Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues

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

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

Pith's one-line read The paper finds that even state-of-the-art LLMs struggle to predict future tutor moves, while tutor strategy is highly indicative of student outcomes.

desk verdict First solid LLM benchmark for future tutor-move prediction, but the paper's second headline claim about tutor moves predicting outcomes is undercut by an internal regression inconsistency. read the letter →

arxiv 2507.06910 v1 pith:D4FJXV36 submitted 2025-07-09 cs.CL cs.CY

Fareya Ikram , Alexander Scarlatos , Andrew Lan This is my paper
classification cs.CLcs.CY
keywords tutoringdialoguestutormovesfuturemovepredictionstudentoutcomedialoguesuccesslargelanguagemodelsmathclassification
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 asks whether modern large language models can predict what a tutor will do next in a math tutoring dialogue, and whether that predicted strategy helps forecast student success. Across two tutoring corpora, the best future-tutor-move F1 scores are only 49% on MathDial and 27% on AlgebraNation, so the authors conclude that LLMs cannot yet forecast tutor strategy reliably. In contrast, student outcome prediction is more tractable, and tutor move annotations carry signal: adding previous move labels raises Llama 3's AlgebraNation dialogue-success F1 from 75.52 to 80.84. The paper positions tutor strategy prediction as an open problem that deserves better models, richer representations, or more work on move taxonomies.

What carries the argument

The central object is the tutor move, a categorical label capturing the pedagogical action of a tutor turn: MathDial uses four moves (probing, focus, telling, generic), AlgebraNation uses sixteen multi-label moves. The paper frames four prediction functions over dialogue history: future tutor move prediction, current move classification, dialogue success prediction, and next-turn student success prediction. The machinery includes a LoRA-fine-tuned Llama 3 model and a zero-shot prompted GPT-4o, together with three label-sequence baselines (a second-order Markov chain, logistic regression, and an LSTM) that isolate the contribution of textual context. Comparing text-only inputs with inputs that include previous move labels is what lets the paper separate the question of whether moves are predictable from the question of whether moves carry outcome signal.

What would settle it

A re-annotation study measuring inter-annotator agreement on the confusable move pairs—probing versus focus in MathDial and giving instruction versus giving explanation in AlgebraNation—would settle whether the low future-move F1 is a model limitation or a label-noise artifact, since poor agreement would mean the prediction ceiling is already near the reliability of the ground truth.

Watch

Extended reading notes

Core claim

The central discovery is an asymmetry: tutor moves are informative about student outcomes, but they are not predictable by current LLMs. On MathDial, Llama 3 with dialogue and previous move labels reaches 49.33 F1 for future tutor moves; on AlgebraNation, GPT-4o reaches 27.32 F1. The errors are systematic—the models most often confuse probing with focus in MathDial and giving instruction with giving explanation in AlgebraNation, the same pairs that annotators found hard to separate. Meanwhile, dialogue success prediction reaches a much higher level, and on the real-student AlgebraNation data the move labels improve Llama 3's F1 by more than five points. The authors take this as evidence that tutor strategy is highly indicative of student outcomes, but that forecasting the next pedagogical move remains beyond the reach of current generative LLMs.

Load-bearing premise

The conclusion depends on the tutor-move labels being a trustworthy operationalization of tutoring strategy, but the paper reports that MathDial annotators struggled to distinguish probing from focus and that the models' top misclassifications fall on exactly those confusable pairs.

Editorial extensions

If this is right

  • Future tutor moves are not reliably forecastable from dialogue history alone with current LLMs, so automated tutors cannot yet anticipate their next best pedagogical action from text.
  • Tutor move annotations are a usable intermediate signal: including them raises Llama 3's AlgebraNation dialogue-success F1 from 75.52 to 80.84.
  • The error pattern is concentrated on semantically adjacent move pairs, so improving or merging ambiguous label definitions is a more promising lever than simply scaling models.
  • Dialogue-level success prediction is substantially easier than turn-level success prediction, and on real student data the move-label gain is clearer than on the synthetic MathDial dialogues.

Reading between the lines

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

  • The paper's 'tutor behavior is unpredictable' conclusion may overstate the model limitation: because annotators themselves confused probing with focus in MathDial, the measured F1 ceiling could partly reflect label noise rather than a fundamental unpredictability of tutor behavior.
  • A direct testable extension would be to re-label both datasets with a merged or behaviorally precise taxonomy and measure whether future-move F1 rises sharply; if it does, annotation granularity, not prediction, is the bottleneck.
  • The success-prediction gain from moves suggests a practical loop the authors only gesture toward: an LLM tutor could select among predicted future moves by estimating which move maximizes the student's predicted success probability.
  • On AlgebraNation, Llama 3 collapses toward the majority label giving_instruction in future-move prediction, so any deployment of these predictions would need calibration or class-balanced training before the outputs are usable.
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 studies four tasks—current and future tutor move prediction, dialogue-level success prediction, and next-turn success prediction—using Llama 3 fine-tuned with LoRA and GPT-4o zero-shot on MathDial and AlgebraNation. It reports that future move prediction is hard (best F1 ≈49% on MathDial, ≈27% on AlgebraNation), that move labels sometimes improve outcome prediction, and that logistic regression coefficients suggest certain moves correlate with success. The paper is first to evaluate generative LLMs for the joint task of predicting tutor strategy and student outcomes.

Significance. If the central claims are robust, the paper provides a useful benchmark for a task that has received little attention, with careful held-out splits, multiple baselines, and error analysis. The qualitative analysis of confusable move categories is informative. However, the internal contradiction in the regression results and the ambiguity in the prompt condition construction need to be resolved before the conclusions about the value of move information and the intrinsic difficulty of strategy prediction can be trusted.

major comments (4)
  1. [Section 3.4, Table 9] Section 3.4 states that 'for MathDial, generic and probing have the strongest positive impact on success, whereas telling has a negative impact on success,' but Table 9 reports a significant negative coefficient for probing (-0.1867, chi-squared 6.7967, p=0.0091), with generic positive (+1.0635) and telling negative (-1.0627). This is an internal contradiction in the evidence for RQ3 and for the abstract's claim that 'tutor strategy is highly indicative of student outcomes.' If Table 9 is correct, the prose misreports the finding; if the prose is correct, the regression output is misreported. Please correct the text or the analysis and re-state the conclusion accordingly.
  2. [Section 3.4, Table 3] The paper claims in Section 3.4 that 'previous move labels improve performance for Llama 3,' but Table 3 shows that on MathDial, adding move labels slightly decreases weighted F1 for dialogue success (70.11 vs 69.96) and for turn success (50.54 vs 49.27); the only consistent improvement is on AlgebraNation dialogue success (75.52 to 80.84). The claim should be qualified to the AlgebraNation dataset, and the authors should report variance (e.g., multiple seeds) or significance tests before concluding that move annotations generally help outcome prediction.
  3. [Section 2.2, Figures 7-8] The zero-shot GPT-4o prompts shown in Figures 7 and 8 embed the previous tutor move labels directly in the dialogue text (e.g., '(generic)', '(focus)' in Figure 7; '["giving_instruction"]' in Figure 8). If the same prompt template is used in the 'Dialogue' condition, then that condition is not actually a text-only input; it already contains the previous move labels, which would conflate the 'Dialogue' and 'Dialogue & Moves' conditions and invalidate the comparison for GPT-4o. Please clarify whether the dialogue-only prompt strips these labels, and if not, re-run the experiments with a properly controlled prompt.
  4. [Section 3.3, Abstract] The conclusion that 'even state-of-the-art LLMs struggle to predict future tutor strategy' rests on the assumption that the ground-truth move labels are reliable. The paper itself notes in Section 3.3 that 'MathDial authors also note that annotators had difficulty differentiating between probing and focus moves' (citing Macina et al., 2023), and the top misclassifications for both models are precisely these confusable pairs (Tables 4-7). Without reporting inter-annotator agreement or a noise-adjusted evaluation, the low future-move F1 (49% MathDial, 27% AlgebraNation) may reflect annotation noise rather than an intrinsic limit of LLMs. Please provide an estimate of label reliability and discuss this alternative interpretation.
minor comments (5)
  1. [Section 2.2, References] The LoRA method is attributed to Saari et al. (2018), which is about Long Range (LoRa) wireless; the Low-Rank Adaptation method should cite Hu et al. (2021).
  2. [Section 3.3, Table 2] The statement 'LLMs improve over baselines for future move prediction' is not true for Llama 3 on AlgebraNation, whose F1 (20.20) is below the LSTM baseline (23.17). Please qualify the claim.
  3. [Section 3.3] The phrase 'GPT-4o was not trained on AlgebraNation' is trivially true for a zero-shot model; presumably the authors mean it was not fine-tuned on that dataset.
  4. [Tables 8-9] The logistic regression analysis does not report confidence intervals or multiple-testing correction; adding these would strengthen the RQ3 claims.
  5. [General] The paper does not state whether code is available; if applicable, please add a reproducibility statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's predictions are evaluated on held-out test splits with external labels and standard baselines; the cited prior work supplies data and related context, not the target result.

full rationale

No circular step is present. RQ1 future-move prediction and RQ2 outcome prediction are evaluated on held-out test splits (MathDial 1,947/537 and AlgebraNation 1,854/464), and the reported F1 values are direct measurements rather than fitted quantities renamed as predictions. The ground-truth tutor move labels come from externally published datasets (MathDial, Macina et al. 2023; AlgebraNation, Lyu et al. 2024), and turn-level student correctness labels are taken from Scarlatos et al. (2025a), a prior published paper by co-authors. Although this is a self-citation, it is used as a data source, not as a load-bearing argument for the paper's central claim, and the labels are externally generated and falsifiable outside the present fitted values. The logistic-regression coefficients in Tables 8 and 9 are descriptive associations, not predictions derived from the target claim, so RQ3 does not reduce to an input. The main conclusion that even state-of-the-art LLMs struggle to predict future tutor moves is an empirical finding independent of any prior claim by the same authors. A separate correctness concern is noted but is not circular: Section 3.4 states that in MathDial 'probing' has a strong positive impact on success, while Table 9 reports a negative coefficient for probing (-0.1867, p=0.0091). This inconsistency affects the reliability of RQ3's interpretation, but it does not make the derivation self-referential.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper adds no new theoretical constructs; it applies existing language models to existing datasets. All predictive results depend on model hyperparameters tuned on validation data, and the correlational findings depend on logistic regression coefficients fitted to the training set. The main implicit load-bearing assumptions are label validity and label reliability, which the paper itself partly questions.

free parameters (7)
  • LoRA rank = 8
    Selected by hyperparameter search over {4,8,16,32} on validation set; affects fine-tuned Llama 3 performance.
  • LoRA alpha = 16
    Set to 16 for final training; standard choice paired with rank 8.
  • Learning rate for Llama 3 = 1e-4
    Selected from {5e-5, 1e-4, 2e-4, 3e-4} on validation set.
  • LSTM hidden dimension = 128
    Selected from {64,128,256,512} on validation set.
  • LSTM dropout = 0.3
    Selected from {0.1,0.3,0.5} on validation set.
  • Multi-label classification threshold = 0.85
    Tuned on validation set for AlgebraNation multi-label predictions.
  • Logistic regression coefficients (Tables 8-9) = Various, e.g., generic +1.0635 in MathDial
    Fitted to training data and used to infer which tutor moves correlate with dialogue success; central to the claim that moves affect outcomes.
assumptions (4)
  • domain assumption MathDial's four move labels and AlgebraNation's 16 move labels are valid operationalizations of tutor strategy.
    All prediction tasks and the regression analysis assume the annotation schemes capture meaningful pedagogical actions; the paper notes annotator confusion between probing and focus in MathDial.
  • domain assumption Turn-level student correctness labels in MathDial, taken from Scarlatos et al. (2025a), are accurate ground truth.
    Used as supervision for turn success and dialogue success prediction; the paper does not re-validate these labels.
  • domain assumption MathDial's LLM-simulated student responses approximate real student behavior well enough for the conclusions to transfer.
    Acknowledged limitation in the Limitations section; affects both the difficulty of move prediction and the interpretation of outcome prediction.
  • domain assumption The random train/test splits are representative and reproducible, despite no seed being reported.
    Results depend on a single split; without a seed, exact replication is difficult.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues." pith.science (2026). https://pith.science/paper/D4FJXV36

@misc{pith2026250706910,
  author       = {Pith},
  title        = {Pith review of: Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4FJXV36}},
  note         = {Machine review of arXiv:2507.06910}
}
read the original abstract

Tutoring dialogues have gained significant attention in recent years, given the prominence of online learning and the emerging tutoring abilities of artificial intelligence (AI) agents powered by large language models (LLMs). Recent studies have shown that the strategies used by tutors can have significant effects on student outcomes, necessitating methods to predict how tutors will behave and how their actions impact students. However, few works have studied predicting tutor strategy in dialogues. Therefore, in this work we investigate the ability of modern LLMs, particularly Llama 3 and GPT-4o, to predict both future tutor moves and student outcomes in dialogues, using two math tutoring dialogue datasets. We find that even state-of-the-art LLMs struggle to predict future tutor strategy while tutor strategy is highly indicative of student outcomes, outlining a need for more powerful methods to approach this task.

Figures

Figures reproduced from arXiv: 2507.06910 by the authors.

Figure 1
Figure 1. Label distribution of tutor move classification for GPT-4o and Llama 3 trained with dialogue and tutor [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Label distribution of future tutor move prediction for GPT-4o and Llama 3 trained with dialogue and tutor [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Left: Distribution of dialogue success classification in MathDial using Llama 3. Right: Distribution of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Tutor move distribution of AlgebraNation dataset. Few classes make up the majority of the distribution. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Tutor move distribution of MathDial dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Left: Distribution of turn-level student success in MathDial. Center: Distribution of final turn student [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Full prompt used to classify the final tutor move in a MathDial dialogue. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Full prompt used to classify the final tutor move in an AlgebraNation dialogue. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 23 canonical work pages

  1. [1]

    Jacobs, and Sidney K

    Mark Abdelshiheed, Jennifer K. Jacobs, and Sidney K. D'Mello. 2024. Aligning tutor discourse supporting rigorous thinking with tutee content mastery for predicting math achievement. In Artificial Intelligence in Education, pages 150--164, Cham. Springer Nature Switzerland

    work page 2024

  2. [2]

    Inneke Berghmans, Lotte Michiels, Sara Salmon, Filip Dochy, and Katrien Struyven. 2014. Directive versus facilitative peer tutoring? a view on students’ appraisal, reported learning gains and experiences within two differently-tutored learning environments. Learning Environments Research, 17:437--459

    work page 2014

  3. [3]

    Koedinger, and Vincent Aleven

    Conrad Borchers, Kexin Yang, Jionghao Lin, Nikol Rummel, Kenneth R. Koedinger, and Vincent Aleven. 2024. https://doi.org/10.5281/zenodo.12729784 Combining dialog acts and skill modeling: What chat interactions enhance learning rates during ai-supported peer tutoring? In Proceedings of the 17th International Conference on Educational Data Mining, pages 117...

    doi:10.5281/zenodo.12729784 2024

  4. [4]

    Kristy Elizabeth Boyer, Eun Young Ha, Michael D Wallis, Robert Phillips, Mladen A Vouk, and James C Lester. 2009. Discovering tutorial dialogue strategies with hidden markov models. In Artificial Intelligence in Education, pages 141--148. IOS Press

    work page 2009

  5. [5]

    Carnegie Learning . 2024. Livehint overview. Online: https://support.carnegielearning.com/help-center/math/livehint/article/livehint-overview/

    work page 2024

  6. [6]

    Jiahao Chen, Zitao Liu, Mingliang Hou, Xiangyu Zhao, and Weiqi Luo. 2024. https://doi.org/10.1145/3627673.3679108 Multi-turn classroom dialogue dataset: Assessing student performance from one-on-one conversations . In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM '24, page 5333–5337, New York, NY, USA. ...

    arXiv 2024

  7. [7]

    Dorottya Demszky and Heather Hill. 2023. The ncte transcripts: A dataset of elementary math classroom transcripts. In Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023), pages 528--538

    2023

  8. [8]

    Dorottya Demszky, Jing Liu, Zid Mancenido, Julie Cohen, Heather Hill, Dan Jurafsky, and Tatsunori B Hashimoto. 2021. Measuring conversational uptake: A case study on student-teacher interactions. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Proces...

    2021

Show all 40 references
  1. [9]

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

    2024 arXiv

  2. [10]

    Ananya Ganesh, Martha Palmer, and Katharina Kann. 2021. https://doi.org/10.18653/v1/2021.findings-acl.418 What would a teacher do? P redicting future talk moves . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4739--4751, Online. Associati...

    2021 doi

  3. [11]

    Thomas Huber, Christina Niklaus, and Siegfried Handschuh. 2023. https://doi.org/10.18653/v1/2023.bea-1.59 Enhancing educational dialogues: A reinforcement learning approach for generating AI teacher responses . In Proceedings of the 18th Workshop on Innovative Use of NLP for B...

    2023 doi

  4. [12]

    Khan Academy . 2023. Supercharge your teaching experience with khanmigo. Online: https://www.khanmigo.ai/

    2023

  5. [13]

    Seanie Lee, Jianpeng Cheng, Joris Driesen, Alexandru Coca, and Anders Johannsen. 2024. https://doi.org/10.18653/v1/2024.naacl-long.6 Effective and efficient conversation retrieval for dialogue state tracking with implicit text summaries . In Proceedings of the 2024 Conference ...

    2024 doi

  6. [14]

    Hang Li, Tianlong Xu, Jiliang Tang, and Qingsong Wen. 2024. https://arxiv.org/abs/2406.13885 Knowledge tagging system on math questions via llms with flexible demonstration retriever . Preprint, arXiv:2406.13885

    2024 arXiv

  7. [15]

    Jionghao Lin, Shaveen Singh, Lele Sha, Wei Tan, David Lang, Dragan Ga s evi \'c , and Guanliang Chen. 2022. Is it a good move? mining effective tutoring strategies from human--human tutorial dialogues. Future Generation Computer Systems, 127:194--207

    2022

  8. [16]

    Baraniuk

    Naiming Liu, Shashank Sonkar, and Richard G. Baraniuk. 2025. https://arxiv.org/abs/2502.15140 Do llms make mistakes like students? exploring natural alignment between language models and human error patterns . Preprint, arXiv:2502.15140

    2025 arXiv

  9. [17]

    Bailing Lyu, Chenglu Li, Hai Li, Wangda Zhu, and Wanli Xing. 2024. Explaining technical, social, and discursive participation in online mathematical discussions. Distance Education, pages 1--24

    2024

  10. [18]

    Jakub Macina, Nico Daheim, Sankalan Chowdhury, Tanmay Sinha, Manu Kapur, Iryna Gurevych, and Mrinmaya Sachan. 2023. M ath D ial: A dialogue tutoring dataset with rich pedagogical properties grounded in math reasoning problems. In Findings of the Association for Computational L...

    2023

  11. [19]

    Hunter McNichols and Andrew Lan. 2025. https://arxiv.org/abs/2503.07928 The studychat dataset: Student dialogues with chatgpt in an artificial intelligence course . Preprint, arXiv:2503.07928

    2025

  12. [20]

    Baptiste Moreau-Pernet, Yu Tian, Sandra Sawaya, Peter Foltz, Jie Cao, Brent Milne, and Thomas Christie. 2024. https://doi.org/10.1145/3657604.3664664 Classifying tutor discursive moves at scale in mathematics classrooms with large language models . In Proceedings of the Eleven...

    2024

  13. [21]

    Andre Nickow, Philip Oreopoulos, and Vincent Quan. 2020. The impressive effects of tutoring on prek-12 learning: A systematic review and meta-analysis of the experimental evidence. Working Paper 27476, National Bureau of Economic Research

    2020

  14. [22]

    Benjamin D Nye, Arthur C Graesser, and Xiangen Hu. 2014. Autotutor and family: A review of 17 years of natural language tutoring. International Journal of Artificial Intelligence in Education, 24:427--469

    2014

  15. [23]

    OpenAI. 2024 a . Gpt-4.1 prompting guide. https://cookbook.openai.com/examples/gpt4-1_prompting_guide. Accessed: 2025-04-21

    2024

  16. [24]

    OpenAI. 2024 b . https://openai.com/index/hello-gpt-4o/ Hello gpt-4o . Accessed: 2025-02-19

    2024

  17. [25]

    Mika Saari, A Muzaffar bin Baharudin, Pekka Sillberg, Sami Hyrynsalmi, and Wanglin Yan. 2018. Lora—a survey of recent research trends. In 2018 41st international convention on information and communication technology, electronics and microelectronics (MIPRO), pages 0872--0877. IEEE

    2018

  18. [26]

    Baker, and Andrew Lan

    Alexander Scarlatos, Ryan S. Baker, and Andrew Lan. 2025 a . Exploring knowledge tracing in tutor-student dialogues using llms. In Proceedings of the 15th Learning Analytics and Knowledge Conference, LAK 2025, Dublin, Ireland, March 3-7, 2025 . ACM

    2025

  19. [27]

    Alexander Scarlatos, Naiming Liu, Jaewook Lee, Richard Baraniuk, and Andrew Lan. 2025 b . https://arxiv.org/abs/2503.06424 Training llm-based tutors to improve student learning outcomes in dialogues . Preprint, arXiv:2503.06424

    2025 arXiv

  20. [28]

    Shashank Sonkar, Kangqi Ni, Sapana Chaudhary, and Richard Baraniuk. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.797 Pedagogical alignment of large language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13641--13650, Miami, F...

    2024 doi

  21. [29]

    Martin, and Tamara Sumner

    Abhijit Suresh, Jennifer Jacobs, Charis Harty, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022. The T alk M oves dataset: K-12 mathematics lesson transcripts annotated for teacher and student discursive moves. In Proceedings of the Thirteenth Language Resources and ...

    2022

  22. [30]

    Ana \"i s Tack, Ekaterina Kochmar, Zheng Yuan, Serge Bibauw, and Chris Piech. 2023. https://doi.org/10.18653/v1/2023.bea-1.64 The BEA 2023 shared task on generating AI teacher responses in educational dialogues . In Proceedings of the 18th Workshop on Innovative Use of NLP for...

    2023 doi

  23. [31]

    LearnLM Team, Abhinit Modi, Aditya Srikanth Veerubhotla, Aliya Rysbek, Andrea Huber, Brett Wiltshire, Brian Veprek, Daniel Gillick, Daniel Kasenberg, Derek Ahmed, Irina Jurenka, James Cohan, Jennifer She, Julia Wilkowski, Kaiz Alarakyia, Kevin R. McKee, Lisa Wang, Markus Kunes...

    2024 arXiv

  24. [32]

    Justin Vasselli, Christopher Vasselli, Adam Nohejl, and Taro Watanabe. 2023. https://doi.org/10.18653/v1/2023.bea-1.63 NAIST eacher: A prompt and rerank approach to generating teacher utterances in educational dialogues . In Proceedings of the 18th Workshop on Innovative Use o...

    2023 doi

  25. [33]

    Deliang Wang, Dapeng Shan, Yaqian Zheng, and Gaowei Chen. 2023. Teacher talk moves in k12 mathematics lessons: Automatic identification, prediction explanation, and characteristic exploration. In Artificial Intelligence in Education, pages 651--664, Cham. Springer Nature Switzerland

    2023

  26. [34]

    Rose Wang and Dorottya Demszky. 2024. https://doi.org/10.18653/v1/2024.naacl-demo.6 Edu- C onvo K it: An open-source library for education conversation data . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: ...

    2024 doi

  27. [35]

    Rose Wang, Qingyang Zhang, Carly Robinson, Susanna Loeb, and Dorottya Demszky. 2024 a . Bridging the novice-expert gap via models of decision-making: A case study on remediating math mistakes. In Proceedings of the 2024 Conference of the North American Chapter of the Associati...

    2024

  28. [36]

    Rose E Wang, Ana T Ribeiro, Carly D Robinson, Susanna Loeb, and Dora Demszky. 2024 b . Tutor copilot: A human-ai approach for scaling real-time expertise. arXiv preprint arXiv:2410.03017

    2024 arXiv

  29. [37]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771

    2019 arXiv

  30. [38]

    Stella Xin Yin, Zhengyuan Liu, Dion Hoe-Lian Goh, Choon Lang Quek, and Nancy F. Chen. 2025. https://doi.org/10.1145/3706468.3706474 Scaling up collaborative dialogue analysis: An ai-driven approach to understanding dialogue patterns in computational thinking education . In Pro...

    2025

  31. [39]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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