acceptodds
Under review as a conference paper at ICLR 2027

RegulaTutor: A Theory-Informed Framework for Long-Term Personalized Learning

Abstract

Long-term personalized tutoring requires language models to translate a learner’s evolving history into appropriate instructional support. Existing LLM-based tutoring systems use memory and retrieval to personalize responses, but the pedagogical rationale linking historical evidence to teaching strategies is not always explicit. We introduce RegulaTutor, a theory-informed framework for long-term personalized tutoring that connects historical evidence, learner interpretation, and learner-facing instruction. Guided by self-regulated learning (SRL) research, its tutoring harness operationalizes a cognitive, affective, metacognitive, and motivational (CAMM) framework for deriving evidence-supported learner interpretations and teaching policies. The resulting policies retain links to their supporting evidence and are updated as new observations arrive. This staged workflow also supports systematic refinement through PairRoute, an offline procedure that uses intermediate execution records to distinguish harness, model, and mixed failures. PairRoute proposes and evaluates targeted harness revisions, then constructs reviewed stage-level supervision for confirmed model failures remaining under the selected harness. We focus this refinement on historical evidence acquisition, which underpins personalized instruction and allows outputs to be checked against learning records and reference answers. RegulaTutor improves historical-evidence accuracy, learner diagnosis, and judged teaching quality on average across evaluated model configurations. PairRoute results show an endpoint improvement in standalone-model historical-evidence accuracy over the refinement trajectory. Together, these results highlight RegulaTutor’s potential to support lifelong personalized education through instructional policies that evolve with learner histories and targeted improvements to the tutoring system.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Intelligent Tutoring Systems, Long-Term Personalized Tutoring, Learner Modeling, Self-Regulated Learning, Trajectory-Guided Refinement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dLnNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes RegulaTutor, an LLM tutoring harness to learn from learners’ long histories. It has three stages 1) an Evidence Resolution Loop (retrieval + mapping + resolution over history chunks), 2) a CAMM-guided learner interpretation step (four rules mapping observed evidence to instructional support), and 3) policy/directive compilation into the final prompt. A separate component, PairRoute, is an offline refinement loop in which two agents diagnose execution-trace failures as harness vs. model failures, propose revisions, and construct stage-level SFT targets.

Critical strengths

The paper identifies a very important in educational data, which is how to use a learner’s long interaction history to inform tutoring decisions. It also evaluates on a recent benchmark across many LLM backbones. On a small local model, structured retrieval over the history improves evidence acquisition and diagnosis compared with giving the model the raw history.

Critical weaknesses
  • The paper opens the framework by claiming it is based on self-regulated learning (SRL) and the CAMM framework. In the actual system, the theory is a few hand-written rules, and the authors themselves call them engineering heuristics. The ablation even shows that removing this part changes the results very little. The gains come from retrieving the right records from the learner’s history. So I would like to better understand, is this because the learning theory is not helpful in terms of llm+harness or it is because the framework instantiates the learning theory incorrectly?
  • I don’t find the motivation and introduction of PairRoute. Is it something proposed in the paper or it is an existing framework (since the authors cited Chen et al. when introducing it in sec2.2). If I understand correctly, PairRoute finds failures, fixes the harness, and then fine-tunes the model. PairRoute is not included in the ablation, and it is never compared with any alternative. The authors themselves say they cannot show that the routing step helps.
  • The paper says RegulaTutor helps across models while most of the gain comes from one small model, which is counted four times in the average. For stronger models, the system often does not help, and on some metrics it hurts (the main text shows the good numbers, and the bad numbers appear only in the appendix).
  • The paper is a bit hard to read and understand. The paper uses many terms, even starting in the abstract, SRL, CAMM, ERL, PairRoute, Router1/2, without clearly defining them or explaining how they fit together.
Questions for authors

Please see my response in weaknesses.

Reviewer_m1xkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RegulaTutor, a tutoring framework that connects learner-history evidence to CAMM-based learner interpretations and evidence-linked teaching policies. It also introduces PairRoute, an offline procedure that diagnoses execution failures, evaluates harness revisions, and constructs stage-level supervision for model updates. Experiments primarily use LongTutor, with component ablations and crossed model–harness evaluations, supplemented by exploratory TeachArena results. The framework improves several history-processing and tutoring metrics, although the effects vary substantially across backbones and configurations.

Critical strengths

The framework makes the connection between evidence and instructional decisions explicit. Keeping supporting records and distinguishing observed difficulties from unsupported claims about motivation or affect are sensible choices for a tutoring system. This makes personalization decisions easier to inspect.

The evaluation usefully separates evidence acquisition, diagnosis, and teaching quality. For example, removing compiled directives improves Evidence but reduces Teaching, showing why better historical recall should not automatically be interpreted as better tutoring. The crossed evaluations also expose compatibility problems rather than reporting only favorable configurations.

Critical weaknesses

The benefit of PairRoute for the complete system remains unclear. In Table 4, standalone-model Evidence improves from 20.0 to 28.2, but under the selected H₁ it decreases from 56.2 at M₀ to 43.7 at M₃. Table 17 further shows that H₁ performs worse than H₀ on Evidence at every Qwen checkpoint, with all four paired intervals excluding zero. These results do not negate the standalone-model gains, but they leave insufficient evidence that the selected refinement trajectory produces a better deployed tutor. The authors acknowledge this limitation, but it still weakens the practical contribution of PairRoute.

The adaptive-teaching mechanism is only partly validated. LongTutor reconstructs learner state from independent history snapshots, so the main evaluation does not test whether a persistent policy changes appropriately as a learner improves or encounters new difficulties. Table 2 provides some support for the policy pathway, but the no-CAMM condition also removes downstream policy/directive processing (Table 11), leaving the theory-specific contribution unresolved. The evidence is therefore stronger for the combined history-processing pipeline than for the proposed mechanism of ongoing instructional adaptation.

Optional presentation suggestion:The conceptual figures are visually elaborate, but offer limited implementation detail. In contrast, the experimental results rely heavily on tables. A plot comparing development and held-out performance across refinement rounds would make the contrasting trends easier to follow.

Questions for authors

What selection criteria justified retaining H₁, and which evaluated model–harness pair best demonstrates the practical benefit of PairRoute? Please reconcile the development-set improvements with the Gold1000 regressions in Tables 4 and 17.

Which existing results demonstrate that persistent policy updates appropriately change instructional support as learner evidence changes? Controlled examples or a small replay study would help distinguish ongoing adaptation from reconstruction on independent snapshots.

Reviewer_uEd9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RegulaTutor, a personalized tutoring framework that transforms learners' historical learning records into personalized instructional decisions through historical evidence retrieval, CAMM-guided learner interpretation, and evidence-linked teaching policies, while supporting policy updates across interactions. The authors also introduce PairRoute, which distinguishes harness failures from model failures based on agent execution traces and addresses them through targeted harness revisions and stage-level supervised fine-tuning (SFT), respectively. Experiments on LongTutor include comparisons across multiple models, component ablations, and crossed model–harness evaluations, supplemented by transfer evaluations on TeachArena. The results demonstrate improvements in certain evidence acquisition and learner diagnosis metrics. However, the framework's advantages over existing tutoring systems and its robustness across different tasks and environments remain unclear.

Critical strengths

S1. Integrating Educational Theory with Historical Evidence-Driven Instructional Decision-Making

This paper incorporates the Cognitive, Affective, Metacognitive, and Motivational (CAMM) framework into personalized tutoring, establishing an explicit connection between historical evidence, learner state interpretation, and instructional strategies. Furthermore, the proposed Evidence-linked Policy mechanism supports dynamic updates to teaching strategies, enhancing the interpretability of instructional decisions.

Critical weaknesses

W1. Insufficient Comparisons with Existing Personalized Tutoring Systems to Demonstrate the Proposed Method's Relative Advantages**

The paper primarily compares RegulaTutor against model-only baselines, while direct comparisons with existing personalized tutoring systems, such as DeepTutor, are very limited. Therefore, the current experiments do not provide sufficient evidence to establish the proposed framework's advantages over existing methods in terms of personalized tutoring quality.

W2. Limited Cross-Model Generalization and Robustness, with Notable Performance Degradation in Several Experiments**

RegulaTutor exhibits inconsistent performance across different backbone models, with noticeable declines in Evidence, Diagnosis, or Teaching scores for some models. Furthermore, RegulaTutor generally underperforms the Direct Baseline on TeachArena's tool-based workflow tasks. Table 4 also shows that the final harness H1 achieves lower Evidence accuracy than H0 across all Qwen checkpoints. These results raise concerns about the framework's generalization ability and robustness, which are not sufficiently discussed or explained in the main text.

Questions for authors

Q1. Why is DeepTutor evaluated only with DeepSeek-v4-pro? Could the authors provide additional comparisons using other backbone models to further demonstrate RegulaTutor's relative advantages?

Q2. RegulaTutor exhibits performance degradation in Evidence, Diagnosis, or Teaching on certain backbone models. Could the authors provide a more detailed analysis of the reasons behind these performance regressions?

Q3. RegulaTutor generally underperforms the Direct Baseline on TeachArena's tool-based workflow tasks. To what extent are these performance degradations caused by interface issues, formatting constraints, and execution failures? After controlling for these factors, does the method still exhibit substantial degradation in transfer performance?

Q4. Table 4 shows that the final harness H1 achieves lower Evidence accuracy than H0 across all four Qwen checkpoints. Could the authors clarify the rationale for selecting H1 and explain the observed performance degradation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.