just on average, because information loss concentrates where the scoring stakes may be highest.
  • Because the pipeline holds prompts, embeddings, features, and classifiers fixed, the observed QWK differences are attributable to summarizer-variant behavior rather than to other pipeline changes.

Reading between the lines

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

  • The obvious next experiment implied by the paper is a truncation baseline: feed the first 512 tokens of each essay through the same embedding-plus-handcrafted pipeline. If truncation reaches QWK near 0.8435, the rationale for summarization collapses; if it falls short, the method gains a real existence proof.
  • The score-band fidelity drop suggests a direct fairness test the authors did not run: compare predicted score distributions against human distributions by band. If underestimation concentrates among the top-scoring essays, the preprocessing is introducing construct-relevant bias rather than just noise.
  • The adaptive 80% resummarization rule could be turned into a length-adaptive compression scheme—targeting different compression ratios for long, high-score essays—making the observed loss pattern a design handle instead of a fixed drawback.
  • Cross-dataset replication on narrative or discipline-specific writing would clarify whether the mid-size-model advantage reflects model characteristics or the argumentative genre of this particular benchmark.
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. This paper proposes a hybrid automated essay scoring (AES) pipeline in which student essays from the ASAP 2.0 dataset are first summarized by one of three GPT-5 variants (GPT-5, mini, nano) with an adaptive 80% resummarization rule to enforce a 512-token cap; the summaries are embedded with Qwen3-Embedding-4B and concatenated with 22 handcrafted features from the original essay, then scored with XGBoost/LightGBM. The paper reports downstream QWK (0.8332–0.8435, Table 4), summarization quality metrics (Table 5), and score-band trends (Figs. 3–6). It concludes that GPT-5 mini gives the best cost–accuracy trade-off and that summarization fidelity declines for higher-scoring essays.

Significance. The study addresses a practical question—whether generative summarization can serve as a viable preprocessing step for fixed-length AES pipelines—and provides a within-pipeline comparison that holds classifier, features, and embedding constant. The paper is unusually explicit about its limitations: it states in §3.4 and §6 that truncation, Longformer, direct-LLM, and SOTA baselines are missing, and it frames the results as an initial controlled evaluation rather than a SOTA benchmark. This transparency is a strength. If the reported effects are real, the cost-efficiency implications (mini outperforming full GPT-5) would be practically useful. However, the central empirical claims currently rest on a narrow, undocumented experimental setup and small QWK differences with no uncertainty quantification.

major comments (4)
  1. [§3.3.1, §2] The asserted 512-token constraint is not justified. The downstream model is XGBoost/LightGBM on a fixed embedding vector, which has no token limit; the context limit of Qwen3-Embedding-4B is never stated. Table 3 shows max essay length ≈1,656 tokens, potentially within a 2k context. Since the paper's motivation is that summarization is needed to avoid truncation, the absence of a truncation or full-essay baseline (explicitly acknowledged in §3.4 and §6) leaves the central framing unsupported. Please either document the real constraint or add the missing baseline(s) and adjust the framing.
  2. [§4, Table 4; §6] The main result—GPT-5 mini QWK 0.8435 vs. GPT-5 0.8350 and nano 0.8332—is a difference of roughly 0.01 QWK, reported on a single split with no confidence intervals, bootstrap, or paired significance test. The paper's own §6 lists these as future work, yet the abstract and §5 state that GPT-5 mini 'achieves the highest agreement' and make a cost-efficiency recommendation. This is a load-bearing claim that needs statistical support (e.g., bootstrap CIs or a paired test on multiple stratified resamples) or must be reworded as a preliminary observation.
  3. [§3.3.1, §3.3.3] The experimental protocol is not fully specified. §3.3.3 says 'The intended reproducible setup is as follows,' which reads as a plan rather than a record; the summarization prompts, data-split seed and sizes, and tuned hyperparameter values (number of estimators, learning rate, max depth, etc.) are not reported, and no code is provided. Without these details, the claim in §3.3.1 that the comparison is 'attributable primarily to model-variant behavior' cannot be verified. Also, the paper uses a subset of ASAP 2.0 (13,845 essays, per Table 3) without describing the preprocessing that led to it; please document the selection criteria.
  4. [§5, Figs. 3–6, Table 3] The score-band decline in summarization quality is presented as evidence that 'more complex writing is harder to compress,' but the figures show band means without error bars or confidence intervals, and the high-score bands have small n (band 6: n=124; band 5: n=776). The interpretation is further confounded with essay length: higher-scoring essays are longer, and the paper does not control for length within bands. The trend is plausible, but the current analysis is insufficient to support the causal claim about complexity.
minor comments (5)
  1. [§2, Table 2] Pricing table 'accessed on July 1, 2026' — no URL or exact source is given; please provide a stable reference or data snapshot.
  2. [§3.2] The claim that Qwen3-Embedding-4B 'supports multilingual inputs' is not demonstrated for English essays; consider noting the model's language coverage or removing the claim.
  3. [§3.6.1, Eq. (2)] The ROUGE-N formula uses 'g∈Ref' but the denominator should sum over reference n-grams; the notation is understandable, but check consistency with the F1 definition and standard ROUGE formulations.
  4. [§4, Fig. 7] The correlation matrix is described in the text but the figure is not visible in this version; if the figure is missing in the final PDF, the text reference dangles.
  5. [References] Several 2025–2026 arXiv references are cited without DOIs or venue information; for a journal submission, please add publication details where available.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical within-pipeline comparison against human labels; no derivation reduces to its inputs.

full rationale

The paper makes no first-principles derivation whose output is equivalent to its inputs. Its central result (Table 4) is an empirical QWK comparison on held-out ASAP 2.0 responses, obtained by varying only the GPT summarizer inside an otherwise fixed hybrid pipeline (Qwen3 embeddings + 22 handcrafted features + XGBoost/LightGBM). QWK is computed from model predictions versus human scores, so no fitted constant is renamed as a prediction, and no equation equates an output to an input by construction. The only self-citation (Hua, Jiao & Wang, 2025) is used in §3.3.1 to motivate the 512-token summarization cap and in §6 to explain classifier choice; it is not used to define the outcome. The paper also explicitly disclaims truncation, Longformer, and direct-LLM baselines as future work rather than claiming they were completed. The possible unsupported nature of the 512-token constraint is a correctness/generalization concern, not a circular derivation. Thus no meaningful circularity is present.

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

The paper introduces no new entities or fitted physical constants. Its central comparison rests on design choices (512-token cap, 80% resummarization, classifier hyperparameters) and on domain assumptions about the validity of ASAP labels and summarization-quality metrics; the QWK values are empirical measurements, not derivations.

free parameters (3)
  • Classifier hyperparameters (XGBoost/LightGBM) = not reported
    Number of estimators, learning rate, max depth/leaves, subsampling, column sampling, and regularization are tuned on validation per §3.3.3, but exact values are omitted; they can change QWK and the ordering among GPT variants.
  • Summary length cap = 512 tokens
    Hand-chosen target to match a claimed downstream embedding constraint (§3.3.1); the actual context limit of Qwen3-Embedding-4B is not stated, so the cap may be unnecessary or arbitrary.
  • Adaptive resummarization step factor = 0.8
    When a summary exceeds 512 tokens, the target is reduced to 80% of the previous target (§3.3.1). This is a hand-selected rule with no ablation showing it preserves content or is optimal.
assumptions (5)
  • standard math Quadratic weighted kappa is the agreed evaluation metric and penalizes larger disagreements quadratically.
    Equation 1 defines QWK; the paper relies on this metric as the measure of human-machine agreement without questioning its appropriateness.
  • domain assumption ASAP 2.0 human holistic scores (1–6) are treated as ground truth for writing proficiency.
    The whole evaluation compares model scores to these labels; no analysis of rater reliability or label noise is provided.
  • domain assumption ROUGE, cosine similarity, entity coverage, keyphrase overlap, FactScore/QuestEval, and redundancy are assumed to measure construct-relevant summary quality.
    Sections 3.6.1–3.6.3 introduce these metrics as valid proxies for information preservation, but no evidence links them to educational construct validity.
  • ad hoc to paper The 512-token summary cap is required because the downstream scoring model is a fixed-length embedding pipeline.
    Section 3.3.1 asserts the cap matches 'the maximum input limit of 512 tokens required by the downstream scoring model,' but the downstream classifiers (XGBoost/LightGBM) consume embeddings, not tokens, and the embedding model's context limit is never given.
  • ad hoc to paper The adaptive 80% resummarization rule preserves semantic fidelity better than naive truncation.
    Section 3.3.1 states this rule improves efficiency and representational consistency, but it is asserted rather than validated against other compression strategies or truncation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cost-efficient generative AI summarization for scalable automated essay scoring in educational assessment." pith.science (2026). https://pith.science/paper/7H5JQMKI

@misc{pith2026260715829,
  author       = {Pith},
  title        = {Pith review of: Cost-efficient generative AI summarization for scalable automated essay scoring in educational assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7H5JQMKI}},
  note         = {Machine review of arXiv:2607.15829}
}
read the original abstract

Automated essay scoring (AES) enables scalable assessment and timely feedback but remains challenged by transformer input-length limitations, which can cause information loss when processing long essays. This study proposes a generative AI-assisted summarization framework to improve long-form essay representation while maintaining scoring reliability. Using the ASAP 2.0 dataset, we generate controlled-length summaries with three GPT-5 variants (GPT-5, GPT-5 mini, and GPT-5 nano) and use them as inputs for downstream AES models. To preserve original writing signals, handcrafted linguistic features extracted from full essays are integrated with summary representations to form a hybrid framework. The approach is evaluated in terms of scoring performance, summarization quality, and computational cost. Scoring reliability is measured using quadratic weighted kappa (QWK), while summary quality is assessed through lexical overlap, semantic similarity, information retention, and redundancy metrics. Results show that GPT-5 mini achieves the highest agreement with human ratings, whereas GPT-5 produces the strongest summarization quality. Summary quality decreases for higher-scoring essays, indicating that more complex writing is more difficult to compress without information loss. These findings reveal trade-offs among model capacity, summary fidelity, cost efficiency, and preservation of educational constructs. This study provides an initial controlled evaluation of GPT-based summarization for AES and identifies important baselines and ablation studies required for future generalization. Overall, generative AI summarization offers a promising approach for scalable writing assessment while requiring careful validation of information preservation and fairness.

Figures

Figures reproduced from arXiv: 2607.15829 by the authors.

Figure 1
Figure 1. Conceptual illustration of word embedding, in which textual units are mapped into a con [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Proposed AES pipeline. Original essays are adaptively summarized by GPT variants, en [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Semantic similarity across essay score levels for different GPT model variants. Lines represent [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: ROUGE-1 F1 score across essay score levels for different GPT model variants. Lines represent [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Entity coverage across essay score levels for different GPT model variants. Lines represent [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Keyphrase overlap across essay score levels for different GPT model variants. Lines represent [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Correlation matrix among summarization evaluation metrics, computed across the evaluated [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 1 canonical work pages

  1. [1]

    2024 , howpublished =

    Kaggle , title =. 2024 , howpublished =

    2024

  2. [2]

    Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and others , journal=

  3. [3]

    Fundamentals of Artificial Intelligence , pages=

    Natural language processing , author=. Fundamentals of Artificial Intelligence , pages=. 2020 , publisher=

    2020

  4. [4]

    arXiv preprint arXiv:2510.22830 , year=

    Exploration of Summarization by Generative Language Models for Automated Scoring of Long Essays , author=. arXiv preprint arXiv:2510.22830 , year=

    arXiv

  5. [5]

    Proceedings of the 22nd

    Xgboost: A scalable tree boosting system , author=. Proceedings of the 22nd

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Lightgbm: A highly efficient gradient boosting decision tree , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    16th International Conference on Educational Data Mining,

    Evaluating quadratic weighted kappa as the standard performance metric for automated essay scoring , author=. 16th International Conference on Educational Data Mining,. 2023 , organization=

    2023

  8. [8]

    Text Summarization Branches Out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text Summarization Branches Out , pages=

Show all 57 references
  1. [9]

    Journal of Physics: Conference Series , volume=

    The implementation of cosine similarity to calculate text relevance between two documents , author=. Journal of Physics: Conference Series , volume=. 2018 , organization=

    2018

  2. [10]

    Sixth Applied Natural Language Processing Conference , pages=

    Sentence reduction for automatic text summarization , author=. Sixth Applied Natural Language Processing Conference , pages=

  3. [11]

    Introduction to the

    Sang, Erik Tjong Kim and De Meulder, Fien , booktitle=. Introduction to the

  4. [12]

    Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing , pages=

    Textrank: Bringing order into text , author=. Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing , pages=

    2004

  5. [13]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (

    Evaluating the factual consistency of abstractive text summarization , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (

    2020

  6. [14]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    QuestEval: Summarization asks for fact-based evaluation , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  7. [15]

    Proceedings of the 55th Annual Meeting of the

    Get to the point: Summarization with pointer-generator networks , author=. Proceedings of the 55th Annual Meeting of the

  8. [16]

    The Phi Delta Kappan , volume=

    The imminence of grading essays by computer , author=. The Phi Delta Kappan , volume=

  9. [17]

    The Journal of Technology, Learning and Assessment , volume=

    An overview of automated scoring of essays , author=. The Journal of Technology, Learning and Assessment , volume=

  10. [18]

    Automated essay scoring with e-rater

    Attali, Yigal and Burstein, Jill , journal=. Automated essay scoring with e-rater

  11. [19]

    Educational Measurement: Issues and Practice , volume=

    A framework for evaluation and use of automated scoring , author=. Educational Measurement: Issues and Practice , volume=. 2012 , publisher=

    2012

  12. [20]

    Advancing Natural Language Processing in Educational Assessment , pages=

    Psychometric considerations when using deep learning for automated scoring , author=. Advancing Natural Language Processing in Educational Assessment , pages=. 2023 , publisher=

    2023

  13. [21]

    Proceedings of the 27th International Conference on Computational Linguistics , pages=

    Automated scoring: Beyond natural language processing , author=. Proceedings of the 27th International Conference on Computational Linguistics , pages=

  14. [22]

    2013 , publisher=

    Handbook of Automated Essay Evaluation: Current Applications and New Directions , author=. 2013 , publisher=

    2013

  15. [23]

    AI Magazine , volume=

    Automated essay evaluation: The Criterion online writing service , author=. AI Magazine , volume=

  16. [24]

    An automated essay scoring systems:

    Ramesh, Dadi and Sanampudi, Suresh Kumar , journal=. An automated essay scoring systems:. 2022 , publisher=

    2022

  17. [25]

    Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages=

    A neural approach to automated essay scoring , author=. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages=

    2016

  18. [26]

    Proceedings of the 54th Annual Meeting of the

    Automatic text scoring using neural networks , author=. Proceedings of the 54th Annual Meeting of the

  19. [27]

    Proceedings of the 21st Conference on Computational Natural Language Learning , pages=

    Attention-based recurrent convolutional neural network for automated essay scoring , author=. Proceedings of the 21st Conference on Computational Natural Language Learning , pages=

  20. [28]

    Advances in Neural Information Processing Systems , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

  21. [29]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle=

  22. [30]

    arXiv preprint arXiv:2108.07258 , year=

    On the opportunities and risks of foundation models , author=. arXiv preprint arXiv:2108.07258 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2303.08774 , year=

    arXiv

  24. [32]

    Sparks of Artificial General Intelligence: Early experiments with

    Bubeck, S. Sparks of Artificial General Intelligence: Early experiments with. arXiv preprint arXiv:2303.12712 , year=

    arXiv

  25. [33]

    Automated assessment and feedback in higher education using generative

    Naseer, Fawad and Khalid, Muhammad Usama and Ayub, Nafees and Rasool, Akhtar and Abbas, Tehseen and Afzal, Muhammad Waleed , booktitle=. Automated assessment and feedback in higher education using generative. 2024 , publisher=

    2024

  26. [34]

    Enhancing academic writing skills and motivation: assessing the efficacy of

    Song, Cuiping and Song, Yanping , journal=. Enhancing academic writing skills and motivation: assessing the efficacy of. 2023 , publisher=

    2023

  27. [35]

    TechTrends , volume=

    Generative artificial intelligence in education and its implications for assessment , author=. TechTrends , volume=. 2024 , publisher=

    2024

  28. [36]

    IEEE Transactions on Learning Technologies , volume=

    Automated essay scoring and revising based on open-source large language models , author=. IEEE Transactions on Learning Technologies , volume=. 2024 , publisher=

    2024

  29. [37]

    Computers & Education: Artificial Intelligence , year=

    Large language models in educational assessment: Opportunities and challenges , author=. Computers & Education: Artificial Intelligence , year=

  30. [38]

    Learning and Individual Differences , volume=

    Kasneci, Enkelejda and Se. Learning and Individual Differences , volume=. 2023 , publisher=

    2023

  31. [39]

    Computers and Education: Artificial Intelligence , volume=

    Applying large language models and chain-of-thought for automatic scoring , author=. Computers and Education: Artificial Intelligence , volume=. 2024 , publisher=

    2024

  32. [40]

    Large language models and automated essay scoring of

    Pack, Austin and Barrett, Alex and Escalante, Juan , journal=. Large language models and automated essay scoring of. 2024 , publisher=

    2024

  33. [41]

    Proceedings of the 18th workshop on innovative use of

    Automated evaluation of written discourse coherence using GPT-4 , author=. Proceedings of the 18th workshop on innovative use of

  34. [42]

    2004 , publisher=

    Graesser, Arthur C and McNamara, Danielle S and Louwerse, Max M and Cai, Zhiqiang , journal=. 2004 , publisher=

    2004

  35. [43]

    Fine-tuning

    Latif, Ehsan and Zhai, Xiaoming , journal=. Fine-tuning. 2024 , publisher=

    2024

  36. [44]

    Automated scoring of junior and senior high essays using

    Latifi, Syed and Gierl, Mark , journal=. Automated scoring of junior and senior high essays using. 2021 , publisher=

    2021

  37. [45]

    RoBERTa: A Robustly Optimized

    Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin , journal=. RoBERTa: A Robustly Optimized

  38. [46]

    He, Pengcheng and Liu, Xiaodong and Gao, Jianfeng and Chen, Weizhu , booktitle=

  39. [47]

    Proceedings of the International Conference on Learning Representations (

    ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators , author=. Proceedings of the International Conference on Learning Representations (

  40. [48]

    Proceedings of the Annual Meeting of the

    Longformer: The Long-Document Transformer , author=. Proceedings of the Annual Meeting of the. 2020 , pages=

    2020

  41. [49]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (

    Can large language models automatically score proficiency of written essays? , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (

    2024

  42. [50]

    Proceedings of the 29th International Conference on Computational Linguistics , pages=

    Automated Essay Scoring via Pairwise Contrastive Regression , author=. Proceedings of the 29th International Conference on Computational Linguistics , pages=. 2022 , address=

    2022

  43. [51]

    2025 , address=

    Wang, Jiong and Liu, Jie , booktitle=. 2025 , address=

    2025

  44. [52]

    2025 , address=

    Li, Xia and Pan, Wenjing , booktitle=. 2025 , address=. doi:10.18653/v1/2025.acl-long.1273 , url=

    2025 doi

  45. [53]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Graph-Based Multi-Trait Essay Scoring , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=. 2025 , address=. doi:10.18653/v1/2025.emnlp-main.1691 , url=

    2025 doi

  46. [54]

    arXiv preprint arXiv:2504.05736 , year=

    Rank-Then-Score: Enhancing Large Language Models for Automated Essay Scoring , author=. arXiv preprint arXiv:2504.05736 , year=

    arXiv

  47. [55]

    Shibata, Takumi and Miyamura, Yuichi , journal=

  48. [56]

    and Chakraborty, Anirban and Wr

    Kucia, Filip J. and Chakraborty, Anirban and Wr. arXiv preprint arXiv:2604.00259 , year=

    arXiv

  49. [57]

    Xia, Wei and Wu, Jin and Shi, Haoran and Wang, Xiangyu and Zheng, Chanjin , journal=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.