Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Teaching a Large Language Model Tutor to Withhold the Answer: A Supervisor Architecture and an Evidence-Driven Method for Tuning Socratic Behavior

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read A deployed LLM tutor can be made to withhold answers reliably by enforcing withholding as a per-turn contract and tuning it against evidence.

desk verdict A value-dense engineering report with honest limitations, but the headline compliance numbers are too brittle for the strength of the claim. read the letter →

arxiv 2608.12292 v1 pith:7IA4BHL4 submitted 2026-08-12 cs.CY

classification cs.CY
keywords intelligenttutoringsystemslargelanguagemodelsSocraticanswerwithholdingguardrailsLLMevaluationpromptinjectionhelpladder
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a large language model tutor can reliably withhold answers it could easily give, and that doing so requires engineering rather than prompting. The authors report a deployed tutor in which a non-LLM policy core sets a per-turn help ceiling on an eight-rung ladder, a deterministic detector strips solution code, and an LLM judge checks risky replies before they reach the student. They tune the system with a fully automated evaluation—scripted student personas, a stronger auditing model, and a recorded reason for every rejection—and report that the final run met all four acceptance gates: no solution reveals, a 0% earnest revise rate, 100% hint-ceiling compliance, and zero exam compromises. The larger claim is that this measure-diagnose-fix loop transfers to any LLM agent that must refuse a capability it has. A sympathetic reader would care because withholding is what makes a tutor educationally useful, not a safety bolt-on.

What carries the argument

The machinery is the pairing of a per-turn contract with an evidence-driven calibration loop. The contract is computed by a non-LLM policy core that reads only trusted learner state and outputs a ceiling on an eight-rung help ladder ($H_0$ acknowledge and encourage, $H_1$ restate, $H_2$ point to a concept, $H_3$ ask a leading question, $H_4$ describe the approach in words, $H_5$ give a worked example, $H_6$ give pseudocode with blanks, $H_7$ show the full solution); a deterministic code-reveal detector strips solution code before the reply ships; and a small-model judge reviews risky turns against the contract, preferring revision over blocking. The calibration loop drives scripted student personas through the live pipeline, re-judges every turn with a stronger auditing model, and records each rejection's stated reason, so failures are fixed by cause rather than by blind prompt edits.

What would settle it

Have a set of human expert raters independently score the same final-run transcripts against the same contract; if they detect solution leaks or over-blocking that the LLM auditor missed, the reported 100% compliance is an artifact of the auditor rather than genuine withholding behavior.

Watch

Extended reading notes

Core claim

The central discovery is that answer-withholding in an LLM tutor is best treated as a per-turn, machine-checkable contract rather than a behavior to be prompted into existence. The system computes each turn's allowed help ceiling from trusted learner state only, never from the student's text, making the binding decision immune to prompt injection; a deterministic detector blocks solution code independently of any model; and a separate LLM judge revises risky replies before they ship. The paper's empirical claim is that after successive fix-by-cause iterations, the tutor reached full compliance on all four acceptance criteria: G1 no solution reveals, G2 a 0% earnest revise rate (gate at most 5%), G3 100% hint-ceiling compliance under adversarial pressure (gate at least 95%), and G4 zero exam compromises. The failures encountered along the way formed an interpretable over-help ladder, from blatant solution leaks, to naming the exact bug, to over-citing general facts, with each fix exposing the next.

Load-bearing premise

The load-bearing premise is that roughly two dozen scripted turns, scored by an LLM auditor rather than by human experts or real students, are a valid stand-in for real student pressure; the paper itself says this does not substitute for real student traffic.

Editorial extensions

If this is right

  • A tutor can pass strict compliance gates on withholding without any human-subjects evaluation, making iteration cheap and fast.
  • The same architecture should transfer to other agents that must refuse a capability, with the over-help ladder as a general diagnostic shape for gross-to-subtle violations.
  • Over-blocking is treated as a measured failure, so tuning pushes toward a warm, still-helpful refusal rather than a generic 'I can't help with that'.
  • The decisive learning-outcome claim, that this withholding preserves practice gains and removes later test loss, remains to be shown with real students and a delayed, tool-removed assessment.

Reading between the lines

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

  • Editorial inference: if the over-help ladder generalizes, then any refusal task has a gross-to-subtle ordering of violations, and a reason-capturing loop should expose that ordering before the system is trusted.
  • Editorial inference: the deterministic detector's blind spots (split-fence, prose, and cross-language reveals) mean the architecture's safety depends on the judge and execution path; in domains where the forbidden capability is prose rather than code, the ceiling becomes the main defense.
  • Editorial inference: a testable extension is to calibrate the LLM auditor against a small set of human expert ratings on the same transcripts, since the reported 100% compliance could be re-scored if the auditor has systematic blind spots.
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 reports a deployed LLM tutoring system that enforces answer-withholding through a supervisor architecture: a non-LLM policy core computes a per-turn help ceiling on an eight-rung hint ladder from trusted learner state only, a deterministic code-reveal detector strips solution code, and an LLM judge checks risky replies against the contract. The authors then describe an automated, no-human-subjects calibration method in which scripted student personas are driven through the live pipeline and re-scored by a stronger auditing LLM, with each rejection's reason recorded. Applying this measure-diagnose-fix loop to a real course tutor produced an 'over-help ladder' of increasingly subtle violations, ending with 0% earnest revise rate (G2), 100% hint-ceiling compliance (G3), and zero deterministic reveals or exam compromises (G1, G4). The paper explicitly does not claim a learning-outcomes result and identifies the scripted evaluation as a limitation.

Significance. If the reported behavior holds, the paper makes a useful engineering contribution: it reframes answer-withholding as a per-turn, inspectable contract rather than a prompt-level aspiration, and it documents a diagnostic loop that separates true withholding failures from measurement artifacts and grounding errors. The non-LLM policy core (P2), the deterministic detector (P3), the explicit test suite, and the unusually candid limitations section are genuine strengths. The over-help ladder observation is a valuable, falsifiable description of how capable models err when instructed to withhold. However, the significance is tempered by the fact that the headline compliance numbers rest on roughly two dozen scripted turns scored by LLM judges with no human ground-truth calibration, so the paper is best read as a systems and calibration case study rather than as evidence of reliable withholding in real deployment.

major comments (3)
  1. [§5, §7] The final compliance claim is load-bearing but statistically thin. One full loop is 'roughly two dozen driven turns' (§4.3), and the final run reports 0/24-ish failures for G3. With zero observed failures in 24 trials, the one-sided 95% confidence upper bound on the true failure rate is about 12%, which does not actually establish that the true rate is below the 95% gate threshold. The paper should either report confidence intervals, run substantially more adversarial turns, or soften the 'full compliance' and 'reliable answer-withholding' language to 'met the acceptance gates in the scripted evaluation.'
  2. [§4.1, §4.3, §3.5] The G2 and G3 gates are scored by the same family of models whose behavior is being tuned: the actor is an LLM, the judge is an LLM, and the auditor is a stronger LLM. The paper even shows in Rung 0 that the auditor was wrong in a structured way. Without human-labeled gold ratings to calibrate the judge and auditor, or an independent deterministic outcome measure, the reported 100% G3 compliance may reflect inter-LLM agreement rather than true withholding behavior. This circularity is acknowledged in §7, but it undermines the strength of the central claim as currently worded. At minimum, the authors should provide a human-rater calibration study or frame the result as 'compliance as judged by the specified automated pipeline.'
  3. [§5, Table 1] The G2 result of 0% earnest revise rate is reported without the number of earnest turns in the final run, the variance across the 'handful of times' the loop was run, or any confidence bound. Since G2 is a gate against over-blocking, which the paper itself calls the more harmful error (P4), a single run with a small number of genuine turns gives little assurance that the production tutor will not over-block honest students at scale. The authors should report per-run counts, run-to-run variation, and a confidence interval or a larger sample.
minor comments (5)
  1. [Abstract, §1] The abstract calls the withholding behavior a 'per-turn, machine-checkable contract,' but G2 and G3 are scored by LLM judgments, not purely mechanical checks; only G1 and G4 are deterministic. Please qualify the language so the reader is not led to expect that every gate is machine-checkable in the strict sense.
  2. [§3.5] The judge is described as 'collusion-resistant' because it never sees the raw student message, which is a good trust-boundary property, but it is unclear how the judge can then assess whether the reply is responsive to the student's actual request. A brief explanation of what contextual signals the judge does receive (e.g., the pre-classifier's intent label or the retrieved sources) would clarify the design.
  3. [Table 1] The entry '54% reported /≈77% real' in the first row of Table 1 is confusing because the '≈77%' value is introduced only in the prose of §5. Please add a footnote or a column clarifying that the 54% figure was a measurement artifact and 77% was the corrected estimate.
  4. [§7] The limitations section says calibration against human raters is 'planned,' but the paper's central claims rely on the uncalibrated LLM auditor. Moving at least a small human-rater calibration into the current study, or explicitly labeling it as future work required before deployment-level claims, would strengthen the paper.
  5. [§4.3] The phrase 'roughly two dozen driven turns plus an audit pass' should be accompanied by the exact number of turns per persona suite and the number of loop repetitions, since the reported pass rates otherwise cannot be interpreted or reproduced by other researchers.

Circularity Check

1 steps flagged · score 3.0 of 10

G2/G3 acceptance scores are audits of the same LLM judge used for tuning, so the reported compliance is partly self-confirmatory; the deterministic gates remain independent.

  1. fitted input called prediction [§4.3, §5, and §7 (Limitations)]
    "for G3 we drive every evasion and injection, then re-judge all driven turns with the stronger auditing model, counting a turn compliant when the auditor does not reject it ... The final run passed both pedagogy gates: a 0% earnest revise rate (G2) and 100% hint-ceiling compliance ... The pedagogy gates pass on roughly two dozen driven turns run a handful of times across the scripted personas, which does not substitute for real student traffic."

    The G2/G3 gates are defined by LLM judge and auditor verdicts, and each iterative fix described in §5 (strict detector, constructive-code floor, 'cite only retrieved sources,' low-rung contract rewording, and even the auditor's own source-persistence fix at Rung 0) was made in response to that same auditor's rejections on the same scripted suite. The final 0% and 100% rates are therefore a measure of convergence to the auditor's preferences on the two dozen turns used for tuning, not an independent held-out prediction of withholding behavior. Because the system (and the auditor) were adjusted until the auditor stopped rejecting, the reported pass rates reduce by construction to the auditor's post-tuning approval, which §7 explicitly concedes does not substitute for real student traffic.

full rationale

The paper's deterministic gates G1 (no solution reveals) and G4 (no exam compromises) are machine-checkable invariants and do not depend on an LLM judge; those claims are self-contained and reproducible from the described architecture. The circularity burden rests entirely on G2 and G3. G3 is defined as 'the fraction of turns that stay at or below the contract’s help ceiling and leak no code, as scored by an independent auditor,' and G2 is defined by the live judge's 'revise' verdicts. The evaluation pipeline then tunes the system until that same auditor and judge pass, and reports the final numbers as full compliance. This is a genuine self-referential loop: the instrument used to define the acceptance gate is also the instrument used to drive the fixes. The paper is unusually candid about this, stating in §7 that the gates 'pass on roughly two dozen driven turns run a handful of times' and that both judge and auditor 'inherit the known reliability limits of LLM-as-judge, including for programming feedback.' No external ground-truth labels or human-raters calibration are used in the reported results, and the small sample (about two dozen turns) means the 100% figure carries wide uncertainty. That said, these are limitations of an empirical evaluation rather than a mathematical derivation; the architecture itself, the deterministic detector, and the policy core are described concretely and are not circular. The over-help ladder is a description of the observed fixes, not a prediction smuggled in. Overall, the central claim of 'full compliance' is partially circular for the two LLM-judged gates, but the deterministic portions and the method's honest disclosure keep the paper from being wholly self-confirming.

Assumptions & free parameters 2 free parameters · 3 assumptions · 3 invented entities

The central claim rests on three families of assumptions: the choice of acceptance thresholds and gate definitions (free parameters), the validity of LLM-as-judge scoring on a tiny scripted-pesonas, and the pedagogical premise that withholding helps learning. The architecture also introduces the help-ladder and supervisor components as new design entities. The paper is transparent about most of these, but the headline pass rates inherit the uncertainty of LLM judges.

free parameters (2)
  • G2 gate threshold (5%) = 5%
    The allowed earnest revise rate is a chosen threshold, not derived. It is an acceptance criterion the authors picked, and the final result of 0% is measured against it.
  • G3 gate threshold (95%) = 95%
    The required hint-ceiling compliance is a chosen threshold. The paper does not derive it from any cost-benefit model.
assumptions (3)
  • domain assumption Scripted personas and LLM-as-judge scores are a valid proxy for real student behavior under pressure.
    The entire calibration and the reported compliance figures rest on this. The paper acknowledges in section 7 that this does not substitute for real student traffic.
  • domain assumption The LLM judge and auditor can distinguish a true over-help violation from a grounding failure or pedagogy-quality issue, after the cause-sorting refinement.
    The cause-sorting discipline assumes the auditor's categorical reasons are meaningful and correct. The paper itself reports the Rung 0 incident where the auditor was wrong in a systematic way.
  • domain assumption The pedagogical premise that withholding answers and forcing productive struggle improves learning.
    The paper draws on Bastani et al., Kapur's productive failure, and feedback meta-analyses. The paper's own contribution is not a direct test of this premise, and it states that the decisive test is future work.
invented entities (3)
  • Eight-rung help ladder H0 through H7
    purpose: Discretizes help into typed rungs so the policy core can compute a per-turn ceiling and the detectors can check compliance.
    The ladder is a design artifact of this paper. It is grounded in the cited help-seeking literature, but the specific rung definitions and the ceiling semantics are new and only validated through the paper's own gates.
  • Supervisor components: policy core, pre-classifier, strategist, code-reveal detector, LLM judge
    purpose: Enforce and check the per-turn withholding contract.
    These are architectural components described in the paper. Their effectiveness is evaluated only via the paper's own scripted personas and LLM judges, not through an external benchmark.
  • Over-help ladder
    purpose: Organizes the failure modes observed during calibration into an ordered list from blatant solution leaks to subtle over-citing.
    This is an interpretative summary of the debugging experience. It is plausible and useful, but no external dataset confirms that all withholding failures follow this exact order.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teaching a Large Language Model Tutor to Withhold the Answer: A Supervisor Architecture and an Evidence-Driven Method for Tuning Socratic Behavior." pith.science (2026). https://pith.science/paper/7IA4BHL4

@misc{pith2026260812292,
  author       = {Pith},
  title        = {Pith review of: Teaching a Large Language Model Tutor to Withhold the Answer: A Supervisor Architecture and an Evidence-Driven Method for Tuning Socratic Behavior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IA4BHL4}},
  note         = {Machine review of arXiv:2608.12292}
}
read the original abstract

An effective large language model (LLM) tutor must often decline to give an answer it could easily produce. In a randomized study, students who used an unguarded chatbot scored higher while practicing but lower on a later test taken without it, whereas a Socratically guarded version of the same model kept the practice gain and removed the later loss [4]. Reliable answer-withholding is therefore central to a tutor's value, yet a capable model pressed by a frustrated student does not withhold reliably on a prompt alone. We report a deployed tutoring system that enforces answer-withholding as a per-turn, machine-checkable contract, and a method for tuning that withholding against evidence. A non-LLM policy core, reading only trusted learner state, sets a per-turn ceiling on an eight-rung help ladder; a deterministic detector strips solution code; and a separate LLM judge checks each risky reply against the contract. We tune the behavior with an automated evaluation that uses no human subjects: scripted student personas are driven through the live pipeline and re-scored by a stronger model, and we record each rejection's stated reason so failures are fixed by cause. Doing so revealed an interpretable "over-help ladder," from blatant solution leaks, to naming the exact bug, to over-citing general facts, with each fix exposing the next. The tutor reached full compliance on all four acceptance criteria. We offer the measure, diagnose, and fix loop as a reusable recipe for any LLM agent that must refuse a capability it has.

Figures

Figures reproduced from arXiv: 2608.12292 by the authors.

Figure 1
Figure 1. Per-turn flow. The student talks only to the actor. Inside the trust boundary, the policy core computes the help [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 6 canonical work pages

  1. [1]

    Vincent Aleven, Elmar Stahl, Silke Schworm, Frank Fischer, and Raven Wal- lace. 2003. Help Seeking and Help Design in Interactive Learning Environ- ments.Review of Educational Research73, 3 (2003), 277–320. doi:10.3102/ 00346543073003277

  2. [2]

    Anderson, Albert T

    John R. Anderson, Albert T. Corbett, Kenneth R. Koedinger, and Ray Pelletier

  3. [3]

    Game the System

    Ryan S. J. d. Baker, Albert T. Corbett, Kenneth R. Koedinger, and Angela Z. Wagner. 2004. Off-Task Behavior in the Cognitive Tutor Classroom: When Students “Game the System”. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’04). 383–390. doi:10.1145/985692.985741

  4. [4]

    Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakçı, and Rei Mariman. 2025. Generative AI Without Guardrails Can Harm Learning: Evidence Teaching an LLM Tutor to Withhold the Answer from High School Mathematics.Proceedings of the National Academy of Sciences 122, 26 (2025), e2422633122. doi:10.1073/pnas.2422633122

  5. [5]

    Benjamin S. Bloom. 1984. The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring.Educational Researcher 13, 6 (1984), 4–16. doi:10.3102/0013189X013006004

  6. [6]

    Michelene T. H. Chi and Ruth Wylie. 2014. The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes.Educational Psychologist 49, 4 (2014), 219–243. doi:10.1080/00461520.2014.965823

  7. [7]

    Corbett and John R

    Albert T. Corbett and John R. Anderson. 1995. Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge.User Modeling and User-Adapted Interaction4, 4 (1995), 253–278. doi:10.1007/BF01099821

  8. [8]

    Jean-Paul Doignon and Jean-Claude Falmagne. 1985. Spaces for the Assessment of Knowledge.International Journal of Man-Machine Studies23, 2 (1985), 175–196. doi:10.1016/S0020-7373(85)80031-6

Show all 32 references
  1. [9]

    Graesser, Patrick Chipman, Brian C

    Arthur C. Graesser, Patrick Chipman, Brian C. Haynes, and Andrew Olney. 2005. AutoTutor: An Intelligent Tutoring System with Mixed-Initiative Dialogue.IEEE Transactions on Education48, 4 (2005), 612–618. doi:10.1109/TE.2005.856149

  2. [10]

    Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. In Proceedings of the 16th ACM Workshop on Artificial Inte...

  3. [11]

    John Hattie and Helen Timperley. 2007. The Power of Feedback.Review of Educational Research77, 1 (2007), 81–112. doi:10.3102/003465430298487

  4. [12]

    Heffernan and Cristina L

    Neil T. Heffernan and Cristina L. Heffernan. 2014. The ASSISTments Ecosystem: Building a Platform that Brings Scientists and Teachers Together for Minimally Invasive Research on Human Learning and Teaching.International Journal of Artificial Intelligence in Education24, 4 (201...

  5. [13]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.Comput. Surveys55, 12, Article 248 (2023), 38 pages. doi:10.1145/3571730

  6. [14]

    Amanpreet Kapoor, Marc Diaz, Stephen MacNeil, Leo Porter, and Paul Denny

  7. [15]

    Manu Kapur. 2008. Productive Failure.Cognition and Instruction26, 3 (2008), 379–424. doi:10.1080/07370000802212669

  8. [16]

    Henley, Paul Denny, Michelle Craig, and Tovi Grossman

    Majeed Kazemitabaar, Runlong Ye, Xiaoning Wang, Austin Z. Henley, Paul Denny, Michelle Craig, and Tovi Grossman. 2024. CodeAid: Evaluating a Classroom Deployment of an LLM-based Programming Assistant that Balances Student and Educator Needs. InProceedings of the 2024 CHI Confe...

  9. [17]

    Charles Koutcheme, Nicola Dainese, Arto Hellas, Sami Sarsa, Juho Leinonen, Syed Ashraf, and Paul Denny. 2025. Evaluating Language Models for Generating and Judging Programming Feedback. InProceedings of the 56th ACM Technical Symposium on Computer Science Education V. 1 (SIGCS...

  10. [18]

    LearnLM Team, Google. 2024. LearnLM: Improving Gemini for Learning. Tech- nical report. arXiv:2412.16429 [cs.CY]

  11. [19]

    Sheese, Jaromir Savelka, and Paul Denny

    Mark Liffiton, Brad E. Sheese, Jaromir Savelka, and Paul Denny. 2024. CodeHelp: Using Large Language Models with Guardrails for Scalable Support in Program- ming Classes. InProceedings of the 23rd Koli Calling International Conference on Computing Education Research. doi:10.11...

  12. [20]

    Jiayu Liu, Zhenya Huang, Tong Xiao, Jing Sha, Jinze Wu, Qi Liu, Shijin Wang, and Enhong Chen. 2024. SocraticLM: Exploring Socratic Personalized Teaching with Large Language Models. InAdvances in Neural Information Processing Systems 37 (NeurIPS 2024)

  13. [21]

    Rongxin Liu, Carter Zenke, Charlie Liu, Andrew Holmes, Patrick Thornton, and David J. Malan. 2024. Teaching CS50 with AI: Leveraging Generative Artificial Intelligence in Computer Science Education. InProceedings of the 55th ACM Technical Symposium on Computer Science Educatio...

  14. [22]

    Antonija Mitrovic. 2012. Fifteen Years of Constraint-Based Tutors: What We Have Achieved and Where We Are Going.User Modeling and User-Adapted Interaction22, 1-2 (2012), 39–72. doi:10.1007/s11257-011-9105-9

  15. [23]

    Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S

    Joon Sung Park, Carolyn Q. Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Robb Willer, Percy Liang, and Michael S. Bernstein

  16. [24]

    Schroeder, Olusola O

    Noah L. Schroeder, Olusola O. Adesope, and Rachel Barouch Gilbert. 2013. How Effective Are Pedagogical Agents for Learning? A Meta-Analytic Review.Journal of Educational Computing Research49, 1 (2013), 1–39. doi:10.2190/EC.49.1.a

  17. [25]

    Kurt VanLehn. 2011. The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems.Educational Psychologist46, 4 (2011), 197–221. doi:10.1080/00461520.2011.611369

  18. [26]

    Wang, Ana T

    Rose E. Wang, Ana T. Ribeiro, Carly D. Robinson, Susanna Loeb, and Dorottya Demszky. 2024. Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. arXiv:2410.03017 [cs.CL]

  19. [27]

    Simon Willison. 2023. The Dual LLM Pattern for Building AI Assistants that Can Resist Prompt Injection. Blog post. https://simonwillison.net/2023/Apr/25/dual- llm-pattern/

  20. [28]

    Benedikt Wisniewski, Klaus Zierer, and John Hattie. 2020. The Power of Feedback Revisited: A Meta-Analysis of Educational Feedback Research.Frontiers in Psychology10, Article 3087 (2020). doi:10.3389/fpsyg.2019.03087

  21. [29]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-Judge with MT- Bench and Chatbot Arena. InAdvances in Neural Information P...

  22. [1995]

    doi:10.1207/s15327809jls0402_2

    Cognitive Tutors: Lessons Learned.Journal of the Learning Sciences4, 2 (1995), 167–207. doi:10.1207/s15327809jls0402_2

  23. [2024]

    arXiv:2411.10109 [cs.AI]

    Generative Agent Simulations of 1,000 People. arXiv:2411.10109 [cs.AI]

  24. [2025]

    arXiv:2504.11146 [cs.HC]

    Exploring Student Behaviors and Motivations using AI TAs with Optional Guardrails. arXiv:2504.11146 [cs.HC]

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.