acceptodds
Under review as a conference paper at ICLR 2027

DualKT: Dual-Timescale Knowledge Tracing via Orthogonal Decomposition and Target-Conditioned Fusion

Abstract

Knowledge tracing (KT) predicts future performance from students' interaction histories. Accurate prediction requires modeling both long-term evolution across the full history and local fluctuations driven by recent interactions. Yet existing methods often encode both scales jointly or directly combine multiple encoders without explicitly modeling the geometric relationship between their representations. We propose DualKT, a dual-timescale model with orthogonal decomposition and target-conditioned fusion. DualKT constructs a structured question representation, encodes the full history into a long-term state with Mamba-2, and uses a target-querying hierarchical Transformer to extract a target-relevant short-term state from recent interactions and chunked full-prefix summaries. It decomposes the short-term state into a long-term-aligned component and an orthogonal residual, then fuses them with the long-term state conditioned on the target question. In five-fold question-level experiments on five public datasets, DualKT ranks among the top two methods in AUC on every dataset and first on four under fixed configurations. Ablations identify the structured question representation as the main source of improvement, with further gains from the dual-timescale branches and learned fusion. These results support explicit modeling and integration of long- and short-term cognitive information for target-specific prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
Knowledge Tracing, Dual-Timescale Modeling, Orthogonal Decomposition, Target-Conditioned Fusion

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

The paper proposes DualKT for question-level knowledge tracing. It combines a structured question representation incorporating question identity, question–KC graph structure, KC embeddings, and a difficulty residual with a Mamba-2 history encoder and a target-querying hierarchical Transformer(The entire work is a combination of existing technologies solid but with less innovation). The Transformer state is decomposed into components parallel and orthogonal to the Mamba state, followed by target-conditioned fusion and a transition-based correction. Experiments on five public datasets report competitive predictive performance, including the highest AUC on four datasets under fixed configurations and three datasets under unified TPE optimization.

Critical strengths

Broad empirical evaluation. The paper evaluates five datasets against 30 baselines and includes ablations, hyperparameter optimization, and a sequence-length sweep. The reported results establish DualKT as a competitive prediction system across several KT benchmarks.

Explicit handling of the prediction protocol. The manuscript describes student-disjoint splits, validation-based checkpoint selection, and question-level evaluation for multi-KC questions. The transition representation also explicitly uses the observed source response rather than the unknown target response.

Critical weaknesses
  1. The experiments do not establish the contribution of the central orthogonal-fusion mechanism. Section 3.4 reports that replacing the proposed fusion with Direct Fusion reduces macro-average AUC by only 0.05 percentage points, whereas removing the structured question representation reduces it by 3.23 percentage points. Moreover, Direct Fusion is a linear projection of concatenated states, while the proposed module includes target-conditioned routing, residual calibration, and additional correction networks. This comparison changes both the decomposition and the fusion architecture, so it cannot attribute the difference specifically to orthogonal projection. Figure 4 provides no uncertainty estimates for these small ablation differences. The largest ablation also jointly removes question identity, graph propagation, difficulty, and question–KC fusion, leaving their individual contributions unresolved. Consequently, the evidence supports the complete system more clearly than the specific mechanism emphasized in the title.
  2. The cognitive interpretation of the geometry is insufficiently justified. Equation 7 is a valid vector decomposition, but a component orthogonal to one learned state direction is not necessarily informationally nonredundant or a distinct cognitive fluctuation. The manuscript does not establish that alignment between the independently normalized branch states measures shared cognitive information. Furthermore, elementwise routing and calibration generally do not preserve orthogonality, and the correction networks in Equations 21–23 are not constrained to the corresponding subspaces. Thus, the initial projection does not establish that the downstream “aligned” and “residual” paths retain the proposed geometric interpretation.
  3. Distinct temporal roles are assumed more strongly than they are demonstrated. Appendix B.2 explicitly states that the Transformer memory covers the complete prefix and differs in resolution rather than using a separate temporal interval. Conversely, the Mamba branch has no explicit constraint requiring slowly varying states. The architecture therefore provides different ways of processing history, but does not establish a separation between persistent knowledge and short-term variation. Figures 6–7 show routing behavior and associations with prediction error on ASSIST2017; these observations do not demonstrate that the branches capture the claimed temporal factors or that orthogonal residuals provide uniquely complementary evidence. The single-branch and shallower ablations also change model capacity, complicating the attribution to temporal complementarity.
  4. The optimized comparisons leave important attribution and reproducibility gaps. Figure 3 tunes four competing methods, excluding strong Table 1 baselines such as DIMKT, UKT, and QIKT. It therefore establishes a ranking within that subset rather than across the broader competitive set. The submission also omits the concrete TPE search ranges, exact trial counts, and selected configurations; Appendix C.5 provides an additional figure rather than these details. No comparison equips a simpler encoder with the same structured question representation, despite that representation accounting for the largest reported ablation effect. Parameter counts, runtime, and memory costs are also absent, leaving the benefit of the additional architectural complexity difficult to assess. Overall, the predictive results are promising, but the submitted evidence does not sufficiently establish the independent value or scientific interpretation of the paper’s central contribution.
Questions for authors

Could the authors clarify how much of the reported improvement comes specifically from the orthogonal decomposition, given that the Direct Fusion comparison also changes the routing and correction components?

What evidence supports interpreting the orthogonal residual as complementary short-term cognitive information, especially since the subsequent element-wise routing and correction modules may alter the original geometric structure?

Could the authors clarify the exact TPE search spaces, trial budgets, and computational overhead used for DualKT and the compared baselines?

Reviewer_2BziOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualKT, a dual-timescale knowledge tracing framework that integrates long-term cognitive evolution and target-relevant short-term information. The model employs Mamba-2 for long-term history modeling and a hierarchical Transformer for short-term retrieval. An orthogonal decomposition mechanism separates short-term representations into aligned and complementary components, which are integrated through target-conditioned fusion. A structured question representation and transition graph further enhance prediction. Experiments on five public datasets demonstrate competitive AUC and ACC performance.

Critical strengths

1.The paper addresses the meaningful problem of modeling long-term and short-term learning dynamics in knowledge tracing.

  1. The orthogonal decomposition and target-conditioned fusion provide an interesting approach to integrating complementary temporal information.

  2. The structured question representation incorporates question identity, KC relations, and difficulty information.

  3. Extensive comparisons with 30 baselines across five public datasets demonstrate competitive predictive performance.

  4. The authors provide ablation studies, long-sequence experiments, and additional analyses of dynamic routing behavior.

Critical weaknesses

1.The methodological novelty is somewhat limited, as dual-timescale modeling, target-aware attention, and adaptive fusion have been investigated in previous KT studies. The distinction from FlucKT and MTKT requires stronger justification.

  1. The ablation results indicate that structured question representation contributes the largest improvement, while replacing orthogonal fusion with direct fusion produces relatively small performance decreases. This weakens the evidence supporting the paper's central innovation.

  2. The orthogonal decomposition is mathematically defined, but its connection to meaningful cognitive information remains insufficiently validated.

  3. The architecture combines multiple components, including Mamba-2, Transformer, graph representations, and several fusion mechanisms, without reporting computational complexity, parameter counts, or inference efficiency.

  4. The experiments primarily evaluate predictive accuracy, without sufficiently examining generalization to unseen questions or students across different educational environments.

Questions for authors

1.How does the proposed dual-timescale modeling fundamentally differ from FlucKT and MTKT beyond architectural implementation?

  1. Since removing structured question representation causes much larger performance degradation than replacing orthogonal fusion, can the authors provide stronger evidence that orthogonal decomposition is essential?

  2. Could the authors separately evaluate the aligned component, orthogonal residual, and residual calibration mechanism to quantify their individual contributions?

  3. What are the parameter counts, training costs, and inference latency compared with strong baselines such as DenoiseKT and FlucKT?

  4. Can the authors provide additional experiments evaluating cross-dataset transfer or generalization to unseen questions to demonstrate the robustness of the learned cognitive representations?

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

This paper presents DualKT, which combines Mamba-2 and hierarchical Transformer encoders to capture long-term and short-term learning states, followed by orthogonal decomposition and target-conditioned fusion. The evaluation is reasonably comprehensive, with comparisons against 30 baselines across five datasets and additional experiments on longer sequences.

Critical strengths

None.

Critical weaknesses

My main reservation concerns the contribution of the proposed orthogonal decomposition. Figure 4 shows that removing the structured question representation reduces AUC by 2.23 to 4.57 percentage points, whereas replacing orthogonal fusion with direct fusion causes a reduction of only 0.02 to 0.09 points. This suggests that the performance gains are driven primarily by question representation. Moreover, the question representation ablation removes several features simultaneously, making their individual contributions difficult to assess. A more convincing evaluation would retain the same question representation and compare orthogonal fusion against parameter-matched adaptive fusion alternatives.

The interpretation of the two cognitive states also deserves scrutiny. The short-term branch retrieves information from both recent interactions and the complete historical prefix, which weakens the claimed distinction between temporal scales. Furthermore, while Equation (7) constructs an orthogonal residual, the subsequent elementwise transformations do not necessarily preserve this property. It remains unclear whether the decomposition captures genuinely complementary cognitive information or simply provides another learned feature transformation.

The experimental results are promising, but the evidence for the main methodological contribution remains limited. I would be more positive if the authors could demonstrate a clear advantage of orthogonal decomposition over simpler fusion strategies under controlled settings.

Questions for authors

Please refer to the weakness section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.