acceptodds
Under review as a conference paper at ICLR 2027

Optimizing Tutor LLMs for Pedagogy While Maintaining Student Engagement

Abstract

Large language models for education are optimized to follow effective pedagogy: they scaffold student learning and help students evaluate and self-correct their reasoning, without giving direct answers. These tutors are typically trained with supervised learning and reinforcement learning against a simulated student that is an assistant LLM. Such simulated students are always enthusiastic and respond at every turn, unlike real students. As a result, when a student becomes less cooperative, current tutors give away solutions too often and struggle to keep the learner engaged. To address this, we train tutors with reinforcement learning against a user model: an LLM trained to imitate human users, rather than being a helpful assistant. Like real students, it can lose interest and quit the conversation at any time, so the tutor must keep it engaged to teach it anything. We train using a combination of two rewards: an engagement reward to keep the student's attention, and a learning reward to maintain pedagogy. We find that the two rewards are complementary, because staying engaged enables longer conversations where learning happens in later turns. Overall, our tutors achieve higher judged learning than prior tutor models while leaking solutions less. We will release two open-source tutors, one for easier and one for harder math problems.

Primary area:
reinforcement learning
Keywords:
LLM tutoring, reinforcement learning, LLM-as-student, user engagement, GRPO, multi-turn RL, AI for education

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_L6KbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates a limitation of training tutor LLMs with assistant LLMs as simulated students. These students are overly cooperative and remain engaged throughout the conversation, unlike real learners who may lose interest or leave. The authors show that a tutor trained in this setting is more likely to reveal solutions when its cooperative simulated student is replaced with a user model that can disengage. They propose EngagementRL, which trains tutors against existing user models capable of ending conversations. The method combines an engagement reward based on the student’s probability of leaving with a learning reward assessed by an LLM judge, while penalizing solution leakage. The authors train two tutor models, UserLM-tutor and OSIM-tutor. Experiments in simulated math tutoring show that both models leak fewer solutions and achieve higher judged learning scores than prior tutor models.

Critical strengths
  1. The paper shows that a tutor trained with consistently cooperative simulated students is more likely to give away solutions when interacting with less engaged students. Figure 2 demonstrates this failure in TutorRL. Although UserLM is an imperfect proxy for real students, this result exposes a concrete weakness in maintaining pedagogical behavior when students participate less actively or may leave the conversation. This finding provides a compelling motivation for incorporating student engagement into tutor training.

  2. The method combines engagement and learning rewards to encourage continued student participation alongside learning progress. The ablations show that the combined objective achieves higher judged learning than either reward alone in both student settings, while also improving retention in the UserLM setting.

  3. The engagement reward is inexpensive to obtain because it uses the student simulator’s probability of ending the conversation without requiring an additional judge. This provides a practical way to incorporate student engagement into tutor training.

Critical weaknesses
  1. The experiments do not sufficiently separate the contribution of the student simulator from the proposed training method. There is no matched comparison using an assistant model as the student with the same combined reward and optimization procedure. It therefore remains unclear how much additional benefit comes from using user models. This limits the evidence that replacing the student simulator is a key factor in improving tutor robustness to student disengagement.

  2. The evaluation metrics that the author propose have several limitations. Quit rate does not distinguish students who leave after learning from those who leave because they lose interest. Engagement and learning are evaluated using similar procedures and some scoring criteria overlap, particularly cognitive engagement and understanding. Whether these metrics adequately distinguish engagement from learning therefore needs further validation. Besides, only the engagement-related metrics were compared with human annotations, and the correlations were weak. The evaluation could also include some other objective measures. For example, solve rate can be affected by solution leakage, but it would provide additional evidence of whether higher learning scores reflect better performance in solving problems.

Questions for authors
  1. Do you have a matched comparison between user models and assistant models as students, using the same training setting and optimization procedure? What additional benefit does using a user model provide under these conditions?

  2. How is misconception repair scored when the student has no initial errors or misconceptions to correct?

  3. Beyond the proposed metrics, is there other objectively verifiable evidence of student learning? Does this evidence support the proposed method’s improvements over the baselines?

  4. In Table 2, the combined objective achieves higher judged engagement than engagement-only training in both student settings. Which dimensions account for this improvement? Could you explain the OSIM results, where the combined objective achieves higher judged engagement but also a higher quit rate compared with engagement-only training?

4 comments
AuthorsOpenReview ↗

Response to questions

We thank the reviewer for the detailed questions. All new results are in the updated paper, and each point below lists where to find them. Happy to discuss more details or add experiments if necessary.

  • Q1: What does a user model add over an assistant model playing the student? Our main baseline already answers this. TutorRL trains its tutor against an assistant LLM prompted to act as a student, which is the setup the reviewer describes. As we discuss in the paper, assistant models are built to be helpful and responsive, so as students they almost never lose interest or leave, and they answer every tutor message in good faith. A tutor trained against such a student never learns to handle a student who is bored, frustrated or ready to give up. Often its fastest route to reward is simply to give the answer away. This is why we train against user models built to behave like real people, including disengaging when the tutoring is poor.

  • Q2: Misconception repair when the student made no mistake. In the original rubric, such a turn effectively scores 0 on this dimension, and the case is common (35–64% of turns). We rescored all test dialogues with an added "not applicable" option and left it out of those turns' average. All scores rise, but the ranking barely changes (rank correlation 0.98 with UserLM, 1.00 with OSIM), and our tutor stays well ahead of TutorRL-7B. (Appendix L, Table 16)

    TutorStudentTurns with nothing to repairLearning (original) ↑Learning (n/a left out) ↑
    TutorRL-7BUserLM35%0.0800.097
    OursUserLM42%0.1420.181
    TutorRL-7BOSIM64%0.1840.253
    OursOSIM62%0.3000.390
  • Q3: Objective evidence of learning. We add a judge-free solve-rate measure: a dialogue counts if a student message contains the correct final answer before any tutor message states it. With OSIM, students reach the answer on their own far more often with our tutor (33.2% vs. 21.2%), and our tutor states the answer less often. With UserLM the gap is small, because UserLM rarely writes a final answer with any tutor, but our tutor states the answer about half as often. (Appendix M, Table 17)

    StudentTutorStudent reaches answer on their own ↑Tutor states the answer ↓
    UserLMTutorRL-7B12.4%23.6%
    UserLMOurs14.6%12.6%
    OSIMTutorRL-7B21.2%28.8%
    OSIMOurs33.2%20.8%
  • Q4: Which engagement dimensions improve, and why OSIM quits more than with engagement-only training. The gain comes mainly from cognitive (+0.80) and behavioral (+0.69) engagement, that is, the student actually working on the problem. The engagement-only tutor keeps OSIM talking without making it work; the learning reward adds that work. The higher quit rate comes almost entirely from students who solved the problem and then left (25% vs. 12%), while students leaving unfinished change little (10% vs. 7%). Students with the combined tutor reach the answer on their own twice as often, and with turns after a quit counted as zero, it still has higher engagement. (Appendix O, Table 19)

    OSIM studentEngagement onlyCombinedDifference
    Behavioral (0–4)2.363.05+0.69
    Affective (0–4)2.562.95+0.39
    Cognitive (0–4)1.912.71+0.80
    Quit rate0.190.35+0.16
    Left unfinished7.4%10.0%+2.6 pts
    Finished, then left11.6%25.2%+13.6 pts
    Reaches answer on their own17%33%+16 pts
AuthorsOpenReview ↗

Critical Weaknesses clarifications

  • Quit rate: leaving after learning vs. giving up. (See Q4 table) As suggested, we now split every early quit by whether the student had already solved the problem. Against TutorRL-7B, far fewer students leave our tutor unfinished (UserLM: 4% vs. 25%; OSIM: 10% vs. 13%). Our extra OSIM quits are all students who had already finished, since OSIM tends to leave once it is done. (Appendix N, Table 18)

  • Overlap between engagement and learning metrics. We agree that cognitive engagement and understanding are related. They are scored in separate judge calls with separate rubrics: cognitive engagement rates depth of processing regardless of correctness, while learning credits only correct, student-produced work. Each score also has two dimensions that ask about different things (behavioral and affective engagement; solution progress and misconception repair). With UserLM, affective engagement correlates only 0.21 per turn with understanding, against 0.43 for cognitive engagement. The two metrics also rank tutors differently: with UserLM, the engagement-only tutor scores far higher on engagement than the learning-only tutor (0.483 vs. 0.398) but not on learning (0.100 vs. 0.106). With OSIM, learning-only training gives much higher learning (0.374 vs. 0.151) but also more quits (0.47 vs. 0.19) and more leaks (0.175 vs. 0.021). If the metrics measured the same thing, they would not separate like this. Table 2; Appendix B (rubrics)

  • Human evaluation, covering learning as well as engagement. We ran a blind pairwise study on Prolific. 20 raters compared our tutor with TutorRL-7B on 100 random test problems, both with the same UserLM student. Raters preferred our tutor for both engagement and learning, and found that TutorRL-7B gives away the answer more often (all p ≤ 0.02). (Section 5, Table 3a; Appendix I, Table 12)

    OursTutorRL-7BSamep
    Keeps student engaged54%26%20%0.002
    Helps student learn56%27%17%0.002
    Overall preference64%36%–0.007
    Gives away the answer ↓21%35%–≤ 0.02

As we approach the end of the discussion period, we would appreciate any last comments, feedback or questions. We have tried to fully address all of the reviewer's concerns and questions. We have added extensive experiments and details to the revised paper. Please let us know if our responses are satisfactory, and we are happy to provide more details or clarifications if necessary.

Thanks for your response! Your clarification has addressed most of my concerns, and the experiments are now more comprehensive. I have updated my score to 3.

Reviewer_LeaKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EngagementRL, a reinforcement learning framework for training LLM-based tutors to maintain student engagement while promoting learning. The authors identify a limitation of existing tutoring approaches: assistant-LLM student simulators tend to remain consistently cooperative, potentially leading to poor generalization when interacting with less engaged students.

To address this issue, the authors train tutors against user simulators (UserLM and OSIM) that can disengage and terminate conversations. The proposed framework combines an engagement reward derived from the simulator's probability of ending the conversation with an LLM-judged learning reward based on solution progress, demonstrated understanding, and misconception repair. It further introduces per-turn credit assignment using a zero-padded group baseline within GRPO.

Experiments on Big-Math and OpenR1-Math compare the approach against existing tutoring models and general-purpose LLMs. The results show improvements in judged engagement and learning, supported by reward ablations, cross-simulator evaluations, and an additional held-out judge analysis on a subset of UserLM dialogues.

Critical strengths

1. Important and well-motivated research problem. The paper identifies a meaningful limitation of existing LLM-based tutoring approaches, namely that assistant-LLM student simulators tend to be overly cooperative and may fail to reflect realistic disengagement behaviors. The authors provide empirical evidence for this issue by evaluating TutorRL under two different student simulators and analyzing seven educational dialogue corpora. This finding highlights an important gap between simulated tutoring environments and more realistic student interactions, providing a strong motivation for the proposed approach.

2. Practical method with supporting experimental evidence. EngagementRL introduces a relatively inexpensive engagement reward derived from the student simulator's probability of terminating the conversation and combines it with a learning reward through turn-level credit assignment. The experiments show that jointly optimizing these objectives improves judged engagement and learning compared with optimizing either objective alone. The ablation studies further support the contributions of the proposed components. These results suggest a practical approach to improving multi-turn tutoring without requiring an additional learned critic.

Critical weaknesses

1. Limited validation of the engagement and learning proxies. The evaluation relies primarily on simulated students and LLM-based judgments rather than independently validated measures of actual learning. The simulator's predicted conversation-termination probability shows only a weak negative correlation with human-rated interestingness (Spearman ρ = -0.14). In addition, the learning metric uses the same judge and rubric during training and the main evaluation. Although the authors provide an additional evaluation with a held-out judge, it covers only 100 problems in the UserLM setting. These limitations raise concerns about how well the reported improvements reflect genuine student engagement and learning beyond the simulated environment.

2. Inconsistent student retention and limited isolation of generalization effects. Although the combined objective achieves strong judged engagement and learning scores, its retention performance is less convincing in the OSIM setting. Specifically, the combined model has a quit rate of 47.4%, compared with 26.2% for TutorRL-7B and 19.0% for the engagement-only variant (Table 2). Since judged engagement is averaged over observed student turns, early termination may also affect the interpretation of the reported engagement scores. Furthermore, the cross-simulator evaluation in Table 3 changes both the student simulator and the problem dataset, making it difficult to isolate generalization to unseen student behaviors. These issues limit the evidence for consistent engagement improvements across different settings.

Questions for authors

Q1. Reliability and validity of the learning evaluation.

The main learning metric is evaluated using gpt-oss-20b, which is also used to provide learning rewards during training. Although Appendix G reports results from a held-out judge, this evaluation covers only 100 Big-Math problems with UserLM.

Could the authors clarify whether the reported learning improvements remain consistent under an independent judge in the OSIM setting? If additional evaluation results or analyses of existing transcripts are available, these would help establish whether the improvements reflect substantive learning behaviors rather than adaptation to the training judge's preferences.

Q2. Student retention and engagement in the OSIM setting.

In Table 2(b), the combined model achieves the highest judged engagement score (0.694), but its quit rate (47.4%) is substantially higher than that of TutorRL-7B (26.2%) and the engagement-only variant (19.0%).

Could the authors explain this discrepancy? In particular, does the engagement improvement persist when early termination is explicitly accounted for, rather than averaging engagement scores only over the observed student turns? A retention-aware analysis or a comparison of engagement scores between completed and prematurely terminated dialogues would help clarify whether the method consistently achieves its stated objective of maintaining student engagement.

3 comments

We thank the reviewer for the insightful questions. All new results are in the updated paper, and each point below lists where to find them.

  • Q1: Held-out judges on the full test set, for both students. We now rescore all 500 test dialogues of every tutor, with both students, using two judges never used in training: Qwen3-32B and Gemma-3-27B-IT. Absolute scores differ between judges, but they rank the tutors almost exactly like the training judge (Spearman 0.92–1.00 for learning, 0.93–0.97 for engagement). Our tutor stays ahead of TutorRL on both metrics, with both students, under both judges, including the OSIM setting. (Section 5, Table 3b; Appendix J, Tables 13–14)

    Tutor (engagement / learning)UserLM, Qwen3-32BUserLM, Gemma-3OSIM, Qwen3-32BOSIM, Gemma-3
    gpt-oss-20b (as tutor)0.09 / 0.020.07 / 0.030.68 / 0.450.62 / 0.53
    TutorRL-7B0.45 / 0.190.41 / 0.220.57 / 0.310.52 / 0.38
    Ours0.56 / 0.260.53 / 0.300.75 / 0.460.67 / 0.54
  • Q2: Retention-aware engagement. A student can end the dialogue early for two reasons: it gives up, or it solves the problem and leaves. We split every early quit by whether the judge had already marked the problem as solved by the student, and we rescore engagement with every turn after a quit counted as zero. With OSIM, all of our extra quits are students who had already solved the problem, and fewer students leave our tutor unfinished (10% vs. 13%). With UserLM, far fewer leave unfinished (4% vs. 25%). After counting early exits as zero, our tutor still has the highest engagement with both students.

    StudentTutorLeft unfinished ↓Finished, then leftEngagement ↑Engagement, turns after quit = 0 ↑
    UserLMTutorRL25%1%0.4000.386
    UserLMOurs4%0%0.4920.490
    OSIMTutorRL13%13%0.5370.486
    OSIMOurs10%25%0.7430.626

    The same holds against the engagement-only tutor (quit rate 0.19): the combined tutor's extra quits are mostly students who finished (25% vs. 12%), and its students reach the answer on their own twice as often (33% vs. 17%).(Section 5; Appendix N, Table 18; Appendix O, Table 19)

  • Transfer with the problem set held fixed. We now test each tutor and its base model with each student on both problem sets, so the student and the problems change one at a time. Changing only the problem set results in each tutor beating its base model on both sets. Our UserLM-trained tutor also transfers to OSIM, a student it never saw: it keeps far more students than its base model. (Appendix P, Table 20)

    OSIM student (learning / engagement / quit)OpenR1 problemsBig-Math problems
    Ours, trained with OSIM0.30 / 0.74 / 0.350.26 / 0.63 / 0.59
    Ours, trained with UserLM0.17 / 0.49 / 0.090.17 / 0.45 / 0.11
    Qwen2.5-7B (base of UserLM tutor)0.15 / 0.48 / 0.590.14 / 0.42 / 0.71
    Qwen3-8B (base of OSIM tutor)0.18 / 0.50 / 0.680.16 / 0.47 / 0.78

    Bold = training setting.

  • Human evaluation. Since the engagement proxies correlate only modestly with existing human labels, we ran a blind pairwise study on Prolific. 20 raters compared our tutor with TutorRL-7B on 100 random test problems, both with the same UserLM student. Raters preferred our tutor on every question and found that TutorRL-7B gives away the answer more often (all p ≤ 0.02). (Section 5, Table 3a; Appendix I, Table 12)

    OursTutorRL-7BSamep
    Keeps student engaged54%26%20%0.002
    Helps student learn56%27%17%0.002
    Overall preference64%36%–0.007
    Gives away the answer ↓21%35%–≤ 0.02

Thanks for the additional experiments and clarifications. They address most of my concerns, especially regarding the independent judges and student retention. I appreciate the extra effort. I just noticed that the OSIM quit rate changed from 47.4% in the original version to 35.2% in the revised version. Could you briefly clarify why the results changed? Overall, I maintain my original rating of 3 (Weak Accept).

The numbers changed because we reran all experiments with a more careful checkpoint selection. In the original version, every tutor was trained for a fixed 300 steps and we evaluated the final checkpoint. In the revised version, we validate every 20 steps on a separate validation split and select the checkpoint with early stopping (Appendix E.6). The numbers are slightly different but the conclusions are unchanged. Thank you again for your time and helpful feedback.

Comments on the paper

3 comments

We thank the reviewer for finding our paper well motivated, and raising detailed questions. All new result details are in the updated paper, and each point below lists where to find them.

  • Q1 Human evaluation. We ran a blind pairwise study on Prolific. 20 raters compared our tutor with TutorRL on 100 random test problems, both with the same UserLM student. Raters preferred our tutor on every question and found that TutorRL gives away the answer more often. All four differences are significant (p ≤ 0.02), and they agree with the LLM-judge ranking. (Section 5, Table 3a; Appendix I, Table 12)

    OursTutorRL-7BSamep
    Keeps student engaged54%26%20%0.002
    Helps student learn56%27%17%0.002
    Overall preference64%36%–0.007
    Gives away the answer ↓21%35%–≤ 0.02
  • Q2 Why the judges disagree on leakage. We now score the full test set with both students and add a third judge, Qwen3-32B. Gemma-3 flags far more dialogues than the other two judges. In 76–85% of the turns that only Gemma-3 flags, its own rationale says the tutor did not give the answer; it flags close step-by-step guidance, which our rubric explicitly says is not a leak. Gemma-3 therefore applies a stricter definition rather than catching more leaked answers. Under all three judges, and under human raters, our tutor leaks less than TutorRL-7B. (Appendix K, Table 15)

    Dialogues with ≥1 leak flag ↓gpt-oss-20b (training)Qwen3-32BGemma-3
    Ours, UserLM5%1%59%
    TutorRL-7B, UserLM24%17%64%
    Ours, OSIM12%4%29%
    TutorRL-7B, OSIM26%9%49%
  • Q3 Misconception repair when the student made no mistake. In the original rubric, such a turn effectively scores 0 on this dimension, and the case is common (35–64% of turns). We rescored all test dialogues with an added "not applicable" option and left it out of those turns' average. All scores rise, but the ranking barely changes (rank correlation 0.98 with UserLM, 1.00 with OSIM), and our tutor stays well ahead of TutorRL-7B. (Appendix L, Table 16)

    TutorStudentTurns with nothing to repairLearning (original) ↑Learning (n/a left out) ↑
    TutorRL-7BUserLM35%0.0800.097
    OursUserLM42%0.1420.181
    TutorRL-7BOSIM64%0.1840.253
    OursOSIM62%0.3000.390
  • Q4 Why our tutor has a higher OSIM quit rate than TutorRL. A student can end the dialogue early for two reasons: it gives up, or it solves the problem and leaves. OSIM tends to leave once it is done. Splitting the quits this way shows that all of our extra quits are students who had already solved the problem; fewer students leave our tutor unfinished (10% vs. 13%). With every turn after a quit counted as zero, our tutor still has higher engagement (0.626 vs. 0.486). (Section 5; Appendix N, Table 18)

    OSIM studentLeft unfinished ↓Finished, then leftEngagement ↑Engagement, turns after quit = 0 ↑
    TutorRL-7B13%13%0.5370.486
    Ours10%25%0.7430.626
  • Q4 Transfer with the problem set held fixed. We now test each tutor and its base model with each student on both problem sets, so the student and the problems change one at a time. Changing only the problem set results in each tutor beating its base model on both sets. Our UserLM-trained tutor also transfers to OSIM, a student it never saw: it keeps far more students than its base model. (Appendix P, Table 20)

    OSIM student (learning / engagement / quit)OpenR1 problemsBig-Math problems
    Ours, trained with OSIM0.30 / 0.74 / 0.350.26 / 0.63 / 0.59
    Ours, trained with UserLM0.17 / 0.49 / 0.090.17 / 0.45 / 0.11
    Qwen2.5-7B (base of UserLM tutor)0.15 / 0.48 / 0.590.14 / 0.42 / 0.71
    Qwen3-8B (base of OSIM tutor)0.18 / 0.50 / 0.680.16 / 0.47 / 0.78

    Bold = training setting.

Dear Reviewers,

Thank you again for your thoughtful reviews. We are glad that all three reviewers found the paper well motivated. Reviewers noted that our motivation rests on a concrete failure case: a pedagogically aligned tutor gives away far more answers once its student can disengage, and our analysis shows that this engagement gap between simulated and real students is systematic. Reviewers also appreciated the design of our engagement reward, which is cheap and practical, without extra judge calls or learned critics. Our new results (held-out judges on the full test set, a human study, and controlled transfer experiments) further show that these gains hold beyond the training judge and setup.

The discussion period ends in a couple of days, so we want to briefly summarize how we addressed the main concerns. We believe we have answered every question and concern raised, and we would be grateful to hear whether our responses resolve your concerns, or whether you have further questions we can still answer.

  1. No human evaluation (raised by all three reviewers). All reviewers noted that our evidence relied on simulated students and LLM judges. We added a blind pairwise study on online recruited crowd raters. Raters compared our tutor with TutorRL-7B (our main comparison baseline), with both tutors talking to the same student.
  • Raters preferred our tutor for keeping the student engaged (54% vs. 26%) and for helping the student learn (56% vs. 27%).
  • They preferred it overall in 64% of pairs.
  • They judged that TutorRL-7B gave away the answer more often (35% vs. 21% of chats).
  • All differences are statistically significant and agree with the LLM-judge ranking.
  1. Same judge in training and evaluation (Reviewers U6pP, LeaK). The held-out judge check in the original paper covered only 100 problems with one student. We now rescore the full test set, with both students, using two judges that were never used in training (Qwen3-32B and Gemma-3). The absolute scores differ between judges, but the tutor rankings stay almost the same as under the training judge. Our tutor stays ahead of TutorRL-7B on engagement and learning in every setting, including OSIM.

  2. Higher quit rate with OSIM (all three reviewers). Reviewers asked why more OSIM students leave our tutor early, even though our tutor scores higher on engagement and learning. A student can leave for two very different reasons: it gives up, or it has solved the problem and is done. OSIM usually leaves as soon as it has solved the problem. When we separate these two cases, fewer students give up with our tutor than with TutorRL-7B. Our higher quit rate comes entirely from students who solved the problem and then left. In other words, more of our students leave because they succeeded, not because they lost interest.

  3. Misconception repair when the student made no mistake (Reviewers U6pP, L6Kb). We added a “not applicable” option for when the student made no mistake, and rescored all test dialogues, leaving those turns out of the average. All scores rise, but the ranking of tutors is essentially unchanged, and our tutor stays well ahead of TutorRL-7B. This shows that our method is robust regardless of whether the student made a mistake or not.

  4. Cross-student transfer mixed two changes (Reviewers U6pP, LeaK). The original transfer experiment changed the student and the problem set at the same time. We now test every tutor with each student on both problem sets, so only one thing changes at a time. Changing only the problem set has little effect. Our UserLM-trained tutor also transfers to OSIM, a student it never saw during training, and keeps far more of those students than its base model.

We would really appreciate it if you could let us know whether these responses address your concerns, and whether you have any remaining questions while there is still time to respond. If your concerns have been resolved, we would be grateful if you would consider updating your assessment. Thank you for your time and effort.

As we approach the end of the discussion period, we would appreciate any last comments, feedback or questions. We have tried to fully address all of the reviewer's concerns and questions. We have added extensive experiments and details to the revised paper. Please let us know if our responses are satisfactory, and we are happy to provide more details or clarifications if necessary.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.