acceptodds
Withdrawn from ICLR 2027

State-conditioned Pedagogical Adaptive Routing for Exercise Recommendation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Exercise recommendation is inherently non-stationary: learner representations are unreliable under sparse interaction histories, the set of pedagogically appropriate exercises changes with knowledge mastery, and the desired balance among consolidation, exploration, and challenge evolves throughout the learning process. However, existing methods often optimize multiple objectives using globally fixed trade-offs and handle representation learning, candidate retrieval, and ranking as separate components, limiting their ability to adapt to both cold-start uncertainty and changing cognitive states. To address these limitations, we propose SPAR, a state-conditioned pedagogical adaptive routing framework that treats exercise recommendation as three interdependent decisions. First, a prototype-guided alignment module regularizes sparse learner representations using behavioral priors derived from similar active learners. Second, a ZPD (Zone of Proximal Development)-inspired retrieval module adapts the target challenge level according to recent performance and constructs a personalized candidate set within an appropriate difficulty range. Third, a state-conditioned dynamic routing network dynamically coordinates relevance, diversity, and difficulty within the retrieved candidate space according to the learner's current cognitive state. Experiments on three public datasets show that SPAR consistently improves ranking performance over strong exercise-recommendation baselines, with particularly pronounced gains for learners with sparse interaction histories.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Exercise Recommendation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NZ1DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPAR, a state-conditioned framework for exercise recommendation that addresses three related problems: unreliable learner representations under sparse interaction histories, difficulty-aware candidate retrieval, and adaptive balancing of multiple recommendation objectives.

SPAR consists of three stages. First, prototype-guided alignment uses representations of active learners to improve those of learners with sparse histories. Second, ZPD-inspired retrieval adjusts the target difficulty according to recent response accuracy and estimated concept mastery. Third, a state-conditioned routing network combines relevance, difficulty, and diversity scores to rank candidate exercises.

The authors evaluate SPAR on three public datasets against eight baselines. The experiments include overall performance comparisons, component ablations, cold-start analysis, parameter sensitivity, and controlled comparisons of retrieval and re-ranking modules. SPAR achieves consistent improvements in offline recommendation metrics, particularly for learners with sparse interaction histories.

Critical strengths
  1. A reasonable approach to coordinating different stages of exercise recommendation. The paper recognizes the dependencies between learner representation, candidate retrieval, and re-ranking. Inaccurate learner representations may lead to inappropriate candidate sets, which cannot be fully corrected by downstream re-ranking. Incorporating learner-state information across these stages is a meaningful approach to improving adaptive exercise recommendation.

  2. Consistent performance improvements with useful component-level evidence. SPAR improves NDCG@10 over the strongest baselines on all three datasets (Table 2). The ablation studies (Table 3) show that each module contributes to performance. More importantly, the retrieval/re-ranking comparisons in Appendix A.3 (Table 4) show that both proposed components can improve existing recommendation pipelines. The cold-start analysis further supports the benefit of prototype alignment for learners with sparse histories.

These experiments provide credible evidence for the effectiveness of the proposed system in offline recommendation.

Critical weaknesses
  1. The training signal for the diversity component is unclear and potentially inconsistent with its definition. In Section 3.5, Eqs. (9)–(11), diversity gain is defined as additional concept coverage relative to a learner's historical exercises. However, Section 4.1 states that candidates containing unseen concepts are excluded from the BCE loss. Under these definitions, candidates used for BCE training would have zero marginal historical concept coverage, leaving the diversity scorer without a candidate-dependent coverage signal from these examples. The paper does not explain whether additional supervision or implementation details resolve this issue. This concern directly affects the proposed multi-objective routing mechanism, which is a central contribution.

  2. The experiments do not sufficiently isolate the benefits of state-conditioned adaptation. The meso ablation replaces ZPD-inspired retrieval with global retrieval, rather than comparing dynamic and fixed difficulty thresholds within the same retrieval framework. Similarly, the micro ablation compares state-conditioned routing with uniform weights, but not with learned state-independent weights. Furthermore, Figure 4 demonstrates changes in routing weights over time, but does not establish that these changes reflect appropriate pedagogical priorities. Given that existing methods already employ multi-stage exercise recommendation, stronger evidence is needed to establish the independent contribution of the proposed state-conditioned mechanisms.

Questions for authors
  1. In the implementation, how does the diversity scorer receive informative training signals for nonzero concept coverage gains, given the exclusion of candidates containing unseen concepts from the BCE loss? Are there additional losses or implementation details not described in the paper?

  2. Could the authors provide controlled comparisons of (a) dynamic versus fixed difficulty thresholds and (b) state-conditioned versus learned global routing weights, while keeping the remaining components unchanged? These results would help isolate the contribution of state adaptation.

1 comment

Thank you very much for your further comments and advice. We will answer your questions and weaknesses point by point:

W1 & Q1: Thanks for your suggestion. In the original submission, Eq. (10) defined marginal coverage relative to the learner’s history. Under that definition, every BCE-eligible exercise has zero marginal coverage gain. Eqs. (9)–(11) in the revised manuscript define coverage over the retrieved candidate pool EcE_c, as used in our experiments. For an exercise eje_j, the contribution is

Δj=Coverage⁡(Ec)−Coverage⁡(Ec∖{ej}).\Delta_j=\operatorname{Coverage}(E_c)-\operatorname{Coverage}(E_c\setminus\{e_j\}).

Under this definition, excluding candidates involving unseen concepts from BCE supervision does not necessarily make their diversity gains zero. For example, an exercise that uniquely covers a previously practiced concept can pass the BCE mask if all its concepts have observed responses. Removing that exercise reduces candidate-list coverage.

We multiply the coverage feature elementwise by the learner’s concept preference vector and pass the result to the diversity scorer. The diversity score contributes to the fused recommendation score used by BCE on eligible candidates. A nonzero coverage feature can therefore carry candidate-dependent gradients to the diversity scorer and the corresponding preference parameters. This learning signal comes from the existing BCE objective; no additional diversity-specific loss is required for this mechanism.

Section 4.1 now specifies that coverage is computed over the full retrieved pool before the BCE mask is applied.

W2 & Q2: Thanks for your suggestion. Table 3 evaluates module-level changes: the meso ablation replaces ZPD-inspired retrieval with global retrieval, while the micro ablation replaces state-conditioned routing with uniform weights. Figure 4 illustrates how routing weights evolve. To isolate the contributions of state adaptation to offline recommendation quality, we compare two sets of matched variants.

(1) Adaptive versus fixed difficulty targets. We compare the adaptive target in Eq. (4) with a single fixed difficulty target selected on the validation set. We use the same mastery estimates, exercise search space, retrieval scoring rule, candidate budget, and downstream re-ranking protocol. The candidate lists can differ because the target determines which exercises are retrieved.

Supplementary comparison of adaptive and fixed difficulty targets

DatasetRetrieval strategyNDCG@10LC-F1@10Recall@10
XES3G5MFixed difficulty target87.3553.1077.60
XES3G5MAdaptive difficulty target88.2954.0078.89
Assist2009Fixed difficulty target84.7053.9080.40
Assist2009Adaptive difficulty target86.5354.1380.94
Assist2012Fixed difficulty target82.2044.8064.70
Assist2012Adaptive difficulty target83.0447.2765.32

(2) State-conditioned versus learned global routing weights. We compare the proposed state-conditioned gate with a learned global softmax weight vector shared across learners. Both variants are trained independently under the same training and validation protocols, using identical candidate lists, scorer architectures, and training objectives.

Supplementary comparison of state-conditioned and learned global routing weights

DatasetRouting strategyNDCG@10LC-F1@10Recall@10
XES3G5MLearned global weights87.4053.3077.80
XES3G5MState-conditioned weights88.2954.0078.89
Assist2009Learned global weights84.2053.9080.70
Assist2009State-conditioned weights86.5354.1380.94
Assist2012Learned global weights82.5046.0064.90
Assist2012State-conditioned weights83.0447.2765.32

On XES3G5M, Assist2009, and Assist2012, the adaptive difficulty target improves downstream NDCG@10 over the fixed target by 0.94, 1.83, and 0.84 points, respectively. State-conditioned routing improves NDCG@10 over learned global weights by 0.89, 2.33, and 0.54 points, respectively. The matched comparisons quantify the contribution of each adaptation to offline recommendation performance under the corresponding protocol.

Reviewer_sAXYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPAR, a state-conditioned framework for personalized exercise recommendation. The framework consists of three main components: prototype-guided learner representation alignment for sparse interaction histories, ZPD-inspired adaptive difficulty-based candidate retrieval, and state-conditioned multi-objective re-ranking that balances relevance, difficulty, and diversity. The authors evaluate SPAR on three public datasets and report improvements over several existing exercise recommendation methods.

Critical strengths
  1. Clear Organization: The paper is well-structured and clearly presented.
  2. High-Quality Figures: The figures are visually appealing and well-designed, effectively illustrating the motivation, overall architecture, and key components of the proposed method.
  3. Comprehensive Experiments: The paper provides extensive experiments on three public datasets, including comparisons with multiple baselines, ablation studies, parameter sensitivity analysis, and cold-start evaluations.
Critical weaknesses
  1. Substantial Overlap in Problem Formulation with NR4DER. The three major problems addressed by SPAR closely correspond to those already investigated in NR4DER (SIGIR 2025): sparse learner interactions and long-tailed distributions, difficulty-aware exercise selection, and personalized diversity-aware re-ranking. SPAR mainly reformulates these existing challenges from a state-conditioned perspective without clearly establishing a fundamentally new research problem.
  2. Limited Methodological Novelty Compared with NR4DER. SPAR largely follows the same technical pipeline as NR4DER. Specifically, it replaces the original representation enhancement mechanism with K-means-based prototype alignment, extends static difficulty thresholds to dynamic thresholds based on recent accuracy, and introduces an additional difficulty scoring branch and dynamic gating mechanism into a similar diversity-aware re-ranking framework. These modifications appear to be incremental extensions rather than substantial methodological innovations.
  3. Insufficient Experimental Evidence to Establish Novelty. Although SPAR achieves performance improvements over NR4DER, the experiments do not sufficiently demonstrate that these gains originate from fundamentally new methodological contributions. More direct comparisons with NR4DER enhanced by individual components, such as adaptive difficulty calibration and dynamic gating, are needed to establish whether SPAR offers meaningful advances beyond straightforward modifications to the existing framework.
  4. Potential Inconsistency Between Diversity Modeling and Training Supervision. In Eqs. (9)–(11), the diversity score is based on the marginal coverage gain of previously unseen concepts. However, Section 4.1 states that candidates involving unseen concepts are excluded from BCE training. Under the stated definitions, this may cause the diversity gain to become zero for supervised training candidates, preventing the diversity scorer from learning meaningful candidate-specific preferences. The authors should clarify how this module receives effective training signals.
Questions for authors
  1. Problem Formulation: Given that NR4DER has already addressed sparse learner interactions, difficulty-aware exercise selection, and personalized diversity-aware re-ranking, what fundamentally new research problem does SPAR aim to solve? How does its state-conditioned formulation differ from the problems investigated in NR4DER?
  2. Methodological Novelty: Beyond replacing or extending individual components, what are the fundamental methodological differences between SPAR and NR4DER? Why should the proposed prototype alignment, adaptive difficulty threshold, and dynamic gating mechanism be considered substantial methodological innovations rather than incremental modifications?
  3. Experimental Validation: Could the authors provide more direct comparisons with enhanced NR4DER variants, such as NR4DER with adaptive difficulty calibration or state-conditioned dynamic gating? Such experiments would help clarify whether SPAR offers substantial improvements beyond straightforward extensions of the existing framework.
  4. Diversity Modeling and Supervision: According to Eqs. (9)–(11), the diversity score relies on the marginal coverage gain of previously unseen concepts. However, Section 4.1 states that candidates involving unseen concepts are excluded from BCE training. Would this cause the diversity gain to be zero for supervised training candidates? If so, how does the diversity scorer learn meaningful candidate-specific diversity preferences?
Reviewer_jWEkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPAR, a state-conditioned exercise recommendation framework that organizes recommendation into three stages: macro-level prototype-guided alignment for sparse-history learners, meso-level ZPD-inspired candidate retrieval, and micro-level state-conditioned multi-objective re-ranking over relevance, difficulty, and diversity. Experiments on three public datasets compare SPAR with eight baselines and report improvements, with additional appendix analyses on cold-start performance and retrieval/re-ranking decoupling.

Critical strengths
  1. The paper provides a reasonably comprehensive empirical evaluation across three public datasets and eight baselines, and the appendix adds useful analyses beyond the main comparison, including retrieval/re-ranking decoupling and cumulative cold-start evaluation. These analyses make the empirical claims more credible than a single main-table comparison would.

  2. The hierarchical formulation that couples sparse-learner representation, candidate construction, and multi-objective re-ranking is a useful way to frame exercise recommendation. In particular, treating representation uncertainty and downstream pedagogical trade-offs as linked decisions is a sensible contribution, even if individual components draw on existing ideas.

Critical weaknesses
  1. The central claim that state-conditioned routing learns an appropriate trade-off among relevance, difficulty, and diversity is not directly validated. Figure 4 shows routing trajectories, and the entropy regularizer encourages non-degenerate weights, but the paper does not compare the learned routing against fixed, random, or rule-based weighting, nor does it isolate whether the routing weights improve pedagogical outcomes beyond the three objective scores themselves. As a result, the evidence supports that the weights change over time, but not clearly that they change in a better way.

  2. The cold-start prototype alignment risk is under-analyzed. Eq. (3) gives shorter histories stronger alignment weights, so an extremely sparse learner may be pulled toward a prototype before the assignment is reliable. Appendix A.4 reports aggregate cold-start gains, but the paper does not report prototype assignment confidence, misassignment rates, or an ablation of the dynamic alignment weight wjw_j. Aggregate improvements therefore do not fully rule out that some very short-history learners are aligned to an inappropriate prototype.

Questions for authors
  1. For learners with at most 5 interactions, how much of SPAR's gain comes from prototype alignment versus state-conditioned diversity routing? Can the authors report a comparison among full SPAR, SPAR without macro alignment, and SPAR with fixed diversity weights for this subgroup?
  2. In the main comparison, SPAR uses its own ZPD-inspired retrieval while baselines use their original retrieval strategies. Although the appendix provides a decoupling experiment on XES3G5M, can the authors report all methods under a unified candidate pool, or decompose the gains from retrieval and re-ranking across all three datasets? This would clarify whether the main-table gains are mainly due to candidate construction rather than the proposed routing mechanism.
2 comments

Thank you very much for your further comments and advice. We will answer your questions and weaknesses point by point:

W1: Thanks for your suggestion. Table 3 includes a uniform-weight control. As described in Section 4.3, the w/o micro variant uses α=β=γ=1/3\alpha=\beta=\gamma=1/3. On XES3G5M, full SPAR improves NDCG@10 from 86.62 to 88.29 over uniform fusion.

We also compare state-conditioned routing with a learned global softmax weight vector shared across learners. Both variants are trained separately under the same training and validation protocols, using identical candidate lists, scorer architectures, and training objectives. On XES3G5M, learned global weights yield 87.40, and state-conditioned routing yields 88.29, a gain of 0.89 points. The corresponding gains are 2.33 points on Assist2009 and 0.54 points on Assist2012. Figure 4 illustrates how the weights vary, and these comparisons quantify the effect of state conditioning on offline recommendation performance. Assessing actual learning outcomes requires measuring those outcomes directly.

Q2: Thanks for your suggestion. Table 2 compares complete recommendation pipelines, including SPAR’s candidate construction. To assess retrieval and re-ranking separately, we extend the stage-swapping comparison in Appendix A.3 to Assist2009 and Assist2012. For each dataset, we cross NR4DER and SPAR retrieval with NR4DER and SPAR re-ranking. We use the same candidate pool when comparing re-rankers and the same re-ranking method when comparing retrievers.

Supplementary comparison of retrieval and re-ranking across three datasets

DatasetRetrievalRe-rankingNDCG@10LC-F1@10Recall@10
XES3G5MNR4DERNR4DER83.4152.8175.21
XES3G5MNR4DERSPAR84.9153.0277.38
XES3G5MSPARNR4DER85.3353.4578.01
XES3G5MSPARSPAR88.2954.0078.89
Assist2009NR4DERNR4DER77.5437.1456.93
Assist2009NR4DERSPAR81.2045.5068.20
Assist2009SPARNR4DER82.6046.8069.50
Assist2009SPARSPAR86.5354.1380.94
Assist2012NR4DERNR4DER80.6446.3649.16
Assist2012NR4DERSPAR81.7546.8555.80
Assist2012SPARNR4DER82.1046.7058.90
Assist2012SPARSPAR83.0447.2765.32

With either candidate pool, SPAR’s re-ranker improves all three metrics on each dataset. SPAR retrieval also improves all three metrics with either re-ranking method. The full SPAR pipeline achieves the highest scores in these comparisons. The matched results show that both stages contribute to the gains and quantify the re-ranking contribution at the module level.

Thank you for the detailed response and for running all the requested additional experiments.

The new results—including the comparison against learned global weights (W1), the stage-swapping decoupling across all three datasets (Q2), and the confidence and subgroup analyses for sparse learners (W2 & Q1)—have satisfactorily addressed my concerns regarding experimental rigor and isolated component effectiveness.

However, as the authors also acknowledge, these results primarily validate improvements on offline ranking metrics rather than directly establishing pedagogical validity or real learning gains. Combined with the relatively incremental nature of the individual components, I still feel the contribution falls slightly short of the acceptance bar.

Therefore, while I appreciate the thorough revisions, I will maintain my original rating.

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

Related papers

Discussion (0)

Sign in to comment.