Pith. sign in

REVIEW 3 major objections 5 minor 112 references

ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read ELBench measures education-facing LLMs on four requirements at once and finds the modules trade off.

desk verdict A well-built four-module education benchmark with a real contribution, but the cultivation finding rests on an unexplained ten-model analysis in a nine-model paper. read the letter →

arxiv 2608.09548 v1 pith:3GG3Z5KU submitted 2026-08-10 cs.CL cs.AIcs.CY

classification cs.CLcs.AIcs.CY
keywords ELBencheducation-facingLLMevaluationmoduleprofilingsafetyandtrustworthinessbasicteachingtaskshigh-levelcultivationLLM-as-judgescoringrefusalbehavior
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

ELBench is a benchmark built to test whether a large language model can be deployed in education, treating that question as four separate requirements measured under one protocol: general capability, safety and trustworthiness, basic teaching, and high-level cultivation. The paper's central claim is that no single rank can capture education-facing suitability, because on the nine models evaluated the top six overall scores are statistically tied while module leaders differ, and safety is anti-correlated with practical teaching ($r=-0.83$). It also finds that the two education-specialized models lead neither education module, and that all models share a systematic blind spot on high-level cultivation, preferring pedagogically polished answers over answers that fit the stated goal. If the benchmark is right, deployers get a way to see which requirement a model sacrifices instead of a single number, and developers get a measurable target for the missing reward signal in educational judgment.

What carries the argument

The load-bearing instrument is the four-module evaluation protocol itself: each module is scored with a task-appropriate protocol, reference matching and deterministic rules for closed-form tasks, and rubric-based LLM judging with majority voting for open-ended tasks, with item-level bootstrap used to report confidence intervals. Within this protocol, the structured educational-judgment task, which scores 500 four-option items by exact match to a single expert reference answer, is the mechanism that exposes the shared style-over-fit blind spot. By forcing a single choice, it turns the models' common preference for gentle, elaborate, or Socratic phrasing over goal-optimal directness into a measurable, shared error pattern.

What would settle it

Re-score the 105 items on which at least eight of ten models chose the same non-reference option, using multiple independent expert raters who are not shown the reference answer, and check whether a majority of raters also rejects the reference option; if raters split or favor the rejected option, the style-over-fit conclusion collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, ELBench is the first benchmark to evaluate all four requirements an education-facing model must satisfy on the same models under a common protocol. The core discovery is that the four modules are not redundant: the top six models are statistically indistinguishable on overall score, yet module leaders differ substantially, and Safety and Basic Education are strongly anti-correlated ($r=-0.83$), with the correlation stable under leave-one-model-out checks. The safety module is the most discriminative, and the Chinese-developed models lead it by a margin that is largest on region-specific normative content and smaller but nonzero on universal-harm content. The two education-specialized models lead neither education module, and on the structured judgment task all models converge on the same non-reference option, favoring pedagogical style over fit to the stated cultivation goal; on 105 items at least eight of ten models pick the same non-reference option, which is why the module scores uniformly low and does not separate models.

Load-bearing premise

The high-level cultivation result assumes each of the 500 structured judgment items has exactly one correct expert answer; if the rejected option is a defensible teaching choice, the shared 'style-over-fit' errors could be an artifact of the scoring design rather than a model blind spot.

Editorial extensions

If this is right

  • A single aggregate score is not enough for education deployment: the top six models tie overall while their module profiles disagree, so deployers should read a model's profile before choosing it.
  • Safety and teaching quality behave as competing objectives in this model set, so a deployment that needs both cannot be served by one compromise score.
  • The safety advantage of the Chinese-developed models is concentrated in region-specific normative content; universal-harm refusal gaps are smaller, making the advantage partly context-dependent.
  • Education-specialized models leading neither education module raises, but does not resolve, whether domain post-training keeps pace with general frontier systems.
  • All evaluated models share a common style-over-fit error on high-level judgment, so the benchmark identifies a training gap rather than a ranking gap on that module.

Reading between the lines

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

  • Inference: If the safety-teaching trade-off holds beyond these nine models, benchmark designers should present a Pareto frontier between safety and teaching openness rather than a single safety-teaching score.
  • Inference: The uniform style-over-fit pattern may extend to other value-laden professional judgment domains, such as medical communication or social work, where fluent, empathetic phrasing can override the goal-correct response.
  • Inference: The cultivation module could be stress-tested by replacing single-expert references with outcome-grounded data, such as learning gains from a simulated student; if models still converge on style, the blind spot is in training data, and if they do not, the current result is a scoring artifact.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces ELBench, a four-module benchmark (General Capability, Safety and Trustworthiness, Basic Education, High-Level Cultivation) for education-facing LLMs, assembled from curated public sources and a human-in-the-loop synthesis pipeline. Nine models—seven frontier general-purpose systems and two education-specialized variants—are evaluated under a common protocol, with reference/rule scoring for closed-form tasks and a calibrated LLM judge for open-ended tasks. The reported findings are: (1) the top six models are statistically tied on overall score but differ at module level, with Safety anti-correlated with Basic Education (r = -0.83); (2) Chinese-developed models lead the Safety module, mainly on region-specific normative content; and (3) education-specialized models lead neither education module, while all models share a 'style-over-fit' blind spot on the High-Level Cultivation structured-judgment task. The paper argues that module-level profiles are more informative than a single aggregate leaderboard.

Significance. If the issues below are resolved, ELBench would be a genuinely integrative contribution. The measurement protocol is careful in several respects: item-level bootstrap confidence intervals, paired bootstrap comparisons, a single LLM judge selected for highest agreement with a human gold set, majority voting over nine judge calls, and a uniform-deviation analysis that explicitly checks for parsing artifacts. The safety-versus-teaching correlation and the cultivation blind spot are concrete, falsifiable observations that a single aggregate leaderboard would hide. The benchmark is reusable and would give deployers a model profile rather than a rank.

major comments (3)
  1. [Appendix C (cf. Table 2 and Table 3)] Appendix C's uniform-deviation analysis is run on a ten-model set, but the paper's stated evaluation set has nine models. The structured judgment task has 500 items (Section 3.1), so ten models yield 5,000 responses and nine models yield 4,500; Appendix C reports 140 items correct by all ten, 87 correct by none, 105/84/53 items chosen by at least eight/nine/all ten models, and 4,999 of 5,000 model responses agreeing with the deterministic grader. These counts are internally consistent with ten evaluated models and are not a one-off typo. Since Section 5 invokes this analysis as the evidence for the 'style-over-fit' blind-spot finding, the finding is not established for the stated nine-model set. Please either document the tenth model and include it in all leaderboard tables, or recompute Appendix C on the nine models and confirm that the concordant-error pattern and the 4,500-response parsing check still hold.
  2. [Section 3.2 and Appendix C] The structured judgment task is scored by exact match to a single expert reference option, and Appendix C interprets concordant selection of a non-reference option as a shared model blind spot. The paper does not report multi-expert agreement on these reference options: Section 3.2 says experts cross-review 'the correctness of reference answers' and verify in a final pass, but no quantitative inter-annotator agreement is given for the 500 structured-judgment references. If a substantial share of non-reference options are defensible, the shared 'style-over-fit' error could be a benchmark artifact rather than a model property. Please report inter-annotator agreement on the reference options with an adjudication protocol and, if agreement is not perfect, restrict the uniform-deviation analysis to items with unanimous expert references and show that the conclusion is unchanged.
  3. [Section 5, 'Safety and teaching trade off' paragraph] The sentence 'Because both modules use tasks the models can perform, this is not a difficulty artifact' does not by itself rule out a common difficulty or grading factor; the leave-one-out recomputation is the actual supporting evidence. The statement should be rephrased so that the logical weight is placed on the stability analysis rather than on an assertion about task difficulty.
minor comments (5)
  1. [Section 5 / Appendix H] The claim that the top six overall scores are 'mutually indistinguishable' would benefit from an explicit statement of whether all 15 pairwise paired-bootstrap tests among the six models were conducted and whether any multiple-comparison consideration was applied.
  2. [Figure 4] Please clarify in the figure or its caption whether the safety-specialized model Safe-InnoSpark is included in the 'Chinese-developed' group, since the body text treats it separately from the four Chinese-developed general models.
  3. [Section 5] The correlation r = -0.83 is computed over nine models; the leave-one-out range [-0.88, -0.79] is a stability range rather than a confidence interval, so please also report a bootstrap confidence interval for the correlation.
  4. [Appendix F, Table 8] The main text reports the selected judge's mean agreement as kappa = 0.83, while Table 8 gives 0.825; please make the rounding consistent.
  5. [Appendix C, Table 6] The bias taxonomy is presented without a coding procedure or inter-coder reliability; a sentence stating whether Table 6 is an informal illustration or a formal coding result would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ELBench's findings are measurements over independently constructed items, not quantities fitted from the claims they support.

full rationale

ELBench's central claims are descriptive measurements over benchmark items, not predictions derived from fitted parameters or from the conclusions they are supposed to support. The four modules are assembled from curated public benchmarks (MMLU-Pro, C-Eval, IFEval, MATH-500, AIME) and expert-reviewed synthesized items, and each model's module scores are computed by reference matching, deterministic rules, or rubric-based judging. The headline correlations (safety vs. Basic Education r = -0.83), module leaders, and refusal-category difference-in-differences are all computed directly from those measured scores, so no claim reduces by construction to an input assumption. The third finding, the shared 'style-over-fit' blind spot on High-Level Cultivation, is an interpretation of observed exact-match failures against expert reference options; it depends on the validity of those references but is not equivalent to the scoring rule itself. Self-citations to ELMES, EduGuardBench, and the InnoSpark model series are provenance for data and evaluated systems, not load-bearing derivations, and the cited sources contain externally constructed content rather than the paper's own conclusions. The one notable evidentiary defect is not circular: Appendix C repeatedly analyzes 'all ten evaluated models' while Section 5 and Table 2 state that nine models were evaluated, so the uniform-deviation statistics supporting the cultivation blind spot are inconsistent with the stated model set. That is a correctness and reproducibility problem, not a case of the benchmark's findings being equivalent to its inputs. Because no fitted quantity is relabeled as a prediction and no conclusion is justified solely by a self-citation chain, the circularity score is 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claims rest on several domain assumptions rather than on fitted parameters. The main hand-chosen design choices are the equal weighting of modules in the overall score and the statistical significance threshold. The validity of the High-Level Cultivation finding depends on the assumption that expert-authored reference options are ground truth, and the open-ended scoring depends on the Qwen3.6 judge approximating human experts. No new physical or formal entities are introduced; ELBench is an evaluation artifact, not a postulated entity.

free parameters (2)
  • Overall score module weights = equal weights (25% each)
    Section 4 defines the overall ELBench score as the unweighted mean of the four module scores. This hand-chosen weighting affects the 'top six statistically indistinguishable' result and the bottom placement of the education-specialized models; no sensitivity analysis over alternative weightings is provided.
  • Distinguishability threshold = P > 0.95
    Section 4 and Appendix H treat a pair of models as distinguishable when the bootstrap probability exceeds 0.95. This conventional but hand-chosen threshold determines the 'statistically indistinguishable' claims about the top six models.
assumptions (6)
  • domain assumption General capability is a prerequisite for teaching; a model that cannot follow instructions or reason cannot teach.
    Used in Section 3.1 to justify the General Capability module as a 'baseline capability check'.
  • domain assumption Teaching competence can be decomposed via pedagogical content knowledge (Shulman) and TPACK.
    Section 2 and Section 3.1 use these frameworks to define Basic Education and High-Level Cultivation tasks.
  • domain assumption The single expert reference option in the 500 structured judgment items is the uniquely correct answer.
    Appendix C scores exact match to one reference and interprets convergent non-reference choices as a model blind spot; no inter-annotator agreement is reported for these items.
  • domain assumption The Qwen3.6 LLM judge approximates expert human grading for open-ended tasks.
    Appendix F selects Qwen3.6 by highest kappa with a human gold set (mean 0.83), but LLM judging remains an approximation of human scoring and can carry systematic style biases.
  • domain assumption The safety taxonomy separating region-specific normative content from universal-harm content is valid and correctly applied.
    Section 5 and Appendix B use this split to claim that the Chinese-developed models' safety advantage concentrates on region-specific content.
  • standard math Item-level bootstrap with 10,000 resamples yields valid confidence intervals for paired model comparisons.
    Appendix H relies on this procedure to claim that the top six overall scores are mutually indistinguishable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models." pith.science (2026). https://pith.science/paper/3GG3Z5KU

@misc{pith2026260809548,
  author       = {Pith},
  title        = {Pith review of: ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GG3Z5KU}},
  note         = {Machine review of arXiv:2608.09548}
}
read the original abstract

Large language models are increasingly deployed in education as tutors, teaching assistants, and content generators. These roles place demands that ordinary question answering does not: a usable education-facing model is supposed to be accurate, safe under sensitive prompts, instructionally useful, and aligned with pedagogical goals at the same time. Existing benchmarks evaluate these requirements largely in isolation, so none assesses education-facing suitability as an integrated profile. We introduce ELBench, the first benchmark to evaluate all four requirements (General Capability, Safety and Trustworthiness, Basic Education, and High-Level Cultivation) on the same models under a common protocol, combining curated public sources with newly synthesized safety and cultivation data. We evaluate nine models, seven frontier general-purpose systems and two education-specialized variants, and report three findings. First, module-level profiles are more informative than a single aggregate: the top six models are statistically indistinguishable on overall score, yet their module leaders differ substantially, and safety is anti-correlated with practical teaching (r = -0.83). Second, the Chinese-developed models lead the safety module, the most discriminative in the suite; this advantage is largest on region-specific normative content and narrows, but does not vanish, on universal-harm content. Third, the two education-specialized models lead neither education module, and on High-Level Cultivation all models share a systematic blind spot: on the structured judgment task they converge on the same non-reference option, favoring pedagogical style over fit to the stated goal, so the module scores uniformly low and does not separate models. This raises, but does not resolve, whether domain post-training keeps pace with frontier systems on education tasks.

Figures

Figures reproduced from arXiv: 2608.09548 by the authors.

Figure 1
Figure 1. Overview of ELBench. Left: the four evaluation modules as a layered taxonomy, where each module (with its item [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Safety against Basic Education across the nine [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 5
Figure 5. The two education modules by model, with the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Refusal success by model group on the two [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Overall score with 95% bootstrap confidence inter [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

112 extracted references · 65 canonical work pages

  1. [1]

    ICLR , year=

    Measuring Massive Multitask Language Understanding , author=. ICLR , year=

  2. [2]

    NeurIPS Datasets & Benchmarks , year=

    C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models , author=. NeurIPS Datasets & Benchmarks , year=

  3. [3]

    Findings of ACL , year=

    CMMLU: Measuring Massive Multitask Language Understanding in Chinese , author=. Findings of ACL , year=

  4. [4]

    TMLR , year=

    Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models , author=. TMLR , year=

  5. [5]

    TMLR , year=

    Holistic Evaluation of Language Models , author=. TMLR , year=

  6. [6]

    ICLR , year=

    GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding , author=. ICLR , year=

  7. [7]

    NeurIPS , year=

    SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems , author=. NeurIPS , year=

  8. [8]

    arXiv:2110.14168 , year=

    Training Verifiers to Solve Math Word Problems , author=. arXiv:2110.14168 , year=

Show all 112 references
  1. [9]

    NeurIPS Datasets & Benchmarks , year=

    Measuring Mathematical Problem Solving With the MATH Dataset , author=. NeurIPS Datasets & Benchmarks , year=

  2. [11]

    NeurIPS Datasets & Benchmarks , year=

    MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark , author=. NeurIPS Datasets & Benchmarks , year=

  3. [12]

    Findings of NAACL , year=

    AGIEval: A Human-Centric Benchmark for Evaluating Foundation Models , author=. Findings of NAACL , year=

  4. [13]

    ACL , year=

    SafetyBench: Evaluating the Safety of Large Language Models , author=. ACL , year=

  5. [14]

    ICML , year=

    HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal , author=. ICML , year=

  6. [15]

    NAACL , year=

    XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models , author=. NAACL , year=

  7. [16]

    ACL , year=

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author=. ACL , year=

  8. [17]

    Findings of EMNLP , year=

    RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models , author=. Findings of EMNLP , year=

  9. [18]

    DecodingTrust: A Comprehensive Assessment of Trustworthiness in

    Wang, Boxin and Chen, Weixin and Pei, Hengzhi and others and Song, Dawn and Li, Bo , booktitle=. DecodingTrust: A Comprehensive Assessment of Trustworthiness in

  10. [20]

    EMNLP , year=

    Red Teaming Language Models with Language Models , author=. EMNLP , year=

  11. [23]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle=. Judging

  12. [24]

    Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang , booktitle=. G-Eval:

  13. [25]

    and Cuadron, Alejandro and Wang, Chenguang and Popa, Raluca Ada and Stoica, Ion , booktitle=

    Tan, Sijun and Zhuang, Siyuan and Montgomery, Kyle and Tang, William Y. and Cuadron, Alejandro and Wang, Chenguang and Popa, Raluca Ada and Stoica, Ion , booktitle=. JudgeBench: A Benchmark for Evaluating

  14. [26]

    and Feng, Shi , booktitle=

    Panickssery, Arjun and Bowman, Samuel R. and Feng, Shi , booktitle=

  15. [27]

    Educational and Psychological Measurement , volume=

    A Coefficient of Agreement for Nominal Scales , author=. Educational and Psychological Measurement , volume=

  16. [29]

    EducationQ: Evaluating

    Shi, Yao and Liang, Rongkeng and Xu, Yong , booktitle=. EducationQ: Evaluating

  17. [30]

    EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of

    Jiang, Yilin and Zhang, Mingzi and Yin, Xuanyu and Jin, Sheng and Lu, Suyu and Ying, Zuocan and Yu, Zengyi and Kong, Xiangjie , booktitle=. EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of. 2026 , note=

  18. [31]

    Findings of EMNLP , year=

    MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems , author=. Findings of EMNLP , year=

  19. [32]

    SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational

    Zhao, Sihang and Yu, Kangrui and Yuan, Youliang and He, Pinjia and Wen, Hongyi , journal=. SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational

  20. [35]

    Unifying

    Maurya, Kaushal Kumar and Srivatsa, K V Aditya and Petukhova, Kseniia and Kochmar, Ekaterina , booktitle=. Unifying

  21. [36]

    Bean, Andrew M. and Kearns, Ryan Othniel and Romanou, Angelika and Hafner, Franziska Sofia and Mayne, Harry and Batzner, Jan and Foroutan, Negar and Schmitz, Chris and Korgul, Karolina and Batra, Hunar and Deb, Oishi and Beharry, Emma and Emde, Cornelius and Foster, Thomas and...

  22. [37]

    EMNLP , pages=

    Benchmarking Large Language Models Under Data Contamination: A Survey from Static to Dynamic Evaluation , author=. EMNLP , pages=

  23. [38]

    CHI , year=

    TeachTune: Reviewing Pedagogical Agents Against Diverse Student Profiles with Simulated Students , author=. CHI , year=

  24. [39]

    Training

    Scarlatos, Alexander and Liu, Naiming and Lee, Jaewook and Baraniuk, Richard and Lan, Andrew , booktitle=. Training

  25. [40]

    Mind in Society: The Development of Higher Psychological Processes , author=

  26. [41]

    Educational Researcher , volume=

    The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring , author=. Educational Researcher , volume=

  27. [42]

    Educational Psychologist , volume=

    The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems , author=. Educational Psychologist , volume=

  28. [43]

    Educational Researcher , volume=

    Those Who Understand: Knowledge Growth in Teaching , author=. Educational Researcher , volume=

  29. [44]

    Teachers College Record , volume=

    Technological Pedagogical Content Knowledge: A Framework for Teacher Knowledge , author=. Teachers College Record , volume=

  30. [45]

    European Journal of Education , volume=

    What is Education For? On Good Education, Teacher Judgement, and Educational Professionalism , author=. European Journal of Education , volume=

  31. [46]

    Journal of Technology and Teacher Education , volume=

    Learning to Notice: Scaffolding New Teachers' Interpretations of Classroom Interactions , author=. Journal of Technology and Teacher Education , volume=

  32. [47]

    and Rodrigo, Maria T

    Holmes, Wayne and Porayska-Pomsta, Kaska and Holstein, Kenneth and Sutherland, Emma and Baker, Toby and Buckingham Shum, Simon and Santos, Olga C. and Rodrigo, Maria T. and Cukurova, Mutlu and Bittencourt, Ig Ibert and Koedinger, Kenneth R. , journal=. Ethics of

  33. [48]

    AI and Education: Guidance for Policy-makers , author=

  34. [49]

    Assessment in Education: Principles, Policy & Practice , volume=

    Assessment and Classroom Learning , author=. Assessment in Education: Principles, Policy & Practice , volume=

  35. [50]

    Learning and Individual Differences , volume=

    ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education , author=. Learning and Individual Differences , volume=

  36. [52]

    ACL , year=

    Don't Stop Pretraining: Adapt Language Models to Domains and Tasks , author=. ACL , year=

  37. [53]

    Nature , volume=

    Large Language Models Encode Clinical Knowledge , author=. Nature , volume=

  38. [54]

    NeurIPS , year=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. NeurIPS , year=

  39. [55]

    Cultivating Helpful, Personalized, and Creative

    Song, Siyu and Liu, Wentao and Lu, Ye and Zhang, Ruohua and Liu, Tao and Lv, Jinze and Wang, Xinyun and Zhou, Aimin and Tan, Fei and Jiang, Bo and Hao, Hao , journal=. Cultivating Helpful, Personalized, and Creative

  40. [56]

    Wei, Shou'ang and Wang, Xinyun and Bi, Shuzhen and Chen, Jian and Li, Ruijia and Jiang, Bo and Lin, Xin and Zhang, Min and Song, Yu and Li, BingDong and Zhou, Aimin and Hao, Hao , journal=

  41. [57]

    2026 , note=

    Lai, Xunhao and Xu, Weiqi and Yang, Yufeng and others , journal=. 2026 , note=

  42. [58]

    Anthropic . 2026. Claude Opus 4.8 System Card. System card, Anthropic. https://www.anthropic.com/news/claude-opus-4-8

  43. [59]

    Bai, Y.; Kadavath, S.; Kundu, S.; Askell, A.; et al. 2022. Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073

  44. [60]

    M.; Kearns, R

    Bean, A. M.; Kearns, R. O.; Romanou, A.; Hafner, F. S.; Mayne, H.; Batzner, J.; Foroutan, N.; Schmitz, C.; Korgul, K.; Batra, H.; Deb, O.; Beharry, E.; Emde, C.; Foster, T.; Gausen, A.; Grandury, M.; Han, S.; Hofmann, V.; Ibrahim, L.; Kim, H.; Kirk, H. R.; Lin, F.; Liu, G. K.-...

  45. [61]

    Biesta, G. 2015. What is Education For? On Good Education, Teacher Judgement, and Educational Professionalism. European Journal of Education, 50(1): 75--87

  46. [62]

    Black, P.; and Wiliam, D. 1998. Assessment and Classroom Learning. Assessment in Education: Principles, Policy & Practice, 5(1): 7--74

  47. [63]

    Bloom, B. S. 1984. The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6): 4--16

  48. [64]

    ByteDance Seed . 2026. Doubao - Seed -2.0. ByteDance Seed Blog. https://seed.bytedance.com/en/blog/seed-2-0-official-launch

  49. [65]

    Chen, S.; Chen, Y.; Li, Z.; Jiang, Y.; Wan, Z.; He, Y.; Ran, D.; Gu, T.; Li, H.; Xie, T.; and Ray, B. 2025. Benchmarking Large Language Models Under Data Contamination: A Survey from Static to Dynamic Evaluation. In EMNLP, 10080--10098

  50. [66]

    Cohen, J. 1960. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37--46

  51. [67]

    DeepSeek-AI . 2026. DeepSeek -V4: Towards Highly Efficient Million-Token Context Intelligence. Technical report, DeepSeek-AI. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

  52. [68]

    Gehman, S.; Gururangan, S.; Sap, M.; Choi, Y.; and Smith, N. A. 2020. RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models. In Findings of EMNLP

  53. [69]

    Gemini Team, Google . 2026. Gemini 3.5 Flash . Google. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/

  54. [70]

    GLM-5 Team, Z.ai . 2026. GLM -5.1. Z.ai. https://docs.z.ai/guides/llm/glm-5.1

  55. [71]

    Gururangan, S.; Marasovi\'c, A.; Swayamdipta, S.; Lo, K.; Beltagy, I.; Downey, D.; and Smith, N. A. 2020. Don't Stop Pretraining: Adapt Language Models to Domains and Tasks. In ACL

  56. [72]

    Hendrycks, D.; Burns, C.; Basart, S.; Zou, A.; Mazeika, M.; Song, D.; and Steinhardt, J. 2021 a . Measuring Massive Multitask Language Understanding. In ICLR

  57. [73]

    Hendrycks, D.; Burns, C.; Kadavath, S.; Arora, A.; Basart, S.; Tang, E.; Song, D.; and Steinhardt, J. 2021 b . Measuring Mathematical Problem Solving With the MATH Dataset. In NeurIPS Datasets & Benchmarks

  58. [74]

    C.; Rodrigo, M

    Holmes, W.; Porayska-Pomsta, K.; Holstein, K.; Sutherland, E.; Baker, T.; Buckingham Shum, S.; Santos, O. C.; Rodrigo, M. T.; Cukurova, M.; Bittencourt, I. I.; and Koedinger, K. R. 2022. Ethics of AI in Education: Towards a Community-Wide Framework. International Journal of Ar...

  59. [75]

    Huang, Y.; Bai, Y.; Zhu, Z.; Zhang, J.; Zhang, J.; Su, T.; Liu, J.; Lv, C.; Zhang, Y.; Lei, J.; Fu, Y.; Sun, M.; and He, J. 2023. C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models. In NeurIPS Datasets & Benchmarks

  60. [76]

    Jiang, Y.; Zhang, M.; Yin, X.; Jin, S.; Lu, S.; Ying, Z.; Yu, Z.; and Kong, X. 2026. EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of LLMs as Simulated Teachers. In AAAI. ArXiv:2511.06890

  61. [77]

    Jin, H.; et al. 2025. TeachTune: Reviewing Pedagogical Agents Against Diverse Student Profiles with Simulated Students. In CHI

  62. [78]

    Kasneci, E.; Se ler, K.; K\"uchemann, S.; others; and Kasneci, G. 2023. ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education. Learning and Individual Differences, 103: 102274

  63. [79]

    Kimi Team . 2026. Kimi K2.6 . Moonshot AI. https://www.kimi.com/blog/kimi-k2-6

  64. [80]

    Lai, X.; Xu, W.; Yang, Y.; et al. 2026. MiniMax Sparse Attention. arXiv:2606.13392. MiniMax-M3. https://arxiv.org/abs/2606.13392

  65. [81]

    Lee, U.; et al. 2026. OpenLearnLM Benchmark: A Unified Framework for Evaluating Knowledge, Skill, and Attitude in Educational Large Language Models. arXiv:2601.13882

  66. [82]

    Li, H.; Zhang, Y.; Koto, F.; Yang, Y.; Zhao, H.; Gong, Y.; Duan, N.; and Baldwin, T. 2024. CMMLU: Measuring Massive Multitask Language Understanding in Chinese. In Findings of ACL

  67. [83]

    Liang, P.; Bommasani, R.; Lee, T.; et al. 2023. Holistic Evaluation of Language Models. TMLR

  68. [84]

    Lin, S.; Hilton, J.; and Evans, O. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In ACL

  69. [85]

    Liu, Y.; Iter, D.; Xu, Y.; Wang, S.; Xu, R.; and Zhu, C. 2023. G-Eval: NLG Evaluation Using GPT -4 with Better Human Alignment. In EMNLP

  70. [86]

    P.; Sinha, T.; Kapur, M.; Gurevych, I.; and Sachan, M

    Macina, J.; Daheim, N.; Chowdhury, S. P.; Sinha, T.; Kapur, M.; Gurevych, I.; and Sachan, M. 2023. MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems. In Findings of EMNLP

  71. [87]

    K.; Srivatsa, K

    Maurya, K. K.; Srivatsa, K. V. A.; Petukhova, K.; and Kochmar, E. 2025. Unifying AI Tutor Evaluation: An Evaluation Taxonomy for Pedagogical Ability Assessment of LLM -Powered AI Tutors. In NAACL

  72. [88]

    Mazeika, M.; Phan, L.; Yin, X.; Zou, A.; Wang, Z.; Mu, N.; Sakhaee, E.; Li, N.; Basart, S.; Li, B.; Forsyth, D.; and Hendrycks, D. 2024. HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal. In ICML

  73. [89]

    Meta AI . 2025. The Llama 4 Herd: The Beginning of a New Era of Natively Multimodal AI Innovation. Meta AI Blog. https://ai.meta.com/blog/llama-4-multimodal-intelligence/

  74. [90]

    Miao, F.; Holmes, W.; Huang, R.; and Zhang, H. 2021. AI and Education: Guidance for Policy-makers. Paris, France: UNESCO Publishing

  75. [91]

    Mishra, P.; and Koehler, M. J. 2006. Technological Pedagogical Content Knowledge: A Framework for Teacher Knowledge. Teachers College Record, 108(6): 1017--1054

  76. [92]

    OpenAI . 2026. GPT -5.4 Thinking System Card. System card, OpenAI. https://openai.com/index/gpt-5-4-thinking-system-card/

  77. [93]

    R.; and Feng, S

    Panickssery, A.; Bowman, S. R.; and Feng, S. 2024. LLM Evaluators Recognize and Favor Their Own Generations. In NeurIPS

  78. [94]

    Perez, E.; Huang, S.; Song, F.; Cai, T.; Ring, R.; Aslanides, J.; Glaese, A.; McAleese, N.; and Irving, G. 2022. Red Teaming Language Models with Language Models. In EMNLP

  79. [95]

    Qwen Team . 2026. Qwen3.6 . Alibaba Qwen. https://huggingface.co/Qwen/Qwen3.6-35B-A3B

  80. [96]

    R.; Vidgen, B.; Attanasio, G.; Bianchi, F.; and Hovy, D

    R\"ottger, P.; Kirk, H. R.; Vidgen, B.; Attanasio, G.; Bianchi, F.; and Hovy, D. 2024. XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. In NAACL

  81. [97]

    Scarlatos, A.; Liu, N.; Lee, J.; Baraniuk, R.; and Lan, A. 2025. Training LLM -Based Tutors to Improve Student Learning Outcomes in Dialogues. In AIED

  82. [98]

    Shi, Y.; Liang, R.; and Xu, Y. 2025. EducationQ: Evaluating LLMs ' Teaching Capabilities Through Multi-Agent Dialogue Framework. In ACL

  83. [99]

    Shulman, L. S. 1986. Those Who Understand: Knowledge Growth in Teaching. Educational Researcher, 15(2): 4--14

  84. [100]

    Singhal, K.; Azizi, S.; Tu, T.; others; and Natarajan, V. 2023. Large Language Models Encode Clinical Knowledge. Nature, 620(7972): 172--180

  85. [101]

    Song, S.; Liu, W.; Lu, Y.; Zhang, R.; Liu, T.; Lv, J.; Wang, X.; Zhou, A.; Tan, F.; Jiang, B.; and Hao, H. 2025. Cultivating Helpful, Personalized, and Creative AI Tutors: A Framework for Pedagogical Alignment using Reinforcement Learning. arXiv:2507.20335

  86. [102]

    Sun, H.; Zhang, Z.; Deng, J.; Cheng, J.; and Huang, M. 2023. Safety Assessment of Chinese Large Language Models. arXiv:2304.10436

  87. [103]

    Y.; Cuadron, A.; Wang, C.; Popa, R

    Tan, S.; Zhuang, S.; Montgomery, K.; Tang, W. Y.; Cuadron, A.; Wang, C.; Popa, R. A.; and Stoica, I. 2025. JudgeBench: A Benchmark for Evaluating LLM -Based Judges. In ICLR

  88. [104]

    A.; and Sherin, M

    van Es, E. A.; and Sherin, M. G. 2002. Learning to Notice: Scaffolding New Teachers' Interpretations of Classroom Interactions. Journal of Technology and Teacher Education, 10(4): 571--596

  89. [105]

    VanLehn, K. 2011. The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist, 46(4): 197--221

  90. [106]

    Vygotsky, L. S. 1978. Mind in Society: The Development of Higher Psychological Processes. Harvard University Press

  91. [107]

    Wang, A.; Pruksachatkun, Y.; Nangia, N.; Singh, A.; Michael, J.; Hill, F.; Levy, O.; and Bowman, S. R. 2019. SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems. In NeurIPS

  92. [108]

    Wang, B.; Chen, W.; Pei, H.; others; Song, D.; and Li, B. 2023. DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models. In NeurIPS Datasets & Benchmarks

  93. [109]

    S.; and Wen, Q

    Wang, S.; Xu, T.; Li, H.; Zhang, C.; Liang, J.; Tang, J.; Yu, P. S.; and Wen, Q. 2024 a . Large Language Models for Education: A Survey and Outlook. arXiv:2403.18105

  94. [110]

    Wang, Y.; Ma, X.; Zhang, G.; et al. 2024 b . MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark. In NeurIPS Datasets & Benchmarks

  95. [111]

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E.; Le, Q.; and Zhou, D. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS

  96. [112]

    Wei, S.; Wang, X.; Bi, S.; Chen, J.; Li, R.; Jiang, B.; Lin, X.; Zhang, M.; Song, Y.; Li, B.; Zhou, A.; and Hao, H. 2025. ELMES : An Automated Framework for Evaluating Large Language Models in Educational Scenarios. arXiv:2507.22947

  97. [113]

    Weidinger, L.; Mellor, J.; Rauh, M.; et al. 2021. Ethical and Social Risks of Harm from Language Models. arXiv:2112.04359

  98. [114]

    xAI . 2025. Grok 4 Model Card. Model card, xAI. https://data.x.ai/2025-08-20-grok-4-model-card.pdf

  99. [115]

    Xu, B.; Bai, Y.; Sun, H.; et al. 2025. EduBench: A Comprehensive Benchmarking Dataset for Evaluating Large Language Models in Diverse Educational Scenarios. arXiv:2505.16160

  100. [116]

    Zhang, M.; et al. 2025. OmniEduBench: A Comprehensive Chinese Benchmark for Evaluating Large Language Models in Education. arXiv:2510.26422

  101. [117]

    Zhang, Z.; Lei, L.; Wu, L.; Sun, R.; Huang, Y.; Long, C.; Liu, X.; Lei, X.; Tang, J.; and Huang, M. 2024. SafetyBench: Evaluating the Safety of Large Language Models. In ACL

  102. [118]

    Zhao, S.; Yu, K.; Yuan, Y.; He, P.; and Wen, H. 2026. SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational LLMs . arXiv:2604.22134

  103. [119]

    P.; Zhang, H.; Gonzalez, J

    Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E. P.; Zhang, H.; Gonzalez, J. E.; and Stoica, I. 2023. Judging LLM -as-a-Judge with MT -Bench and Chatbot Arena. In NeurIPS Datasets & Benchmarks

  104. [120]

    Zhou, J.; Lu, T.; Mishra, S.; Brahma, S.; Basu, S.; Luan, Y.; Zhou, D.; and Hou, L. 2023. Instruction-Following Evaluation for Large Language Models. arXiv:2311.07911

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.