Pith. sign in

REVIEW 2 major objections 5 minor 51 references

Automatic Generation of Inference Making Questions for Reading Comprehension Assessments

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper demonstrates that GPT-4o, prompted with a taxonomy of bridging inferences, can produce multiple-choice reading-comprehension questions that expert reviewers judge usable in grades 3–12, but it does not reliably generate the…

desk verdict Solid empirical study and a genuinely interesting null result on inference-type targeting, but the 'suitable for operational use' claim outruns the evidence. read the letter →

arxiv 2506.08260 v1 pith:PEMHQPAJ submitted 2025-06-09 cs.CL cs.AI

classification cs.CLcs.AI
keywords readingcomprehensioninferencemakingautomaticitemgenerationlargelanguagemodelsGPT-4ochain-of-thoughtpromptingbridgingdiagnosticassessment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish a practical pipeline for automatically generating diagnostic reading-comprehension questions that target specific kinds of inference-making, and to measure where that pipeline works and where it fails. The authors introduce a three-type taxonomy of bridging inferences—pronominal bridging, text-connecting, and gap-filling—validate it against an existing diagnostic item bank, and then use few-shot prompting of GPT-4o to write new multiple-choice questions. Expert review of 357 generated items found 93.8% acceptable in general quality, with the best condition reaching 96.7%, yet only 42.6% matched the inference type the prompt asked for. The paper concludes that automatic generation plus human screening is a scalable path toward diagnostic reading-comprehension assessments, while making clear that type-targeted generation remains an unsolved problem.

What carries the argument

The carrying machinery is a taxonomy of bridging-inference questions—pronominal, pronominal bridging, text-connecting, and gap-filling—used in three places: to annotate an existing item bank, to structure hand-written training examples, and to define the evaluation rubric. Generation is done by few-shot prompting GPT-4o with inference-type-specific step-by-step instructions (find a pronoun, generate a multiple-choice question with distractors, follow additional rules), with a chain-of-thought variant that also asks the model to output a text hint and a reasoning note. The evaluation rubric scores each generated item on general quality, correct inference type, and adequacy of the model's reasoning, and the paper's headline numbers come from majority votes of three expert raters after an adjudication round.

What would settle it

A study that administers a sample of the expert-approved generated items to a large grade 3–12 student sample and calibrates them with item response theory would settle the operational-readiness claim: if expert-approved items show poor discrimination, unexpected difficulty, or differential functioning across student groups, then expert review alone is not sufficient.

Watch

Extended reading notes

Core claim

On the authors' own terms, the central finding is a contrast. GPT-4o, given a passage and a small set of manually written examples, can produce multiple-choice reading-comprehension items that expert reviewers rate highly: 93.8% of 357 generated items passed the general quality bar (correct key, clearly wrong distractors, developmentally appropriate for grades 3–12), with the best condition reaching 96.7%. But when the prompt asks for a particular bridging-inference type, the model matches the requested type only 42.6% of the time overall (best condition 46.1%), and text-connecting questions are especially unreliable at 24.1%. The paper also reports that the mix of inference types in the generated pool nonetheless resembles the mix in an existing expert-written item bank, and that chain-of-thought prompting gives no clear quality advantage over standard prompting. The authors conclude that automatic generation plus expert screening is a viable path for expanding diagnostic item banks, while type-targeted generation remains a hard, unsolved part of the task.

Load-bearing premise

The paper's operational-readiness claim rests entirely on expert human judgment; no student response data were collected, so the items' real difficulty, discrimination, and fairness are unknown.

Editorial extensions

If this is right

  • A human-screened LLM pipeline can supply a substantial share of new multiple-choice reading-comprehension items: 93.8% of all 357 generated questions and 96.7% in the best condition passed expert review on general quality.
  • More training examples improved generation quality, while chain-of-thought prompting gave no clear advantage over standard few-shot prompting.
  • Per-item control of inference type is not yet reliable, with the best condition matching the requested type only 46.1% of the time and text-connecting questions being hardest at 24.1%.
  • The overall distribution of inference types in the generated item bank closely matched the operational item bank, so aggregate coverage may be acceptable even though individual targeting fails.
  • Human expert evaluation remains a necessary gate; the model's reasoning was judged adequate for only about 37–39% of the chain-of-thought items, which may explain why adding chain-of-thought did not help.

Reading between the lines

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

  • If expert judgments do not track real item behavior, the 93.8% claim would overstate readiness; a practical validation would re-run the pipeline on a new passage set and compare expert scores with actual student response data.
  • The decoupling of general quality from type accuracy suggests an alternative design the paper does not pursue: prompt for any good inference item, then classify its inference type after generation, using the taxonomy as a post-hoc filter rather than a generation constraint.
  • The fact that text-connecting was hardest for both the model and the human question-writers points to an underlying construct difficulty; a testable extension is whether supplying more explicit text-connecting training examples, or constraining candidate noun-phrase pairs, closes that gap.
  • A testable extension of the distribution-matching result is to accept the natural type mix of an LLM-generated item bank and use post-hoc annotation to balance parallel test forms, rather than forcing per-item type control.
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

2 major / 5 minor

Summary. The paper introduces a taxonomy of bridging-inference types for reading comprehension (pronominal, pronominal bridging, text-connecting, gap-filling), validates it by annotating an operational item bank, and then uses the taxonomy to prompt GPT-4o via few-shot and chain-of-thought prompting to generate multiple-choice inference questions for new expository passages. The generated items are evaluated by three expert raters on general item quality, inference-type accuracy, and reasoning quality. The main empirical findings are that 93.8% of the 357 generated items are judged to be of good general quality, but only 42.6% match the requested inference type, and chain-of-thought prompting provides no clear benefit. The paper concludes that LLM-based item generation with human review is a promising path toward scalable diagnostic RC assessments.

Significance. If the results hold, the paper makes several useful contributions: a literature-grounded taxonomy of inference types that appears to capture meaningful distinctions in operational RC items; a new task formulation for generating questions targeting specific inference types; and a carefully designed human evaluation protocol with high inter-rater agreement (Fleiss' kappa 0.57-0.83). The public release of the example item bank is a concrete reproducibility asset. The work also provides a cautionary data point about LLMs' limited ability to control inference type, which is relevant to the educational NLP community. However, the central practical claim - that items are 'suitable for operational use' - is not fully supported by the evidence collected, as detailed in the major comments.

major comments (2)
  1. [Abstract; Section 5 (Results)] The claim that generated questions are 'suitable for operational use in grade 3-12 contexts' (Abstract and Section 5) overstates what the evaluation establishes. The rubric in Table 2 assesses answer correctness, distractor plausibility, and developmental safety; it does not measure item difficulty, discrimination, or bias, which are prerequisites for operational scoring and for valid inferences about student ability. The Limitations section explicitly acknowledges that the study 'did not include pilot testing in real-world settings to evaluate how the generated items perform with actual student responses.' Expert review alone does not establish operational readiness. The evidence supports the weaker claim that expert raters judged the items as acceptable, and the wording should be adjusted throughout, including the abstract and conclusion.
  2. [Section 4.3 (Human Evaluation)] The three raters are also the authors who designed the prompting framework and wrote the training examples. The quality ratings are therefore not independent of the generation pipeline, and the high inter-rater agreement among co-authors does not rule out shared expectations or confirmation bias. Because the 93.8% quality figure rests entirely on these ratings, the paper should either recruit external annotators or explicitly discuss this threat to validity and justify why the conclusions are not affected. At minimum, the authors should state this limitation and consider adding a blinded re-rating by a non-author as a robustness check.
minor comments (5)
  1. [Section 3.1] There is a typo in the sentence 'demonstrate its its value for future item development' - the word 'its' is repeated.
  2. [Section 5] In the paragraph discussing the distribution of inference types, 'the collect of items' should read 'the collection of items.'
  3. [Section 5] The statement that 'our results indicate no statistically significant differences in generation performance across the various prompting conditions' is not accompanied by any test statistic, effect size, or p-value. Please report the statistical details or qualify the claim as descriptive.
  4. [Section 4.2] The frequency penalty sweep from 0 to 0.3 is mentioned as yielding an optimum at 0.2, but no data are shown to support this choice. A supplementary table or figure of generation quality across penalty values would strengthen reproducibility.
  5. [Figure 5] The stacked bar chart is difficult to read because the legend keys and the 'true type' labels are visually similar and overlap. Consider replacing the figure with a confusion matrix or adding percentage labels to each segment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical human-evaluation study whose central quality claims rest on independent expert ratings, not on fitted parameters or self-referential definitions.

full rationale

This paper makes no formal derivation or first-principles prediction; it reports an empirical pipeline (taxonomy construction, few-shot GPT-4o item generation, and human evaluation). The taxonomy is adapted from external psycholinguistic literature (Graesser et al., 1994; Cain and Oakhill, 1999; Haviland and Clark, 1974) and used consistently for annotation, prompt construction, and evaluation; using the same taxonomy as both generation target and evaluation criterion is a standard operationalization and does not make the outcome entailed by the input. The headline claim that 93.8% of items were 'good-quality' comes from three expert raters independently scoring 357 generated items under an explicit rubric (Table 2), with agreement reported in Table 3; it is not computed from any equation or fitted parameter. The temperature and frequency-penalty settings are prompt-engineering choices, and the paper does not use them to predict any of its reported outcomes. Self-citations (Sabatini et al., 2019; Zelikman et al., 2023; Yeatman et al., 2024) are contextual, used for background or future-work directions, and are not load-bearing for the central empirical claims. The main legitimate concern is that 'suitable for operational use' outruns the evidence: the Limitations section explicitly states, 'Our study did not include pilot testing in real-world settings to evaluate how the generated items perform with actual student responses,' so item difficulty, discrimination, and bias were not measured. That is an external-validity / overclaim issue, not circularity, because the claimed conclusion is not definitionally equivalent to the measured rubric. No circular step can be exhibited with a specific reduction, so the appropriate finding is no significant circularity.

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

The central claim rests on the validity of the taxonomy and the reliability of expert judgment. No fitted parameters are involved in the scientific claim; the free parameters listed are prompt-engineering choices that affect generation diversity and reproducibility but do not load the main conclusion.

free parameters (3)
  • temperature = 0
    Set to 0 for reproducibility as stated in Section 4.2. Not fit to data but affects sampling determinism.
  • frequency_penalty = 0.2
    Tuned to consistently generate three diverse items without harming quality (Section 4.2).
  • number_of_training_examples = 4 or 6 passages
    Experimental condition; results show 6 improves performance slightly. Not a parameter fit to the outcome.
assumptions (3)
  • domain assumption The four-part taxonomy (pronominal, pronominal bridging, text-connecting, gap-filling) captures meaningful distinctions in inference-making for reading comprehension.
    Grounded in prior literature (Cain and Oakhill, Graesser et al.) and validated via annotation of an operational item bank with kappa = 0.83, but still a modeling choice.
  • domain assumption Expert human evaluation is a valid measure of item quality and operational suitability.
    Three expert raters with high agreement, but no student response data validate that the items actually function as intended for the target population.
  • domain assumption GPT-4o is representative of current LLM capabilities for this task.
    Only one model was tested, as acknowledged in Limitations. The findings may not generalize to other LLMs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic Generation of Inference Making Questions for Reading Comprehension Assessments." pith.science (2026). https://pith.science/paper/PEMHQPAJ

@misc{pith2026250608260,
  author       = {Pith},
  title        = {Pith review of: Automatic Generation of Inference Making Questions for Reading Comprehension Assessments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEMHQPAJ}},
  note         = {Machine review of arXiv:2506.08260}
}
read the original abstract

Inference making is an essential but complex skill in reading comprehension (RC). Some inferences require resolving references across sentences, and some rely on using prior knowledge to fill in the detail that is not explicitly written in the text. Diagnostic RC questions can help educators provide more effective and targeted reading instruction and interventions for school-age students. We introduce a taxonomy of inference types for RC and use it to analyze the distribution of items within a diagnostic RC item bank. Next, we present experiments using GPT-4o to generate bridging-inference RC items for given reading passages via few-shot prompting, comparing conditions with and without chain-of-thought prompts. Generated items were evaluated on three aspects: overall item quality, appropriate inference type, and LLM reasoning, achieving high inter-rater agreements above 0.90. Our results show that GPT-4o produced 93.8% good-quality questions suitable for operational use in grade 3-12 contexts; however, only 42.6% of the generated questions accurately matched the targeted inference type. We conclude that combining automatic item generation with human judgment offers a promising path toward scalable, high-quality diagnostic RC assessments.

Figures

Figures reproduced from arXiv: 2506.08260 by the authors.

Figure 1
Figure 1. Overview of automatic item generation and human evaluation. We use GPT-4o to generate bridging￾inference RC items for given reading passages via few-shot prompting, comparing conditions with and without chain-of-thought prompts. We prompt each inference type separately: pronominal bridging, text-connecting, and gap-filling inferences. Human evaluation focuses on general item quality, inference type appropriateness, … view at source ↗
Figure 2
Figure 2. Distribution of different inference types in an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Few-shot prompt for generating pronomi￾nal bridging inference questions. The system prompt (beige background) defines the inference type and out￾lines expert-inspired steps. Training examples (provided in the prompt) follow. In the standard condition, only the question and answer key (green) are shown; in the CoT condition, text hints and reasoning (blue) are also included. A new passage is provided in the user prom… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Examples of LLM-generated RC items via Chain-of-Thought prompting with 6 training passages. Left: [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Human evaluation of inference-type accuracy. Each panel displays the distribution of true inference types [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of item inference type cover [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Few-shot prompting using Chain-of-Thought for generating text-connecting inference. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Few-shot prompting using Chain-of-Thought for generating gap-filling inference. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 36 canonical work pages

  1. [1]

    Manish Agarwal, Rakshit Shah, and Prashanth Mannem. 2011. Automatic question generation using discourse cues. In Proceedings of the Sixth Workshop on Innovative Use of NLP for Building Educational Applications, pages 1--9

  2. [2]

    Jun Araki, Dheeraj Rajagopal, Sreecharan Sankaranarayanan, Susan Holm, Yukari Yamakawa, and Teruko Mitamura. 2016. Generating questions and multiple-choice answers using semantic analysis of texts. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1125--1136

  3. [3]

    Nischal Ashok Kumar, Nigel Fernandez, Zichao Wang, and Andrew Lan. 2023. https://doi.org/10.18653/v1/2023.bea-1.22 Improving reading comprehension question generation with data augmentation and overgenerate-and-rank . In Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023), pages 247--259, Toronto, Can...

  4. [4]

    Yigal Attali, Andrew Runge, Geoffrey T LaFlair, Kevin Yancey, Sarah Goodwin, Yena Park, and Alina A Von Davier. 2022. The interactive reading task: Transformer-based automatic item generation. Frontiers in Artificial Intelligence, 5:903077

  5. [5]

    Nihat Bayat and G \"o khan C etinkaya. 2020. The relationship between inference skills and reading comprehension. Education and Science

  6. [6]

    Claudine Bowyer-Crane and Margaret J Snowling. 2005. Assessing children's inference generation: What do tests of reading comprehension measure? British journal of educational psychology, 75(2):189--201

  7. [7]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

  8. [8]

    Kate Cain. 2022. Children's reading comprehension difficulties. The science of reading: A handbook, pages 298--322

Show all 51 references
  1. [9]

    Kate Cain and Jane V Oakhill. 1999. Inference making ability and its relation to comprehension failure in young children. Reading and writing, 11:489--503

  2. [10]

    Kate Cain, Jane V Oakhill, Marcia A Barnes, and Peter E Bryant. 2001. Comprehension skill, inference-making ability, and their relation to knowledge. Memory & cognition, 29(6):850--859

  3. [11]

    Michael Flor and Brian Riordan. 2018. A semantic role-based approach to open-domain automatic question generation. In Proceedings of the thirteenth workshop on innovative use of NLP for building educational applications, pages 254--263

  4. [12]

    Bilal Ghanem, Lauren Lutz Coleman, Julia Rivard Dexter, Spencer von der Ohe, and Alona Fyshe. 2022. https://doi.org/10.18653/v1/2022.findings-acl.168 Question generation for reading comprehension assessment by modeling how and what to ask . In Findings of the Association for C...

  5. [13]

    Arthur C Graesser, Murray Singer, and Tom Trabasso. 1994. Constructing inferences during narrative text comprehension. Psychological review, 101(3):371

  6. [14]

    Susan E Haviland and Herbert H Clark. 1974. What's new? acquiring new information as a process in comprehension. Journal of verbal learning and verbal behavior, 13(5):512--521

  7. [15]

    Yufang Hou. 2020. https://doi.org/10.18653/v1/2020.acl-main.132 Bridging anaphora resolution as question answering . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1428--1438, Online. Association for Computational Linguistics

  8. [16]

    Yufang Hou, Katja Markert, and Michael Strube. 2018. Unrestricted bridging resolution. Computational Linguistics, 44(2):237--284

  9. [17]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

  10. [18]

    Kevin Hwang, Kenneth Wang, Maryam Alomair, Fow-Sen Choa, and Lujie Karen Chen. 2024. Towards automated multiple choice question generation and evaluation: aligning with bloom’s taxonomy. In International Conference on Artificial Intelligence in Education, pages 389--396. Springer

  11. [19]

    Walter Kintsch. 1998. Comprehension: A paradigm for cognition. Cambridge university press

  12. [20]

    Hideo Kobayashi, Yufang Hou, and Vincent Ng. 2022. https://aclanthology.org/2022.coling-1.64/ End-to-end neural bridging resolution . In Proceedings of the 29th International Conference on Computational Linguistics, pages 766--778, Gyeongju, Republic of Korea. International Co...

  13. [21]

    Saurabh Kulshreshtha and Anna Rumshisky. 2022. Reasoning circuits: Few-shot multihop question generation with structured rationales. arXiv preprint arXiv:2211.08466

  14. [22]

    Ghader Kurdi, Jared Leo, Bijan Parsia, Uli Sattler, and Salam Al-Emari. 2020. A systematic review of automatic question generation for educational purposes. International Journal of Artificial Intelligence in Education, 30:121--204

  15. [23]

    Kunze Li and Yu Zhang. 2024. Planning first, question second: An llm-guided method for controllable question generation. In Findings of the Association for Computational Linguistics ACL 2024, pages 4715--4729

  16. [24]

    Yunting Liu, Shreya Bhandari, and Zachary A Pardos. 2025. Leveraging llm respondents for item evaluation: A psychometric analysis. British Journal of Educational Technology, 56(3):1028--1052

  17. [25]

    Xinyi Lu and Xu Wang. 2024. Generative students: Using llm-simulated student profiles to support question item evaluation. In Proceedings of the Eleventh ACM Conference on Learning@ Scale, pages 16--27

  18. [26]

    Kawshik Manikantan, Makarand Tapaswi, Vineet Gandhi, and Shubham Toshniwal. 2024. https://arxiv.org/abs/2411.07466 Identifyme: A challenging long-context mention resolution benchmark . Preprint, arXiv:2411.07466

  19. [27]

    Vaibhav Mavi, Anubhav Jangra, and Adam Jatowt. 2024. https://arxiv.org/abs/2204.09140 Multi-hop question answering . Preprint, arXiv:2204.09140

  20. [28]

    Nikahat Mulla and Prachi Gharpure. 2023. Automatic question generation: a review ofmethodologies, datasets, evaluation metrics, and applications. Progress in Artificial Intelligence , 12:1--32

  21. [29]

    Edward J O'Brien, Anne E Cook, and Robert F Lorch. 2015. Inferences during reading. Cambridge University Press

  22. [30]

    Onkar Pandit and Yufang Hou. 2021. https://doi.org/10.18653/v1/2021.naacl-main.327 Probing for bridging inference in transformer language models . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langua...

  23. [31]

    E Margaret Perkoff, Abhidip Bhattacharyya, Jon Cai, and Jie Cao. 2023. Comparing neural question generation architectures for reading comprehension. In Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023), pages 556--566

  24. [32]

    Yin Poon, John Sie Yuen Lee, Yu Yan Lam, Wing Lam Suen, Elsie Li Chen Ong, and Samuel Kai Wah Chu. 2024. https://aclanthology.org/2024.sighan-1.3/ Few-shot question generation for reading comprehension . In Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing...

  25. [33]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://doi.org/10.18653/v1/D16-1264 SQ u AD : 100,000+ questions for machine comprehension of text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-...

  26. [34]

    Manav Rathod, Tony Tu, and Katherine Stasaski. 2022. https://doi.org/10.18653/v1/2022.bea-1.26 Educational multi-question generation for reading comprehension . In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages...

  27. [35]

    Ina R \"o siger. 2018. https://aclanthology.org/L18-1058/ BASHI : A corpus of W all S treet J ournal articles annotated with bridging links . In Proceedings of the Eleventh International Conference on Language Resources and Evaluation ( LREC 2018) , Miyazaki, Japan. European L...

  28. [36]

    John Sabatini, Jonathan Weeks, Tenaha O'Reilly, Kelly Bruce, Jonathan Steinberg, and Szu-Fu Chao. 2019. SARA Reading Components Tests, RISE forms: Technical Adequacy and Test Design . ETS Research Report Series, 2019(1):1--30

  29. [37]

    Andreas S \"a uberli and Simon Clematide. 2024. https://aclanthology.org/2024.readi-1.3/ Automatic generation and evaluation of reading comprehension test items with large language models . In Proceedings of the 3rd Workshop on Tools and Resources for People with REAding DIffi...

  30. [38]

    Franz Schmalhofer, Mark A McDaniel, and Dennis Keefe. 2002. A unified model for predictive and bridging inferences. Discourse Processes, 33(2):105--132

  31. [39]

    Murray Singer, Peter Andruslak, Paul Reisdorf, and Nancy L Black. 1992. Individual differences in bridging inference processes. Memory & cognition, 20(5):539--548

  32. [40]

    Murray Singer and Gilbert Remillard. 2004. Retrieving text inferences: Controlled and automatic influences. Memory & Cognition, 32(8):1223--1237

  33. [41]

    Katherine Stasaski, Manav Rathod, Tony Tu, Yunfang Xiao, and Marti A Hearst. 2021. Automatically generating cause-and-effect questions from passages. In Proceedings of the 16th Workshop on Innovative Use of NLP for Building Educational Applications, pages 158--170

  34. [42]

    Richard Thurlow and Paul van den Broek. 1997. Automaticity and inference generation during reading comprehension. Reading & Writing Quarterly: Overcoming Learning Difficulties, 13(2):165--181

  35. [43]

    Masaki Uto, Yuto Tomikawa, and Ayaka Suzuki. 2023. https://doi.org/10.18653/v1/2023.bea-1.10 Difficulty-controllable neural question generation for reading comprehension using item response theory . In Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educ...

  36. [44]

    Paul van den Broek, Katinka Beker, and Marja Oudega. 2015. Inference generation in text comprehension: Automatic and strategic processes in the construction of a mental representation . In Edward J. O'Brien, Anne E. Cook, and Robert F. Lorch Jr., editors, Inferences during rea...

  37. [45]

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

  38. [46]

    Ying Xu, Dakuo Wang, Mo Yu, Daniel Ritchie, Bingsheng Yao, Tongshuang Wu, Zheng Zhang, Toby Jia-Jun Li, Nora Bradford, Branda Sun, et al. 2022. Fantastic questions and where to find them: Fairytaleqa--an authentic dataset for narrative comprehension. arXiv preprint arXiv:2203.13947

  39. [47]

    Jason D Yeatman, Jasmine E Tran, Amy K Burkhardt, Wanjing Anya Ma, Jamie L Mitchell, Maya Yablonski, Liesbeth Gijbels, Carrie Townley-Flores, and Adam Richie-Halford. 2024. Development and validation of a rapid and precise online sentence reading efficiency assessment. In Fron...

  40. [48]

    Eric Zelikman, Wanjing Ma, Jasmine Tran, Diyi Yang, Jason Yeatman, and Nick Haber. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.135 Generating and evaluating tests for k-12 students with language model simulations: A case study on sentence reading efficiency . In Proceedi...

  41. [49]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476--15488

  42. [50]

    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...

  43. [51]

    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 7, 2026 · model on record in the stance chip above.