acceptodds
Under review as a conference paper at ICLR 2027

CT-LFSSM: Multiscale Continuous-Time State-Space Learning and Forgetting in Knowledge Tracing

Abstract

Knowledge Tracing (KT) aims to model students' evolving knowledge from historical interactions and predict their future performance. However, existing methods often encode irregular temporal gaps as auxiliary inputs rather than explicitly evolving latent knowledge over elapsed time. They also tend to absorb forgetting, response evidence, and learning into a single state transition, making these effects difficult to distinguish. Independently parameterized items further limit statistical sharing across items associated with the same knowledge component. To address these limitations, we propose CT-LFSSM, a Continuous-Time Learning–Forgetting State-Space Model. CT-LFSSM represents each knowledge component with fast-, medium-, and slow-timescale latent states. It propagates these states over real elapsed time through personalized exponential decay, with half-lives constrained to increase from the fast to the slow timescale. Before each response, an event-conditioned convex gate aggregates the three states into a current mastery estimate. After observing the response, CT-LFSSM uses the prediction residual to revise the inferred pre-interaction state, while separately consolidating practice-induced learning gains across multiple timescales. This separation assigns distinct update paths to forgetting, response-driven belief revision, and learning. For response prediction, CT-LFSSM further decomposes item effects into shared knowledge-component-level parameters and regularized item-specific residuals. Experiments on four real-world datasets show that CT-LFSSM achieves state-of-the-art performance across all datasets, while remaining robust across varying temporal gaps and student interaction sequence lengths.

Primary area:
learning on time series and dynamical systems
Keywords:
Knowledge Tracing, Continuous-Time Modeling, State-Space Models, Multi-Timescale Learning, Forgetting Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ADw6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a refreshing and well-engineered perspective on Knowledge Tracing by bridging continuous-time state-space dynamics, psychometric item response theory, and causal update decoupling. The benchmark re-evaluation is exceptionally rigorous, and the paper is well-written. However, the empirical justification for its central claim—that continuous multiscale dynamics drive the main performance leap—remains partially compromised by the dominant contribution of the episodic memory module and the lack of direct single-vs-multiscale ablations.

Critical strengths
  1. The problem motivation is clear and relevant. Explicitly modeling continuous-time decay over latent mastery states is more principled than treating Δt only as an auxiliary feature.

  2. The combination of ordered multi-timescale states, event-conditioned gating, and predict-observe-update separation is conceptually meaningful. In particular, separating response-driven evidence revision from practice-induced learning gain is a useful modeling idea.

  3. The causal execution order is carefully described, and the appendix provides an online inference procedure. If implemented as stated, it reduces the risk of future-label leakage.

  4. The experimental protocol for the first three datasets is relatively rigorous: the authors rerun 14 baselines under a pyKT-compatible protocol, use five-fold validation, and provide ablations, sequence-length slices, and time-gap slices.

  5. The paper is generally well structured, with a complete symbol table, implementation details, reproducibility statement, and an explicit discussion of some limitations such as statistical identifiability.

Critical weaknesses
  1. The central claim is not sufficiently supported by targeted ablations. The paper claims that multi-timescale continuous-time state evolution is the core contribution, but Table 4 does not directly ablate: single-scale vs. two-scale vs. three-scale states; continuous-time decay vs. Δt-only features; learnable gating vs. fixed mixing; the hard ordering constraints; and the separation of evidence revision from learning gain. Without these controls, the main claim remains under-supported.

  2. Episodic memory and hierarchical 4PL appear to dominate the reported gains. In Table 4, removing episodic memory and hierarchical item-response causes the largest drops, while removing hierarchical state update causes only a small drop. This raises a legitimate concern that performance may be driven mainly by an external memory module and 4PL parameterization rather than by the proposed multi-timescale continuous-time state-space mechanism. A controlled experiment attaching episodic memory to strong baselines such as simpleKT or AKT is needed.

  3. Direct continuous-time and multi-scale baselines are missing, and the XES3G5M comparison is not fully apples-to-apples. CT-NCM is the most direct continuous-time forgetting baseline but is absent from the main comparison; DMKT/FlucKT/TFKT are also not compared. On XES3G5M, CT-LFSSM uses timestamp and multi-KC inputs, while several published baselines may not have the same input conditions. Strong baselines should be rerun under matched timestamp/multi-KC settings.

  4. Statistical significance and time-modeling evidence are insufficient. The paper fixes a single random seed (3407), and the reported five-fold standard deviations are fold-level rather than seed-level. Many gains are only 0.3–1.25 percentage points, close to baseline variance. Multi-seed runs, paired significance tests, or bootstrap confidence intervals are needed. In addition, ASSIST2009 sets Δt = 0 and cannot support the continuous-time claim; time-gap results should be added on ASSIST2012 and XES3G5M.

Questions for authors

1.Can you provide targeted ablation results comparing single-scale, two-scale and three-scale states, continuous-time decay against Δt-only features, learnable gating against fixed mixing, the effect of hard ordering constraints, and the separation of evidence revision from learning gain to validate the central claim of multi-timescale continuous-time state evolution?

2.Could you report controlled experiments that attach your episodic memory module to strong baselines such as simpleKT and AKT, to clarify whether the performance gains mainly come from episodic memory and hierarchical 4PL rather than the proposed multi-timescale continuous-time state-space mechanism?

3.Will you add comparisons against direct continuous-time and multi-scale baselines including CT-NCM, DMKT, FlucKT and TFKT, and rerun the relevant baselines on XES3G5M under matched timestamp and multi-KC input settings for a fair comparison?

4.Can you supply experimental results with multiple random seeds and paired significance tests or bootstrap confidence intervals to verify whether the reported performance improvements are statistically significant? Could you also present time-gap experimental results on ASSIST2012 and XES3G5M to support the continuous-time modeling claim, given that Δt is fixed to 0 in ASSIST2009?

6 comments

We sincerely thank the reviewer for the thorough assessment of the central contribution, component attribution, baseline fairness, and statistical reliability of CT-LFSSM. We address W1–W4 and Q1–Q4 in the order raised. Our additional evidence comprises five-fold active-timescale, fixed-mixing, unconstrained-half-life, and unified-update controls; an explicit-decay ablation on real-timestamp data; a four-model analysis of response-driven belief revision and practice-induced learning; additional direct temporal baselines; and paired, student-cluster bootstrap analysis.

W1 — Targeted evidence for multi-timescale continuous-time state evolution

CT-LFSSM maintains three KC-level latent states whose elapsed-time evolution is governed by student–KC-personalized half-lives. An event-conditioned convex gate reads these states before the current response; after observing the response, a signed evidence-based correction is followed by a nonnegative practice-induced learning transition. The corresponding mechanisms can be tested separately while retaining episodic memory and the hierarchical 4PL observation model. We evaluated the S1 Fast-only and S2 Fast + Slow active-scale controls on Algebra2005, retaining the same overall architecture and parameter budget. Inactive state slots were masked during readout and post-response updates, while episodic retrieval and hierarchical 4PL remained unchanged. Full CT-LFSSM uses all three timescales. The five-fold evaluation uses seed 3407 and 93,481 held-out test targets per fold.

Table R4.1. Algebra2005 Full three-scale result and newly trained S1/S2 active-scale controls (five-fold mean ± fold-level population SD).

Active KC timescalesAUC ↑ACC ↑
S1 — Fast0.8224 ± 0.00030.8097 ± 0.0013
S2 — Fast + Slow0.8266 ± 0.00050.8126 ± 0.0013
S3 — Fast + Medium + Slow (Full CT-LFSSM)0.8304 ± 0.00050.8168 ± 0.0016

Full CT-LFSSM achieves the highest reported mean AUC (0.8304) and ACC (0.8168). Relative to S1, the reported mean differences are +0.0080 AUC and +0.0071 ACC; relative to S2, they are +0.0038 AUC and +0.0042 ACC. Among the matched active-scale controls, S2 also improves AUC over S1 in all five folds (+0.0042 on average). These results support using multiple retention horizons and the complete three-scale design. The remaining controls retain three latent-state slots and isolate the readout or response-update parameterization. FixedMix replaces the event-conditioned scale gate by the constant vector (1/3, 1/3, 1/3). NoHalfLifeOrder replaces cumulative positive increments by independent trainable log half-lives, retaining the same initialization and numerical bounds. UnifiedUpdate replaces the sequential evidence-revision/learning transitions by a single signed saturating update while reusing the other model modules.

Table R4.2. Algebra2005 Full CT-LFSSM and additional mechanism controls (five-fold mean ± fold-level population SD).

ConfigurationAUC ↑ACC ↑
Full CT-LFSSM0.8304 ± 0.00050.8168 ± 0.0016
FixedMix: constant-scale mixing0.8247 ± 0.00040.8094 ± 0.0026
NoHalfLifeOrder: independent half-lives0.8266 ± 0.00060.8133 ± 0.0016
UnifiedUpdate: single post-response transition0.8270 ± 0.00090.8139 ± 0.0015

Full CT-LFSSM attains the highest mean AUC and ACC in Table R4.2. Relative to FixedMix, the reported mean differences are +0.0057 AUC and +0.0074 ACC, supporting event-conditioned scale mixing. The half-life hierarchy guarantees Hu,n,k,f≤Hu,n,k,m≤Hu,n,k,sH_{u,n,k,f} \le H_{u,n,k,m} \le H_{u,n,k,s}, preserving the meanings of the three retention horizons. Personalized half-life medians are 0.6183/14.8189/281.2777 days on ASSIST2012 and 0.3540/6.8189/104.2161 days on Algebra2005. NoHalfLifeOrder and UnifiedUpdate further test the half-life parameterization and the choice of separate response-update transitions.

We separately isolated explicit latent-state decay on ASSIST2012 by setting forgetting_mode=none while retaining the other elapsed-time inputs, episodic aging, dynamic scale gate, and response-update architecture. Both models were selected on validation AUC and evaluated on the same 524,251 held-out test targets (fold 0).

Table R4.3. ASSIST2012 explicit latent-decay control, fold 0 (Full S3 versus S3-NoDecay).

ModelAUC ↑ACC ↑NLL ↓Brier ↓ECE ↓
Full S30.78290.75740.50200.16570.0155
S3-NoDecay0.78210.75660.50290.16610.0173
Full − NoDecay+0.0008+0.0008−0.0009−0.0004−0.0018

Because NoDecay still receives elapsed-time information outside the latent decay operator, this is a focused test of state evolution rather than a removal of timestamps from the complete model. It improves all five aggregate metrics in this fold. Together, the scale-count, gating, and NoDecay controls isolate three distinct aspects of the central state-space mechanism.

W2 — Separating the effects of episodic memory, hierarchical 4PL, and latent-state dynamics

Episodic retrieval, hierarchical item-response, and continuous-time KC states operate at different locations in the predict–observe–update computation. Episodic retrieval adds a bounded history-dependent observation-logit residual; hierarchical 4PL shares KC-level item parameters with item-specific residuals; the latent state evolves between interactions and is revised/updated after responses. Their contributions can therefore coexist rather than having to be attributed to a single module.

The submitted Table 4 establishes that episodic memory and hierarchical item-response are individually valuable. We present those effects explicitly, rather than treating the w/o hierarchical state update row as a test of the entire temporal-state model: that particular ablation replaces hierarchical evidence/consolidation parameterizations but retains three states and continuous-time decay.

Table R4.4. Submitted Table 4: actual held-out test AUC after component removal (five-fold mean ± fold-level population SD). Each w/o row denotes a model retrained without that component.

ModelASSIST2012 AUC ↑ASSIST2009 AUC ↑Algebra2005 AUC ↑
Full CT-LFSSM0.7833 ± 0.00050.7873 ± 0.00190.8304 ± 0.0005
w/o Episodic Memory0.7699 ± 0.00080.7511 ± 0.00060.8037 ± 0.0017
w/o Hierarchical Item-Response0.7730 ± 0.00060.7680 ± 0.00220.8135 ± 0.0014
w/o Cognition and Acquisition Gates0.7796 ± 0.00150.7818 ± 0.00160.8243 ± 0.0005
w/o Outcome Conditioning0.7822 ± 0.00030.7857 ± 0.00170.8267 ± 0.0007
w/o Hierarchical State Update0.7816 ± 0.00070.7863 ± 0.00170.8261 ± 0.0007

Removing episodic memory or hierarchical item-response consistently reduces AUC in all three datasets. The narrower hierarchical state-update ablation also lowers AUC, but retains the latent timescale count and the continuous-time decay mechanism; their specific effects are tested separately in Table R4.1–R4.3.

More focused attribution comes from the new experiments in Table R4.1–R4.3: episodic retrieval and hierarchical observation remain present while the active states, their mixing, or their explicit decay are

changed. Thus the performance difference attributable to these state-space operations is measured with the two strong prediction components held fixed.

We retrained ASSIST2012 controls that selectively remove the two post-response transitions, retaining episodic memory and hierarchical item-response modeling. NoRevision disables belief revision, NoLearning disables learning gain, and NoBoth disables both. Table R4.5 reports the submitted Full result and these five-fold controls.

Table R4.5. ASSIST2012 submitted Full result and five-fold post-response pathway controls (mean ± fold-level population SD; 524,251 test targets per fold).

ModelAUC ↑ACC ↑
Full CT-LFSSM0.7833 ± 0.00050.7582 ± 0.0003
NoRevision0.7817 ± 0.00040.7568 ± 0.0004
NoLearning0.7780 ± 0.00050.7553 ± 0.0002
NoBoth0.7762 ± 0.00030.7547 ± 0.0001

Full CT-LFSSM attains the strongest reported mean AUC and ACC in Table R4.5. Its mean AUC exceeds NoRevision by +0.0016, NoLearning by +0.0053, and NoBoth by +0.0071. When belief revision is absent, retaining practice-induced learning increases AUC by +0.0055 (NoRevision − NoBoth); when practice-induced learning is absent, retaining belief revision raises AUC by +0.0018 (NoLearning −NoBoth). These two conditional effects are positive in all five folds, showing how the signed evidence correction and nonnegative practice-acquisition transitions contribute to future predictions.

W3 — Direct temporal baselines and explicit input conditions on XES3G5M

We expanded the direct temporal/forgetting-oriented comparisons with CT-NCM, DMKT, FlucKT, and LefoKT to ASSIST2012, ASSIST2009, and Algebra2005. Table R4.6 reports completed five-fold AUC and ACC (mean ± fold-level population SD), complementing the original submission’s broader baseline evaluation.

Table R4.6. Expanded direct temporal and multiscale baseline results on three datasets (five-fold mean ± fold-level population SD).

ModelASSIST2012 ACC ↑ASSIST2012 AUC ↑ASSIST2009 ACC ↑ASSIST2009 AUC ↑Algebra2005 ACC ↑Algebra2005 AUC ↑
CT-NCM (2022)0.7569 ± 0.00250.7795 ± 0.00310.7355 ± 0.00150.7663 ± 0.00160.7994 ± 0.00060.8032 ± 0.0008
DMKT (2025)0.7284 ± 0.01080.7527 ± 0.01440.7131 ± 0.00510.7416 ± 0.00670.7886 ± 0.00210.7958 ± 0.0024
FlucKT (2025)0.7577 ± 0.00100.7809 ± 0.00190.7373 ± 0.00280.7731 ± 0.00330.8101 ± 0.00120.8280 ± 0.0010
LefoKT (2025)0.7559 ± 0.00310.7792 ± 0.00350.7400 ± 0.01000.7858 ± 0.00250.8109 ± 0.00090.8272 ± 0.0010
CT-LFSSM0.7582 ± 0.00030.7833 ± 0.00050.7419 ± 0.00140.7873 ± 0.00190.8168 ± 0.00160.8304 ± 0.0005

CT-LFSSM obtains the highest mean AUC and ACC among these direct baselines on all three datasets. For TFKT, our reproduction attempt did not yield a reliably validated comparison due to unresolved issues in the released implementation and its reported experimental configuration. Only completed baseline results are included.

For XES3G5M, the submitted Appendix B.3 reports CT-LFSSM under the official five-fold benchmark alongside 19 published benchmark baselines. These baseline values are not newly rerun results. To clarify the information-condition issue, we now show the benchmark’s explicit question-ID, KC-ID, and timestamp flags for representative methods.

Table R4.7. XES3G5M published official-benchmark results and reported input flags (submitted Appendix B.3).

ModelQuestion IDKC IDTimestampAUC ↑ACC ↑
DKT-FNoYesYes0.79400.8209
SAINTYesYesYes0.80740.8177
IEKTYesYesNo0.82800.8316
CT-LFSSMYesYesYes0.83150.8380

In the published benchmark comparison, CT-LFSSM exceeds IEKT by +0.0035 AUC and +0.0064 ACC. SAINT has all three input flags marked as available in the benchmark table; CT-LFSSM exceeds its reported results by +0.0241 AUC and +0.0203 ACC. The table makes the reported question, KC, and timestamp input categories visible alongside the published benchmark metrics.

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

This paper considers how time intervals between practice interactions may lead to different levels of forgetting and thus different probabilities of correct responses. The authors propose CT-LFSSM, which represents each knowledge concept with fast-, medium-, and slow-timescale states and personalizes students’ forgetting rates by adjusting the base half-lives. They also take question difficulty, discrimination, guessing, and slipping into consideration. An event-conditioned convex gate combines the three timescales to predict correctness on the next interaction. After observing a response, the knowledge state is updated through response-driven belief revision and practice-induced learning, intended to correct the previous state estimate and capture learning from practice, respectively.

Critical strengths
  1. The paper addresses an important aspect of knowledge tracing that students with similar practice histories may have different levels of retained knowledge because of different time intervals between interactions. Explicitly modeling how knowledge states evolve over elapsed time is well motivated.

  2. Maintaining multiple memory timescales is an intuitive way to represent knowledge with different retention durations. The half-life parameterization gives the forgetting dynamics a clear interpretation, while personalized half-lives allow the model to represent differences in retention across students and knowledge concepts.

Critical weaknesses
  1. The method section is difficult to follow. Many new symbols and equations are introduced before their roles and connections to the overall framework are sufficiently explained, while key definitions are deferred to the appendix. A clearer high-level explanation of each mechanism and its place in the prediction and update process would substantially improve readability.

  2. The model combines many components and modeling assumptions (e.g., three memory timescales, personalized half-lives, event-conditioned mixing, hierarchical item-response modeling, episodic memory, and separate revision and learning updates). While the authors ablate several components, the justification for other central design choices remains limited, and their individual contributions are not directly evaluated. It therefore remains unclear which mechanisms drive the reported improvements and whether the full architectural complexity is necessary.

Questions for authors
  1. Although the appendix provides the initial half-life values, how different are the learned fast, medium, and slow half-lives on each timestamped dataset? Their ordering is guaranteed, but could they converge to similar values? How do these half-lives compare with the observed time gaps? In particular, the counts or proportions of interactions in the intervals used in Table 3 would help clarify whether long gaps are common or relatively rare.

  2. What evidence supports the need for three memory timescales rather than a single timescale? The current ablations do not seem to directly address this choice.

  3. How are belief revision and practice-induced learning distinguished during training, given that both are optimized through response prediction? Could the two updates compensate for each other, and what evidence supports the benefit of keeping them separate?

6 comments

We sincerely thank the reviewer for the detailed and constructive comments on the clarity and necessity of CT-LFSSM’s modeling components. We address the concerns in the same order as the review: the presentation and execution flow of the method (W1), the contribution of its individual mechanisms (W2), the learned half-life ranges and their relationship to observed time gaps (Q1), the empirical need for multiple memory timescales (Q2), and the distinction between response-driven belief revision and practice-induced learning (Q3). Our responses integrate the submitted mathematical definitions and visualizations with new checkpoint-level analyses and completed, validation-selected held-out experiments. We have also revised the manuscript accordingly for your reference and sincerely appreciate your time and consideration.

W1 — Clarifying the roles and connections of the model components

The central organizing principle of CT-LFSSM is a single predict–observe–update process. Its state before interaction nn contains the previous post-response KC states Zu,n−1+Z^{+}_{u,n-1}, student profile hu,n−1h_{u,n-1}, and bounded episodic memory Mu,n−1\mathcal{M}_{u,n-1}. The model first predicts the current response using only information available beforehand; it then uses the observed response to update its online state. We will place the following compact roadmap before the detailed equations in Section 4 and retain the implementation details in Appendix A.

Table R3.1. Execution order, mathematical operation, and component role in CT-LFSSM.

StageEquation / operationPurpose and information boundary
1. Evolve stateZ+→Z−Z^{+}\rightarrow Z^{-}, Eq. (8)Apply elapsed-time decay through personalized fast-, medium-, and slow-timescale half-lives; no current response is used.
2. Read latent stateω=softmax⁡fω(⋅)\omega=\operatorname{softmax} f_{\omega}(\cdot), Eqs. (9)–(10)Combine three KC states with an event-conditioned convex gate.
3. Form observationHierarchical 4PL + episodic logit, Eqs. (11)–(14)Combine KC-shared item parameters and past-only episodic retrieval to predict pu,np_{u,n}.
4. Observe responseϵu,n=yu,n−pu,n\epsilon_{u,n}=y_{u,n}-p_{u,n}The observed response first becomes available after the prediction.
5. Revise beliefZ−→Z~Z^{-}\rightarrow\tilde{Z}, Eq. (15)Correct the inferred pre-interaction knowledge state using signed prediction evidence.
6. Learn from practiceZ~→Z+\tilde{Z}\rightarrow Z^{+}, Eqs. (16)–(17)Add nonnegative practice-induced learning via multi-timescale consolidation.
7. Persist historyUpdate hh, append to M\mathcal{M}Make the updated profile and memory available only to later interactions.

The division between the two prediction-time components is also explicit. The three-scale gate reads the continuously evolving KC-level latent state, while episodic retrieval contributes a bounded residual to the observation logit, not to the latent state itself. Thus, the model separates what it infers about current knowledge from how past specific interactions adjust the current response prediction.

This roadmap connects Section 4.1 (state evolution), Sections 4.2.1–4.2.2 (pre-response readout and observation), and Section 4.2.3 (post-response transitions). We will also move short definitions of the most frequently used symbols—Z±Z^{\pm}, Hu,n,k,rH_{u,n,k,r}, ωu,n,r\omega_{u,n,r}, ϵu,n\epsilon_{u,n}, Ek,rE_{k,r}, and Gu,n,rG_{u,n,r}—into the main text, while leaving low-rank constructions and numerical parameterizations in Appendix A. The existing Figure 2 already depicts the process; the roadmap makes the temporal and mathematical dependencies explicit.

W2 — Identifying the contributions of central design choices

The architecture contains complementary mechanisms with distinguishable functions: personalized continuous-time latent-state evolution, event-conditioned scale readout, response-dependent state transitions, KC-shared item-response modeling, and past-only episodic retrieval. Beyond the submitted module ablations, we have now performed focused interventions on active scale count, dynamic mixing, explicit state decay, and the two post-response branches.

Table R3.2. Submitted Table 4: held-out test AUC (five-fold mean ± fold-level population SD). Each w/o row reports the performance of a model trained without the specified component; lower AUC than Full indicates a performance reduction upon its removal.

ModelASSIST2012 AUC ↑ASSIST2009 AUC ↑Algebra2005 AUC ↑
Full CT-LFSSM0.7833 ± 0.00050.7873 ± 0.00190.8304 ± 0.0005
w/o Episodic Memory0.7699 ± 0.00080.7511 ± 0.00060.8037 ± 0.0017
w/o Hierarchical Item-Response0.7730 ± 0.00060.7680 ± 0.00220.8135 ± 0.0014
w/o Cognition and Acquisition Gates0.7796 ± 0.00150.7818 ± 0.00160.8243 ± 0.0005
w/o Outcome Conditioning0.7822 ± 0.00030.7857 ± 0.00170.8267 ± 0.0007
w/o Hierarchical State Update0.7816 ± 0.00070.7863 ± 0.00170.8261 ± 0.0007

Full CT-LFSSM achieves higher mean AUC than every listed ablation across all three datasets. For example, retaining episodic memory improves mean AUC by +0.0134, +0.0362, and +0.0267 on ASSIST2012, ASSIST2009, and Algebra2005, respectively; retaining hierarchical item-response modeling yields gains of +0.0103, +0.0193, and +0.0169.

The original hierarchical-update ablation changes the evidence-strength and consolidation parameterizations while retaining three states, elapsed-time decay, and the event-conditioned scale gate. The new experiments therefore target distinct questions that this ablation did not isolate:

  1. Number of active latent timescales: S1 versus S2 versus S3, with episodic memory and item-response modeling unchanged (Q2; Table R3.5).
  2. Event-conditioned readout: Full S3 versus three-scale uniform FixedMix on Algebra2005 (Q2).
  3. Explicit state decay: Full S3 versus S3-NoDecay on timestamped ASSIST2012, with other elapsed-time inputs retained.
  4. Post-response branch contributions: Full, NoRevision, NoLearning, and NoBoth on ASSIST2012 (Q3).

These results quantify the effects of different state-space operations without removing the episodic retrieval or hierarchical observation capacity in the targeted controls. In particular, a large episodic-memory ablation effect and a measurable contribution from multiscale state dynamics are compatible: the two components affect different points of the prediction–update computation.

Q1 — How different are the learned half-lives, and how do they compare with observed gaps?

We examined the learned KC-level base half-lives and student–KC-personalized half-lives of CT-LFSSM on ASSIST2012 and Algebra2005, using checkpoints selected on the validation sets. Table R3.3 reports their actual magnitudes and the paired ratios between adjacent timescales, allowing the extent of separation to be assessed directly.

Table R3.3. Learned half-life medians (days) and medians of paired adjacent-scale ratios. Personalized values summarize sampled student–KC pairs; the ratios are computed per pair, not by dividing the displayed marginal medians.

DatasetHalf-lifeFast (days)Medium (days)Slow (days)Median M/FMedian S/M
ASSIST2012Base KC0.28596.0806102.172521.138516.7881
ASSIST2012Personalized0.618314.8189281.277723.834818.9909
Algebra2005Base KC0.24574.495463.713918.343214.1945
Algebra2005Personalized0.35406.8189104.216119.253015.0202

The learned timescales do not collapse to similar retention horizons: among the examined samples, no adjacent-scale ratio falls below 2, and the typical fast-to-slow span exceeds two orders of magnitude. These measurements provide quantitative support for the separation illustrated in Appendix C.1.

We also counted the relative frequencies of actual elapsed-time intervals in the fold-0 validation sets using the same four regimes as the submitted Algebra2005 Table 3.

Table R3.4. Observed elapsed-time gap proportions in fold-0 validation interactions.

Dataset≤1 hour1 hour–1 day1–7 days>7 days
ASSIST201291.2792%1.8216%3.7277%3.1715%
Algebra200597.6257%0.7767%1.3119%0.2857%

For completeness, the held-out Algebra2005 test set contains 91,410, 509, 1,202, and 360 interactions in these respective bins, totaling 93,481 unique test targets; the >7-day bin is only 0.3851% of that test set. The ASSIST2012 test set, in turn, has 16,409 interactions beyond seven days among 524,251 targets (3.1300%).

The magnitudes of the learned half-lives can be related to these data regimes. On ASSIST2012, the personalized medium-scale median of 14.82 days is close to the median observed gap of 14.66 days among validation interactions exceeding seven days, while the slow-scale median of 281.28 days covers a much longer retention regime. On Algebra2005, the personalized medium median of 6.82 days is near the upper end of the 1–7-day range, and the slow median of 104.22 days extends beyond the largest observed validation gap of 89.83 days. These results show that the model learns distinct sub-day, multi-day/week, and longer-term retention horizons, rather than three nearly identical decay rates. The validation-gap distributions also explain why long-gap performance needs to be interpreted in light of its relatively sparse supervision.

Q2 — What evidence supports multiple memory timescales rather than one?

We evaluated one- and two-active-scale controls on Algebra2005 with seed 3407, a five-fold held-out protocol, and the same episodic memory, hierarchical item-response model, and overall parameter budget. S1 uses Fast only; S2 uses Fast + Slow. Inactive scales are masked from readout and response-triggered updates, and update regularizers are normalized by the number of active scales. Full CT-LFSSM uses all three timescales.

Table R3.5. Algebra2005 Full three-scale result and additional S1/S2 active-scale controls (five-fold mean ± fold-level population SD).

Active state scalesAUC ↑ACC ↑
S1: Fast only0.8224 ± 0.00030.8097 ± 0.0013
S2: Fast + Slow0.8266 ± 0.00050.8126 ± 0.0013
S3: Fast + Medium + Slow (Full CT-LFSSM)0.8304 ± 0.00050.8168 ± 0.0016

The complete three-scale CT-LFSSM achieves the highest reported mean AUC (0.8304) and ACC (0.8168), exceeding S1 by +0.0080 AUC and +0.0071 ACC and S2 by +0.0038 AUC and +0.0042 ACC. S2 also improves AUC over S1 in all five folds (+0.0042 on average).

The event-conditioned scale gate provides an additional source of evidence. The FixedMix control replaces dynamic mixing with uniform 1/3, 1/3, 1/3 weights, obtaining 0.8247 ± 0.0004 AUC and 0.8094 ± 0.0026 ACC on Algebra2005. Full CT-LFSSM achieves higher reported mean AUC (+0.0057) and ACC (+0.0074). Together with the learned half-life separation in Q1 and adaptive gate pattern in Appendix C.2, this supports the use of distinct retention horizons and event-conditioned readout.

Reviewer_1gshOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CT-LFSSM, a continuous-time knowledge tracing framework that represents student mastery via three latent states with different forgetting rates. It separates response-driven belief revision from practice-induced learning and introduces a hierarchical item-response model. The results are good across four benchmarks, although the evidence for the central learning and forgetting mechanisms is less convincing than the overall predictive performance suggests.

Critical strengths
  1. The distinction between belief revision and knowledge acquisition is well motivated.
  2. The hierarchical item-response formulation is a sensible design choice. Sharing parameters across knowledge components while retaining item-specific variation can help address data sparsity, and the ablation results support its effectiveness.
Critical weaknesses
  1. Table 4 reveals that removing episodic memory reduces AUC by 0.0362 on ASSIST2009 and 0.0267 on Algebra2005, whereas removing the hierarchical state update reduces AUC by only 0.0010 and 0.0043, respectively. This raises a question about whether the performance gains primarily originate from the proposed state dynamics or the additional memory and observation components.

  2. Table 3 shows that CT-LFSSM achieves an AUC of 0.7619 for gaps exceeding seven days, substantially below HawkesKT at 0.7953. Similarly, LPKT outperforms CT-LFSSM for gaps between one and seven days. Given that long-term retention is a key motivation, these results deserve more discussion. More importantly, comparisons across time bins alone do not establish whether the proposed decay mechanism models forgetting more accurately.

  3. The paper discusses CT-NCM, LefoKT, FlucKT, and TFKT but does not include them in the main experimental comparison. This weakens the state-of-the-art claim, particularly because these methods address temporal evolution or forgetting directly. The XES3G5M comparison also relies on previously published benchmark results rather than evaluations against recent methods. The related work could additionally discuss HiTSKT and recent comparative studies of forgetting functions.

Questions for authors
  1. How much improvement comes specifically from the three continuous-time states when episodic memory and the item-response model remain unchanged?

  2. Why does CT-LFSSM fall behind HawkesKT for long temporal gaps despite its explicit multiscale forgetting formulation?

5 comments

We sincerely thank the reviewer for the careful assessment of CT-LFSSM, particularly the attribution of predictive gains to episodic memory versus latent-state dynamics (W1/Q1), the interpretation of long-gap performance and explicit forgetting (W2/Q2), and the coverage of directly relevant baselines and related work (W3). We address these points individually using the submitted ablations, newly completed five-fold state and update-path controls, a direct no-decay comparison on timestamped data, exact elapsed-time sample counts, and additional temporal-KT baseline results.

W1 — Which components account for the predictive gains?

The episodic memory and continuous-time state-space dynamics serve different roles in the predict–observe–update computation. Episodic retrieval supplies a bounded, similarity-conditioned residual to the current observation logit. In contrast, the KC-level fast-, medium-, and slow-timescale states evolve between observations through personalized half-life decay; after each observed answer, response-driven belief revision and practice-induced learning alter the latent state used for future predictions. The hierarchical item-response component provides shared KC-based observation parameters. These mechanisms therefore provide distinct sources of predictive information.

The submitted Table 4 already provides a useful component-level comparison. For clarity, Table R2.1 reports the actual five-fold AUC of the full model and each model trained without one component, rather than presenting positive AUC differences as if they were the scores of the ablated models. A lower AUC in a w/o row indicates that removing that component reduces predictive performance. Importantly, the original w/o hierarchical state update disables the hierarchical evidence-strength and consolidation parameterizations (use_hierarchical_evidence_update=False, use_hierarchical_consolidation=False); it does not remove the three latent states, their continuous-time decay, or dynamic scale readout. Its smaller performance reduction therefore measures a narrower intervention than removal of all state dynamics.

Table R2.1. Submitted Table 4: five-fold test AUC (mean ± fold-level population SD). Each w/o row is the AUC after removing the specified component, not an AUC gain.

ModelASSIST2012 AUC ↑ASSIST2009 AUC ↑Algebra2005 AUC ↑
Full CT-LFSSM0.7833 ± 0.00050.7873 ± 0.00190.8304 ± 0.0005
w/o Episodic Memory0.7699 ± 0.00080.7511 ± 0.00060.8037 ± 0.0017
w/o Hierarchical Item-Response0.7730 ± 0.00060.7680 ± 0.00220.8135 ± 0.0014
w/o Cognition and Acquisition Gates0.7796 ± 0.00150.7818 ± 0.00160.8243 ± 0.0005
w/o Outcome Conditioning0.7822 ± 0.00030.7857 ± 0.00170.8267 ± 0.0007
w/o Hierarchical State Update0.7816 ± 0.00070.7863 ± 0.00170.8261 ± 0.0007

Relative to w/o Episodic Memory, retaining episodic retrieval increases mean AUC by +0.0134, +0.0362, and +0.0267 on ASSIST2012, ASSIST2009, and Algebra2005, respectively. The corresponding gains from retaining hierarchical item-response modeling are +0.0103, +0.0193, and +0.0169. The gains of the narrower hierarchical response-state-update parameterization are +0.0017, +0.0010, and +0.0043. These complementary gains motivate the more focused state-transition controls below.

To examine the response-driven state-transition pathways more directly, we retrained three controls on ASSIST2012: NoRevision (no residual-based belief revision), NoLearning (no practice-induced learning gain), and NoBoth (both paths disabled). All controls retain the episodic memory and hierarchical observation model. Each model used seed 3407, validation-AUC checkpoint selection, and five-fold held-out evaluation with 524,251 targets per fold. Table R2.2 reports the submitted Full result alongside the new controls.

Table R2.2. ASSIST2012 Full result and post-response pathway ablations (five-fold mean ± fold-level population SD).

ModelAUC ↑ACC ↑Full − variant ΔAUC
Full CT-LFSSM0.7833 ± 0.00050.7582 ± 0.0003—
NoRevision (learning retained)0.7817 ± 0.00040.7568 ± 0.0004+0.0016
NoLearning (revision retained)0.7780 ± 0.00050.7553 ± 0.0002+0.0053
NoBoth (neither retained)0.7762 ± 0.00030.7547 ± 0.0001+0.0071

Full CT-LFSSM achieves the highest reported mean AUC and ACC. Relative to NoBoth, keeping only practice-induced learning improves mean AUC by +0.005495, while keeping only response-driven belief revision improves it by +0.001785. These results make the state-update contribution visible even while episodic retrieval and the item-response model remain present. The independent contributions of the continuous-time state evolution and scale readout are evaluated with still more targeted controls in Q1 below.

W2 — Interpretation of time-gap results

We audited the elapsed-time distribution of all held-out Algebra2005 test targets and paired the exact bin counts with the submitted Table 3 metrics. The resulting sample sizes provide a concrete basis for interpreting AUC and ACC across time horizons, especially for the relatively small long-gap subset.

Table R2.3. Algebra2005: exact held-out target counts and relevant time-gap AUC/ACC comparisons from submitted Table 3. The gap bins use the time between consecutive valid interactions.

Elapsed-time binTest targetsCT-LFSSM AUCHawkesKT AUCLPKT AUCCT-LFSSM ACCHawkesKT ACC
≤1 hour91,4100.82610.81700.81870.81530.8141
1 hour–1 day5090.79400.79400.79080.71880.6688
1–7 days1,2020.78660.78580.80020.72300.7030
>7 days3600.76190.79530.75760.66770.6669

The >7-day interval contains 360 of 93,481 test targets (0.3851%) across 101 student sequences and 267 distinct items; its median elapsed time is 13.96 days. By comparison, 91,410 interactions (97.7846%) lie in the ≤1-hour bin. Thus, the longest-range AUC reflects a small, distinct evaluation subset. In that bin, HawkesKT leads in AUC (0.7953 versus 0.7619), while the two models have nearly identical ACC (0.6669 versus 0.6677). In the 1–7-day bin, LPKT has higher AUC (0.8002 versus 0.7866), whereas CT-LFSSM has higher ACC (0.7230 versus 0.7155). The exact target counts make clear which gap comparisons have substantial evaluation support and which depend on much smaller subsets.

The sampling imbalance is also visible before testing: in Algebra2005, the mean share of supervised training targets with gaps >7 days is 0.3078% across folds; for ASSIST2012 it is 3.1486%. The corresponding fixed held-out test proportions are 0.3851% and 3.1300%. These counts characterize the limited long-gap supervision in Algebra2005 and help interpret the variability of the longest-range ranking. We additionally isolated the decay mechanism itself on ASSIST2012; the controlled results are provided in Q2 rather than inferred solely from cross-model gap comparisons.

W3 — Direct temporal baselines, XES3G5M, and related work

We have expanded the direct comparison with CT-NCM, DMKT, FlucKT, and LefoKT from ASSIST2012 to all three datasets: ASSIST2012, ASSIST2009, and Algebra2005. Table R2.4 reports the completed five-fold AUC and ACC results (mean ± fold-level population SD). These comparisons extend the baseline analysis of the submitted manuscript and Appendix B.2.

Table R2.4. Expanded direct temporal/multiscale baseline comparison on three datasets (five-fold mean ± fold-level population SD; higher is better).

MethodASSIST2012 ACC ↑ASSIST2012 AUC ↑ASSIST2009 ACC ↑ASSIST2009 AUC ↑Algebra2005 ACC ↑Algebra2005 AUC ↑
CT-NCM (2022)0.7569 ± 0.00250.7795 ± 0.00310.7355 ± 0.00150.7663 ± 0.00160.7994 ± 0.00060.8032 ± 0.0008
DMKT (2025)0.7284 ± 0.01080.7527 ± 0.01440.7131 ± 0.00510.7416 ± 0.00670.7886 ± 0.00210.7958 ± 0.0024
FlucKT (2025)0.7577 ± 0.00100.7809 ± 0.00190.7373 ± 0.00280.7731 ± 0.00330.8101 ± 0.00120.8280 ± 0.0010
LefoKT (2025)0.7559 ± 0.00310.7792 ± 0.00350.7400 ± 0.01000.7858 ± 0.00250.8109 ± 0.00090.8272 ± 0.0010
CT-LFSSM0.7582 ± 0.00030.7833 ± 0.00050.7419 ± 0.00140.7873 ± 0.00190.8168 ± 0.00160.8304 ± 0.0005

CT-LFSSM achieves the highest mean AUC and ACC in all three datasets among these direct baselines. The consistency across different learning datasets is particularly informative: the mean AUC advantages over the strongest competing model in each dataset are +0.0024 (ASSIST2012), +0.0015 (ASSIST2009), and +0.0024 (Algebra2005), while the ACC advantages are +0.0005, +0.0019, and +0.0059. Together with the controlled ablations in W1 and Q1, these results support the empirical benefit of the proposed combination of continuous-time state evolution, multiscale readout, and response-dependent transitions.

We also attempted to reproduce TFKT. Although its source code is public, the repository's README provides AKT rather than TFKT execution commands and does not document the data splits and evaluation settings needed to match our five-fold protocol on these datasets. We therefore exclude TFKT from the numerical comparison.

On XES3G5M, the submitted Appendix B.3 reports results from the published official five-fold benchmark rather than a new matched-input rerun of all 19 baselines. Table R2.5 reproduces representative results and the input categories documented for those models. CT-LFSSM obtains 0.8315 AUC and 0.8380 ACC, compared with 0.8280 AUC and 0.8316 ACC for the strongest reported baseline, IEKT.

Table R2.5. Selected XES3G5M official-benchmark results and documented input categories (submitted Appendix B.3).

ModelQuestion IDKC IDTimestampAUC ↑ACC ↑
DKT-FNoYesYes0.79400.8209
SAINTYesYesYes0.80740.8177
IEKTYesYesNo0.82800.8316
CT-LFSSMYesYesYes0.83150.8380

In line with the shared emphasis on temporal learning dynamics and forgetting in HiTSKT (Ke et al., 2024) and Sha et al. (2026), CT-LFSSM’s distinguishing contribution is to model forgetting directly within each student–KC latent state through personalized fast-, medium-, and slow-timescale half-lives, combine these states through an event-conditioned gate before prediction, and separate evidence-based belief revision from practice-induced learning after observing the response. The learned half-life ranges and targeted ablations in Q1 provide direct evidence for these design choices.

Q1 — What is the specific contribution of continuous-time states when episodic memory and item response are retained?

We evaluated the Fast-only (S1) and Fast + Slow (S2) active-scale controls under matched parameter budgets, retaining episodic memory, hierarchical item-response, and response-update mechanisms. Inactive state slots are masked. Full CT-LFSSM uses all three timescales. All five-fold results use the Algebra2005 held-out evaluation with 93,481 test targets per fold.

Table R2.6. Algebra2005 Full three-scale result and five-fold S1/S2 active-scale controls, with episodic memory and item-response modeling retained.

Active latent timescalesAUC ↑ACC ↑
S1: Fast only0.8224 ± 0.00030.8097 ± 0.0013
S2: Fast + Slow0.8266 ± 0.00050.8126 ± 0.0013
S3: Fast + Medium + Slow (Full CT-LFSSM)0.8304 ± 0.00050.8168 ± 0.0016

Full CT-LFSSM achieves the highest reported mean AUC (0.8304) and ACC (0.8168). Relative to S1, these means are higher by +0.008036 AUC and +0.007097 ACC; relative to S2, by +0.003839 AUC and +0.004201 ACC. Among the matched active-scale controls, S2 improves over S1 in AUC in all five folds (+0.004197 on average). Together, these results show gains from multiple retention horizons while retaining episodic memory and item-response capacity.

The trained half-lives further confirm that the model actually learns different retention regimes. Across the two timestamped datasets, the median personalized fast, medium, and slow half-lives are:

Table R2.7. Median personalized half-lives and adjacent-scale ratios from trained CT-LFSSM checkpoints.

DatasetFast median (days)Medium median (days)Slow median (days)Medium/FastSlow/Medium
ASSIST20120.618314.8189281.277723.834818.9909
Algebra20050.35406.8189104.216119.253015.0202

In the inspected model export, no sampled adjacent-scale ratio is below 2, and no sampled half-life is saturated at its numerical boundary. Hence the three latent horizons do not simply collapse to nearly identical forgetting rates.

Finally, a second controlled test keeps all three states but replaces the event-conditioned gate with uniform fixed mixing. On Algebra2005, this FixedMix variant achieves 0.824665 ± 0.000422 AUC and 0.809405 ± 0.002635 ACC. The Full three-scale model records higher mean AUC (+0.005735) and ACC (+0.007395) than the FixedMix control. Together, the active-scale and gate comparisons isolate the predictive contributions of multiple latent retention horizons and their event-conditioned combination while leaving the memory and observation modules in place.

Reviewer_jhA8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CT-LFSSM, a continuous-time learning–forgetting state-space model for knowledge tracing. It maintains fast, medium, and slow latent states for each knowledge component, models their decay over real elapsed time using personalized half-lives, and combines them through an event-conditioned gate. After observing a response, the model first revises the knowledge states using the prediction residual and then incorporates practice-induced learning gains. The prediction module further uses episodic memory and a hierarchical item-response model with knowledge-component-level parameters and item-specific residuals.

Critical strengths

S1. The paper addresses a clear and relevant problem in knowledge tracing.

Critical weaknesses

W1. The novelty of this paper is not sufficiently clear. (D1)

W2. The methodology is not sufficiently solid. (D2, D3)

W3. The experiments are not sufficient. (D4, D5, D6)

Questions for authors

D1. The proposed framework combines several existing mechanisms, including continuous-time forgetting, multiscale state representation, and post-response state updates. However, the paper does not sufficiently clarify what fundamentally distinguishes this combination from prior knowledge tracing methods. Although Section 2.2 discusses related approaches such as CT-NCM, FlucKT, DMKT, and TFKT, some directly relevant methods are not included in the main experimental comparisons. Please clarify the methodological novelty beyond combining existing components and provide more direct comparisons with closely related methods where feasible.

D2. The 4PL equations use different difficulty parameterizations. Equations (4) and (14) appear to use different parameterizations of item difficulty, particularly in how it interacts with item discrimination. Please clarify the intended formulation and ensure that the definition and interpretation of item difficulty are consistent throughout the paper and appendix.

D3. The sequence-length analysis groups students by their complete histories, whereas model inputs are limited to 200 interactions and evaluation uses rolling windows. Please clarify whether latent states and episodic memory are carried across windows, and how much prior history is available for each prediction in the longest-sequence group.

D4. The ablations do not separately test the contributions of continuous-time decay and the three-timescale state representation. Please include single-timescale and no-decay variants on datasets with timestamps, and specify which components are removed in “w/o hierarchical state update.”

D5. The paper motivates the hierarchical item-response model partly by the difficulty of estimating parameters for infrequently observed items. However, Table 4 only removes the entire module and reports aggregate performance. Please report results stratified by item frequency to assess whether the proposed parameter sharing benefits infrequent items.

D6. Table 3 reports performance across time-gap bins on Algebra2005 but omits the number of test interactions in each bin. Please report these counts, particularly for the greater-than-seven-day bin, to support the interpretation of the long-gap results.

7 comments

We sincerely thank the reviewer for the detailed assessment of CT-LFSSM. We address three sets of concerns in sequence: methodological differentiation and direct temporal-KT comparisons (W1/D1); the 4PL-style observation model and rolling-window evaluation protocol (W2/D2–D3); and targeted evidence for multi-timescale state evolution, low-frequency item prediction, and elapsed-time evaluation (W3/D4–D6). In addition to the submitted ablations, we provide four newly evaluated temporal-KT baselines across three datasets, active-scale and matched no-decay controls, a complete evaluation-window audit, item-frequency analyses with a paired student-cluster bootstrap, and exact time-gap sample counts. We have also revised the manuscript accordingly for your reference and sincerely appreciate your time and consideration.

W1 — Methodological Contribution

D1. What distinguishes CT-LFSSM from related temporal KT methods, and how does it compare empirically?

We thank the reviewer for this important question. Existing temporal KT methods have advanced the modeling of continuous-time forgetting and heterogeneous knowledge dynamics. For example, CT-NCM models continuous-time knowledge evolution, while DMKT and FlucKT explore different forms of temporal heterogeneity. Building on these advances, CT-LFSSM focuses on three interconnected challenges: (1) modeling knowledge retention across heterogeneous timescales, (2) distinguishing evidence-based knowledge-state revision from practice-induced learning, and (3) maintaining reliable response prediction under sparse item-level observations. We address these challenges within a unified continuous-time predict–observe–update framework, with the effectiveness of the corresponding modeling decisions examined through the original and targeted ablation experiments.

The following discussion directly connects each modeling challenge to its proposed solution and supporting experimental evidence.

Table R1.1. Original ablation study on ASSIST2012, ASSIST2009, and Algebra2005 (five-fold mean ± population standard deviation; higher is better).

ModelASSIST2012 AUCASSIST2012 ACCASSIST2009 AUCASSIST2009 ACCAlgebra2005 AUCAlgebra2005 ACC
Full CT-LFSSM0.7833 ± 0.00050.7582 ± 0.00030.7873 ± 0.00190.7419 ± 0.00140.8304 ± 0.00050.8168 ± 0.0016
w/o hierarchical item-response0.7730 ± 0.00060.7519 ± 0.00120.7680 ± 0.00220.7270 ± 0.00070.8135 ± 0.00140.8070 ± 0.0016
w/o cognition and acquisition gates0.7796 ± 0.00150.7558 ± 0.00090.7818 ± 0.00160.7376 ± 0.00120.8243 ± 0.00050.8111 ± 0.0019
w/o outcome conditioning0.7822 ± 0.00030.7574 ± 0.00020.7857 ± 0.00170.7408 ± 0.00060.8267 ± 0.00070.8138 ± 0.0015
w/o episodic memory0.7699 ± 0.00080.7507 ± 0.00070.7511 ± 0.00060.7207 ± 0.00090.8037 ± 0.00170.7900 ± 0.0068
w/o hierarchical state update0.7816 ± 0.00070.7569 ± 0.00050.7863 ± 0.00170.7412 ± 0.00070.8261 ± 0.00070.8117 ± 0.0015

Challenge 1: Modeling knowledge retention across heterogeneous timescales. Knowledge acquired through irregular interactions may exhibit different retention rates, making it difficult to represent knowledge evolution through a single temporal horizon. CT-LFSSM addresses this by maintaining fast-, medium-, and slow-timescale states with personalized half-lives, explicit continuous-time decay, and event-conditioned scale integration.

The targeted five-fold experiments on Algebra2005 directly support these choices. The complete three-scale model achieves an AUC of 0.830400, outperforming S1 (Fast only, 0.822364) and S2 (Fast + Slow, 0.826561) by 0.008036 and 0.003839, respectively. Retaining all three scales but replacing event-conditioned integration with uniform FixedMix reduces AUC to 0.824665, corresponding to a 0.005735 advantage for the full model. Furthermore, removing explicit continuous-time state decay decreases AUC from 0.782923 to 0.782120 in the paired ASSIST2012 Fold-0 experiment. These findings provide direct evidence for the predictive contributions of multiple retention horizons, adaptive scale integration, and explicit temporal decay.

Challenge 2: Distinguishing knowledge-state revision from practice-induced learning. A student response provides evidence about existing knowledge while the interaction itself may also induce learning. Treating both effects as an undifferentiated state transition makes their distinct roles difficult to represent. CT-LFSSM therefore separates Response-Driven Belief Revision from Practice-Induced Learning, allowing evidence-based correction and nonnegative acquisition to operate sequentially.

The five-fold ASSIST2012 experiments directly evaluate these two pathways. The full model achieves an AUC of 0.783300, compared with 0.781708 for NoRevision, 0.777998 for NoLearning, and 0.776214 for NoBoth, corresponding to improvements of 0.001592, 0.005302, and 0.007086, respectively. More importantly, on Algebra2005, the separate-update design outperforms UnifiedUpdate (0.830400 versus 0.827037 AUC; +0.003363), providing further evidence for explicitly distinguishing the two update operations. The original ablations in Table R1.1 offer complementary support: removing outcome conditioning decreases AUC by 0.0011–0.0037, while removing hierarchical state updates decreases AUC by 0.0010–0.0043 across the three datasets.

Challenge 3: Maintaining reliable response prediction under sparse item-level observations. Individual items often have limited response records, making purely item-specific estimation unreliable. CT-LFSSM addresses this challenge through KC-shared hierarchical item-response modeling with regularized item-specific variation, complemented by historical episodic evidence.

The original ablations provide direct evidence for the contributions of these two components. Removing hierarchical item-response modeling decreases AUC by 0.0103, 0.0193, and 0.0169 on ASSIST2012, ASSIST2009, and Algebra2005, respectively. Removing episodic memory produces further AUC decreases of 0.0134, 0.0362, and 0.0267. The targeted item-frequency analysis in D5 further shows that hierarchical item-response modeling improves AUC by 0.0067 on ASSIST2012 and 0.0079 on Algebra2005 for items with only 1–5 supervised training records. These results support the effectiveness of the hierarchical observation model and historical evidence, including under limited item-level supervision.

Expanded Empirical Comparison

To further evaluate the effectiveness of these modeling decisions in combination, we additionally compared CT-LFSSM with four directly relevant temporal KT methods—CT-NCM, DMKT, FlucKT, and LefoKT—under a five-fold evaluation protocol on ASSIST2012, ASSIST2009, and Algebra2005.

Table R1.2. Comparison with additional temporal KT baselines in AUC (five-fold mean ± population standard deviation; higher is better).

ModelASSIST2012ASSIST2009Algebra2005
CT-NCM (2022)0.7795 ± 0.00310.7663 ± 0.00160.8032 ± 0.0008
DMKT (2025)0.7527 ± 0.01440.7416 ± 0.00670.7958 ± 0.0024
FlucKT (2025)0.7809 ± 0.00190.7731 ± 0.00330.8280 ± 0.0010
LefoKT (2025)0.7792 ± 0.00350.7858 ± 0.00250.8272 ± 0.0010
CT-LFSSM0.7833 ± 0.00050.7873 ± 0.00190.8304 ± 0.0005

Table R1.3. Comparison with additional temporal KT baselines in ACC (five-fold mean ± population standard deviation; higher is better).

ModelASSIST2012ASSIST2009Algebra2005
CT-NCM (2022)0.7569 ± 0.00250.7355 ± 0.00150.7994 ± 0.0006
DMKT (2025)0.7284 ± 0.01080.7131 ± 0.00510.7886 ± 0.0021
FlucKT (2025)0.7577 ± 0.00100.7373 ± 0.00280.8101 ± 0.0012
LefoKT (2025)0.7559 ± 0.00310.7400 ± 0.01000.8109 ± 0.0009
CT-LFSSM0.7582 ± 0.00030.7419 ± 0.00140.8168 ± 0.0016

CT-LFSSM achieves the highest reported mean AUC and ACC among these methods on all three datasets. Compared with the strongest competing method for each dataset and metric, CT-LFSSM improves mean AUC by 0.0024, 0.0015, and 0.0024, and mean ACC by 0.0005, 0.0019, and 0.0059, respectively.

Overall, the original and targeted ablations directly connect the three identified challenges to the effectiveness of their corresponding solutions, while the expanded baseline comparisons demonstrate the competitive predictive performance of the complete framework. We also attempted to reproduce TFKT. Although its source code is public, the repository's README provides AKT rather than TFKT execution commands and does not document the data splits and evaluation settings needed to match our five-fold protocol on these datasets. We therefore exclude TFKT from the numerical comparison.

W2 — Mathematical Consistency and Evaluation Protocol

D2. Why do Equations (4) and (14) look different, and how is difficulty defined?

Equation (4) is the conventional four-parameter logistic (4PL) response function presented as background in Section 3.2. Equation (14) is our CT-LFSSM observation model, which preserves that 4PL response family while specifying how the input mastery and item parameters are constructed and adding an event-specific episodic correction. Their different roles, rather than a change in the difficulty parameterization, explain the additional terms.

The standard 4PL form in Equation (4) is

pu,nIRT=gq+(1−gq−sq)σ(aq(mu,n−bq)).p^{\mathrm{IRT}}_{u,n}=g_q+(1-g_q-s_q)\sigma(a_q(m_{u,n}-b_q)).

Here bqb_q is the difficulty location on the mastery scale, aq>0a_q>0 is discrimination, and gqg_q and sqs_q determine the lower and upper asymptotes. In CT-LFSSM, the mastery input is not a static ability variable. It is constructed from elapsed-time-decayed KC states, an event-conditioned scale gate, and the current item/KC context:

mu,n,kKC=∑r∈f,m,sωu,n,rzu,n,k,r−+νk.m^{\mathrm{KC}}_{u,n,k}=\sum_{r\in\\{f,m,s\\}}\omega_{u,n,r}z^{-}_{u,n,k,r}+\nu_k.

mu,n=cu,ncog∑kρu,n,kpredmu,n,kKC+βu,n.m_{u,n}=c^{\mathrm{cog}}_{u,n}\sum_k\rho^{\mathrm{pred}}_{u,n,k}m^{\mathrm{KC}}_{u,n,k}+\beta_{u,n}.

We also estimate item characteristics through KC-level sharing rather than fitting four unrelated sets of parameters separately for every item. With normalized item–KC weights πqk\pi_{qk},

bq=∑kπqkbk+δqb.\displaystyle b_q=\sum_k\pi_{qk}b_k+\delta^b_q.

aq=0.1+softplus⁡(∑kπqkαk+δqa).\displaystyle a_q=0.1+\operatorname{softplus}\left(\sum_k\pi_{qk}\alpha_k+\delta^a_q\right).

gq=gmax⁡σ(∑kπqkθkguess).\displaystyle g_q=g_{\max}\sigma\left(\sum_k\pi_{qk}\theta^{\mathrm{guess}}_k\right).

sq=smax⁡σ(∑kπqkθkslip).\displaystyle s_q=s_{\max}\sigma\left(\sum_k\pi_{qk}\theta^{\mathrm{slip}}_k\right).

Equation (14) then introduces a bounded, past-only episodic residual at the observation logit:

ℓu,n=aq(mu,n−bq)+δu,nepi.\displaystyle \ell_{u,n}=a_q(m_{u,n}-b_q)+\delta^{\mathrm{epi}}_{u,n}.

pu,n=gq+(1−gq−sq)σ(ℓu,n).\displaystyle p_{u,n}=g_q+(1-g_q-s_q)\sigma(\ell_{u,n}).

This extension serves two purposes. First, KC sharing enables item difficulty and discrimination to draw on observations from related items, which is especially useful when individual items have few supervised occurrences (quantified in D5). Second, episodic retrieval contributes context-specific evidence to the current response logit without rewriting persistent item difficulty or changing the underlying KC state. Placing δu,nepi\delta^{\mathrm{epi}}_{u,n} after the discrimination-scaled mastery–difficulty term keeps its contribution in logit units and preserves the 4PL guessing/slipping asymptotes. All inputs to this observation are available before the current response.

Crucially, the psychometric core is unchanged: both equations use aq(mu,n−bq)=aqmu,n−aqbqa_q(m_{u,n}-b_q)=a_qm_{u,n}-a_qb_q, not aqmu,n−bqa_qm_{u,n}-b_q. Setting δu,nepi=0\delta^{\mathrm{epi}}_{u,n}=0 in Equation (14), with the same mastery and item parameters, recovers Equation (4) exactly. Thus, Equation (14) is an episodic-augmented, KC-factorized instantiation of the standard 4PL response function, not a different definition of difficulty. The revised manuscript clarifies this baseline-to-proposed-model relationship alongside Equation (14).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for the careful and constructive feedback. We have provided detailed point-by-point responses, completed additional controlled experiments and analyses, and incorporated the corresponding revisions into the manuscript.

Following Reviewer jhA8's comments, we clarified the mathematical relationship between the standard 4PL formulation and our hierarchical observation model, audited the rolling-window evaluation protocol, and added targeted evidence for the temporal mechanism and KC-level parameter sharing. Following Reviewer 1gsh's suggestions, we expanded the direct temporal-baseline comparisons and added a controlled real-timestamp decay intervention. Following Reviewer yPPY's comments, we reorganized the method around a compact predict-observe-update roadmap and added targeted tests of scale count, dynamic mixing, learned half-lives, and post-response updates. Following Reviewer ADw6's assessment, we further consolidated the component-level evidence, clarified baseline input conditions and provenance, and added paired student-cluster bootstrap analysis.

The revised manuscript now connects the main modeling choices of CT-LFSSM to explicit mechanism-level evidence and presents the evaluation protocol and statistical interpretation more transparently.

Core Contributions of Our Work

CT-LFSSM models knowledge tracing as a continuous-time, multi-timescale state-space process. For each knowledge component, it maintains fast-, medium-, and slow-timescale latent states, propagates them over real elapsed time through personalized exponential decay, and combines them with an event-conditioned convex gate. The learned personalized half-life medians are 0.62 / 14.82 / 281.28 days on ASSIST2012 and 0.35 / 6.82 / 104.22 days on Algebra2005, showing distinct short-, medium-, and long-term retention regimes.

The model follows a strict predict-observe-update process. Before observing the current response, CT-LFSSM evolves the latent states, performs event-conditioned multiscale readout, and forms the response probability using hierarchical 4PL-style item-response modeling and past-only episodic retrieval. After the response is observed, a signed prediction-residual pathway revises the inferred pre-interaction state, followed by a separate nonnegative practice-induced learning transition. This gives forgetting, evidence-based belief revision, and learning distinct temporal and mathematical roles.

For the observation model, CT-LFSSM combines KC-level parameter sharing with regularized item-specific variation, while episodic retrieval contributes context-specific evidence without rewriting the persistent KC state. Across four real-world datasets, the manuscript reports state-of-the-art performance, improving over the strongest baselines by 0.35-1.25 percentage points in AUC and 0.30-1.14 percentage points in accuracy.

Clarifications and Manuscript Revisions

RevisionWhat is now made explicit
Method execution flow (yPPY, W1)Section 4 is organized around the predict-observe-update sequence, with the main information boundary and component roles stated before the detailed equations; lower-level parameterizations remain in Appendix A.
4PL consistency (jhA8, D2)Equations (4) and (14) use the same psychometric core, with difficulty entering as aq(mu,n−bq)a_q(m_{u,n}-b_q). Equation (14) extends the standard 4PL through dynamic mastery, KC-level parameter sharing, and a bounded past-only episodic logit residual.
Rolling-window evaluation (jhA8, D3)A prediction sees at most 199 earlier interactions within a 200-interaction segment. Each window reconstructs its latent state, profile, and episodic memory from the history visible in that window. The ASSIST2012 audit covers exactly 524,251 held-out targets with no missing or duplicated predictions.
Temporal baselines and benchmark provenance (1gsh, W3; ADw6, W3)The revised discussion distinguishes rerun three-dataset comparisons from the official XES3G5M benchmark and makes question-ID, KC-ID, and timestamp input conditions explicit.
Statistical reporting (ADw6, W4)The paper states that the reported standard deviations are fold-level population SDs at fixed seed 3407 and supplements the item-response analysis with paired student-cluster bootstrap intervals.
Ablation presentation (jhA8, yPPY, ADw6)The original component ablations are separated from the new targeted scale-count, FixedMix, NoHalfLifeOrder, UnifiedUpdate, response-branch, and explicit-decay controls, so each intervention has a clear interpretation.

Additional Experiments and Analyses

Added analysisMain result
Number of active timescales - Algebra2005Full S3 reaches 0.8304 AUC, exceeding S1 (Fast only) by +0.0080 and S2 (Fast + Slow) by +0.0038. S2 also exceeds S1 in AUC in all five folds.
Event-conditioned scale mixing - Algebra2005Full exceeds uniform FixedMix by +0.0057 AUC and +0.0074 ACC, directly supporting adaptive scale readout.
Ordered half-life parameterization - Algebra2005Full exceeds NoHalfLifeOrder by +0.0038 AUC, while the learned half-lives remain clearly separated across the three retention horizons.
Separate post-response transitionsOn Algebra2005, Full exceeds the capacity-preserving UnifiedUpdate control by +0.0034 AUC. On ASSIST2012, Full exceeds NoRevision, NoLearning, and NoBoth by +0.0016, +0.0053, and +0.0071 AUC, respectively.
Explicit continuous-time decay - ASSIST2012In the controlled fold-0 Full-vs-NoDecay intervention, retaining latent-state decay improves all five aggregate metrics, including +0.0008 AUC overall. Across elapsed-time bins, the largest AUC gain is +0.0030 for gaps longer than 30 days.
Episodic memory and hierarchical item-responseThe original ablations show AUC gains from episodic memory of +0.0134 / +0.0362 / +0.0267, and from hierarchical item-response modeling of +0.0103 / +0.0193 / +0.0169, on ASSIST2012 / ASSIST2009 / Algebra2005, respectively.
Paired student-cluster bootstrapHierarchical item-response yields mean paired AUC gains of +0.0093 on ASSIST2012 (95% CI [+0.0089, +0.0099]) and +0.0130 on Algebra2005 (95% CI [+0.0111, +0.0149]).
Direct temporal/multiscale baselinesAgainst CT-NCM, DMKT, FlucKT, and LefoKT, CT-LFSSM achieves the highest mean AUC and ACC on ASSIST2012, ASSIST2009, and Algebra2005. Its AUC margins over the strongest of these baselines are +0.0024 / +0.0015 / +0.0024, respectively.
Real time-gap supportThe revised analysis reports exact test counts for each elapsed-time regime. In Algebra2005, only 360 / 93,481 targets fall beyond 7 days, while ASSIST2012 contains 16,409 / 524,251 targets beyond 7 days, providing substantially richer long-gap supervision for the real-time analyses.
XES3G5M official benchmarkUnder the official five-fold benchmark, CT-LFSSM reaches 0.8315 AUC / 0.8380 ACC, compared with 0.8280 / 0.8316 for IEKT, while the revised table also reports the documented question-ID, KC-ID, and timestamp input flags.

Taken together, the revised paper now presents CT-LFSSM at three complementary levels: overall predictive performance, controlled mechanism-level evidence, and real-time/statistical analysis. The additional experiments make the roles of the multiscale state dynamics, event-conditioned readout, hierarchical observation model, episodic evidence, and post-response transitions directly visible in the empirical results.

Thank you again for your time, careful reading, and constructive feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.