acceptodds
Under review as a conference paper at ICLR 2027

DARIF: Delay and Density Aware Representation for Irregular Multivariate Time Series Forecasting

Abstract

In irregular multivariate time series (IMTS), measurements are recorded asynchronously across variables, creating two challenges for forecasting. First, existing methods usually perform information interaction among temporally aligned or nearby observations, which makes cross-variable delayed interactions with long time intervals difficult to model directly. Second, existing methods usually treat irregular sampling as a difficulty in modeling, and lack the utilization of observation density information brought by irregular sampling. To address the above challenges, we propose Delay- and Density-Aware Representation Learning for Irregular Multivariate Time Series Forecasting (DARIF) with two corresponding components: the Delay-Aware Module and Density-Aware Module. The Delay-Aware Module maintains a separate state for each source variable, decays these states according to the elapsed interval between active patches, and uses each target variable's query to weight information from previously observed sources. The Density-Aware Module estimates a continuous observation-time density for each variable and temporal patch using locally adaptive bandwidths, then encodes the resulting distributions for forecasting. Compared with state-of-the-art methods, DARIF achieves improvements of up to 6.8% across four benchmarks.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Irregular Multivariate Time Series Forecasting

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_49iUOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In order to capture asynchronous variable dependencies and account for varying sampling densities, the paper proposes D2ARIF, trying to address the challenge using the density encoder and decay factor. Experiments demonstrate competitive performance against state-of-the-art baselines.

Critical strengths
  1. Motivation is clear and sound, and the manuscript is likely written by human.
Critical weaknesses
  1. While the motivation is clear, the proposed method is not that elegant: The decay approach for asynchronous variable dependency learning is an implicit modeling rather than direct message passing among observations, and the density-aware module fuses its results with the delay-aware one only via concatenation+MLP (only realize 1+1=2 instead of 1+1>2). Even though the density-aware module is designed to be plug-and-play, it can be integrated with other modules in a more complementary way.
  2. No source code available during the review stage, and the experimental setting setup is vague, making the results less convincing. The experiment seem to be based on code pipeline PyOmniTS, but there's no clear explanation or citation. Also, using 10−210^{-2} to demenstrate benchmark results indicates the performance gain of the proposed method is marginal over baselines.
  3. The manuscript is not carefully examed: No citation for the four datasets; Inconsistent use of "Density-Aware Module" and "DensAM"; Typos like "moudel".
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 49iU-W1/W2/W3

Weakness 1

Graph-based methods such as ASTGI [1] explicitly select interaction partners through kNN-based sparse observation-level connections. If a relevant observation is excluded from the selected neighborhood, its interaction cannot be captured through a direct message passing edge. In contrast, D²ARIF allows each target variable to retrieve information from all previously observed source-variable memories without preselecting neighbors or constructing edges.

To examine whether the Delay-Aware and Density-Aware Modules could be coupled more tightly, we evaluated three variants in which density information directly modulates memory updates. Density-modulated writing: after layer normalization and a linear projection, the density representation is added to the input of the sigmoid write gate in Eq. (14) to control the strength of the current observation’s memory update. Density-modulated decay: the corresponding projection is added to the input of the softplus function governing the decay rate in Eq. (12) to modulate the retention of historical information. Joint density modulation of writing and decay: both modifications are applied simultaneously. These results support concatenation followed by an MLP as sufficient to combine the modules’ complementary signals on the benchmarks.

Each cell reports MSE / MAE in units of 10−210^{-2}.

Fusion strategyHumanActivityUSHCNPhysioNetMIMIC-III
Density-modulated writing4.560±0.14 / 11.36±0.1216.36±0.69 / 22.55±0.4031.76±0.15 / 37.84±0.0842.37±1.24 / 40.03±0.62
Density-modulated decay4.547±0.11 / 11.38±0.1815.78±0.49 / 21.96±0.3731.88±0.30 / 37.95±0.0643.11±0.32 / 39.89±0.13
Joint density-modulated writing and decay4.161±0.03 / 11.28±0.0115.47±0.42 / 22.13±0.3631.75±0.10 / 37.84±0.0448.11±0.62 / 43.95±0.40
Concatenation+MLP4.142±0.04 / 11.10±0.0614.15±0.57 / 20.97±0.4130.69±0.09 / 35.30±0.0942.50±0.28 / 38.61±0.69

[1] Liu, X., et al. ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting. ICLR, 2026.

Weakness 2

Thank you for raising this issue. Our implementation builds on the official TFMixer repository, which we cited in the manuscript. Since TFMixer was developed on top of PyOmniTS, our implementation also indirectly depends on PyOmniTS. We inadvertently omitted a separate citation to PyOmniTS. Our code is available for review in the https://anonymous.4open.science/r/D2ARIF-FCC3/. Upon acceptance, we will release it publicly and submit it to the PyOmniTS project. We will add this citation in the revision and provide more detailed experimental settings below.

Appendix A.4 reports the hardware and main training settings (AdamW, learning rate 10−310^{-3}, weight decay 0.01, up to 200 epochs). Appendices A.1–A.4 describe synthetic data generation, splitting, normalization, and validation-based selection.

For the real-world experiments, we follow TFMixer’s preprocessing and evaluation. We reserve 10% of records for testing and split the remainder 90/10 for training and validation (approximately 81% / 9% / 10% overall). We split HumanActivity records before constructing windows. All methods use the same splits, forecasting targets, and evaluation protocol. We select checkpoints by validation MAE and compute test MSE and MAE only on observed future targets, excluding missing and padded positions. Reproduced and additional experiments report means and standard deviations across seeds 2024, 2025, and 2026.

DatasetHistory windowForecasting horizon
HumanActivity3000 msObservations in the following 300 ms
USHCN3 yearsNext three observation timestamps
PhysioNet 201236 hoursNext three observation timestamps
MIMIC-III36 hoursNext three observation timestamps

The ×10−2\times 10^{-2} display scale preserves relative differences and the resulting comparisons between methods. For example, in Table 1, the USHCN MSE decreases from 15.18 to 14.15, corresponding to a relative improvement of 6.79%. Such relative comparisons are invariant to the display scale.

Weakness 3

We thank the reviewer for pointing out these specific issues. We will cite the original sources for all four datasets. Regarding the abbreviation DensAM, the caption of Table 3 already states that "DensAM denotes the Density-Aware Module." We used this abbreviation only to avoid an overly long table entry. In the revised manuscript, we will consistently use Density-Aware Module instead of the abbreviation. And we have also carefully proofread the manuscript and corrected spelling mistakes, including “moudel,” as well as other grammatical issues.

Thanks for your explanations. The primary weakness of the paper is still the novelty part. Bringing your insights into paper makes it an appealing one, but unfortunately there seems to lack such highlights in this manuscript. Therefore, I would still keep my score as it is.

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

This paper proposes D²ARIF, a framework for irregular multivariate time series forecasting that combines a Delay-Aware Module and a Density-Aware Module. The former maintains source-specific memory states with elapsed-time decay and target-dependent retrieval to capture delayed cross-variable dependencies. The latter employs adaptive kernel density estimation to encode local observation-time distributions within temporal patches. The two representations are fused for forecasting.

The method is evaluated on four real-world datasets against 16 baselines. Additional experiments include module ablations, integration of the density module into five existing backbones, and controlled synthetic experiments targeting delayed dependencies and observation-time distributions. D²ARIF achieves the best results on six of eight main benchmark metrics.

Critical strengths
  1. The paper addresses two relevant challenges in irregular time series forecasting: delayed cross-variable interactions and nonuniform observation-time distributions. Maintaining source-specific memory states and explicitly encoding local timestamp distributions are reasonable modeling choices that may provide useful inductive biases.

  2. The paper compares against 16 baselines on four datasets and includes ablation studies, synthetic experiments, and backbone integration experiments. In particular, the Density-Aware Module improves MSE across all 20 reported backbone–dataset combinations, providing evidence of its practical applicability beyond the proposed architecture.

Critical weaknesses
  1. The Delay-Aware Module primarily combines existing Delta Rule memory updates, temporal decay, and attention-based retrieval, while the Density-Aware Module builds on established adaptive KDE techniques. The paper does not clearly demonstrate a substantial methodological advance beyond combining these existing mechanisms.

  2. The experiments do not separately evaluate source-specific memory, elapsed-time decay, or target-specific retrieval. Similarly, adaptive KDE is not compared with simpler alternatives such as fixed-bandwidth KDE or timestamp histograms. Consequently, the necessity of the proposed mechanisms remains unclear.

  3. Equations (2) and (6) use sqrt(2πw) as the Gaussian kernel normalization factor, whereas the correct expression is sqrt(2π)w when w denotes the standard deviation. This error may affect density estimation and bandwidth optimization, and its consistency with the actual implementation needs verification.

  4. The most closely related density-based method, HEADS, is not included in the real-world benchmark comparisons. Moreover, important details regarding dataset splits, forecasting configurations, preprocessing, and experimental repetitions are insufficiently documented, making it difficult to assess experimental fairness and reproduce the results.

I am open to reconsidering my rating if the authors can adequately address my concerns, particularly regarding the algorithmic novelty, mathematical correctness, and experimental reliability, and provide convincing supporting evidence during the rebuttal.

Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer rubM-W1/W2

Weakness 1

We thank the reviewer for this important comment. The methodological contribution of our work lies in addressing two problems that existing approaches do not adequately resolve.

First, direct access to long-delayed asynchronous source histories. Existing methods model cross-variable dependencies through local interactions, kNN-based sparse observation-level connections (e.g., ASTGI [1]), or shared states that mix information across variables [2, 3]. The former may exclude relevant long-delayed observations from direct message passing, while the latter does not maintain independently accessible source-specific histories for target-specific retrieval. In contrast, D²ARIF combines per-source memory, elapsed-time decay, and target-specific retrieval, allowing each target to access all previously observed source-variable memories without preselecting neighbors (Eqs. (12)–(16)).

Second, explicit modeling of within-patch timestamp distributions. Existing methods rely on counts, masks, or window-level density [4], which may overlook differences in observation timing, while GraFITi [5] and TFMixer [6] encode timestamps without separately modeling their within-patch distributions. In contrast, D²ARIF uses locally adaptive KDE to capture each variable's timestamp distribution as a separate feature (Eqs. (1)–(8)). This design reduces error by approximately 15%–20% compared with HEADS [4] under fixed observation counts but different temporal layouts (Appendix Table 7), and improves all backbone–dataset combinations, including GraFITi [5] and TFMixer [6] (Table 3).

[1] Liu, X., et al. ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting. ICLR, 2026.

[2] Rubanova, Y., Chen, R. T. Q., & Duvenaud, D. Latent Ordinary Differential Equations for Irregularly-Sampled Time Series. NeurIPS, 2019.

[3] Schirmer, M., Eltayeb, M., Lessmann, S., & Rudolph, M. Modeling Irregular Time Series with Continuous Recurrent Units. ICML, 2022.

[4] Hu, Y., et al. HEADS: Temporal Observation Density Boosts Irregular Multivariate Time Series Forecasting. DASFAA, 2026.

[5] Yalavarthi, V. K., et al. GraFITi: Graphs for Forecasting Irregularly Sampled Time Series. AAAI, 2024.

[6] Qiu, X., et al. Bridging Time and Frequency: A Joint Modeling Framework for Irregular Multivariate Time Series Forecasting. ICML, 2026.

Weakness 2

Thank you for suggesting these ablations. We tested each design independently by sharing memory across sources at matched total memory-state capacity, removing elapsed-time decay, or making retrieval target-agnostic.

Relative to the corresponding ablations, source-specific memory reduces MSE by 6.12% on HumanActivity and 36.09% on USHCN. Elapsed-time decay reduces MSE by 28.32% on USHCN and 6.49% on MIMIC-III. Target-specific retrieval reduces MAE by 6.17% on PhysioNet and MSE by 5.32% on MIMIC-III. These results indicate that each design contributes to forecasting accuracy, with its impact varying across datasets.

Each cell reports MSE / MAE in units of 10−210^{-2}.

MethodHumanActivityUSHCNPhysioNetMIMIC-III
Shared-state memory4.412±0.03 / 11.16±0.0422.14±0.79 / 24.21±0.4130.97±0.25 / 37.56±0.0943.58±0.36 / 39.26±0.14
No temporal decay4.243±0.04 / 11.12±0.0619.74±1.74 / 24.84±0.8032.11±0.05 / 37.50±0.0445.45±1.00 / 40.05±0.55
Target-agnostic retrieval4.265±0.04 / 11.15±0.0514.46±0.63 / 21.66±0.4731.11±0.27 / 37.62±0.0444.89±0.96 / 39.37±0.51
D²ARIF4.142±0.04 / 11.10±0.0614.15±0.57 / 20.97±0.4130.69±0.09 / 35.30±0.0942.50±0.28 / 38.61±0.69

For the Density-Aware Module, we tested fixed-bandwidth KDE and timestamp histograms. To isolate the effect of locally adapting the smoothing scale, the fixed-bandwidth variant uses one validation-selected bandwidth per dataset, shared across samples, variables, and patches. Histograms test a discretized representation using normalized counts in predefined within-patch time bins.

Averaged across datasets, D²ARIF reduces MSE / MAE by 12.8% / 10.0% versus fixed-bandwidth KDE, 6.3% / 8.1% versus histograms, and 8.4% / 8.5% versus observation counts and masks. The gains support locally adaptive density estimation over the simpler timestamp representations evaluated here.

Each cell reports MSE / MAE in units of 10−210^{-2}.

MethodHumanActivityUSHCNPhysioNetMIMIC-III
Fixed-bandwidth KDE4.456±0.24 / 12.27±0.0721.57±1.61 / 25.34±0.4131.91±0.21 / 37.51±0.0645.26±0.73 / 41.72±0.24
Timestamp histogram4.467±0.11 / 12.54±0.1816.24±2.41 / 22.72±1.0931.91±0.13 / 37.50±0.0743.04±0.30 / 41.67±0.29
D²ARIF4.142±0.04 / 11.10±0.0614.15±0.57 / 20.97±0.4130.69±0.09 / 35.30±0.0942.50±0.28 / 38.61±0.69
AuthorsOpenReview ↗

Response to Reviewer rubM-W3/W4

Weakness 3

We thank the reviewer for carefully examining the equations. We rechecked Eqs. (2) and (6) and confirmed that the normalization factor in the original manuscript already uses the correct form identified by the reviewer. To make the notation unambiguous, we will rewrite 12πw\frac{1}{\sqrt{2\pi}w} as 1w2π\frac{1}{w\sqrt{2\pi}} in the revised manuscript.

Weakness 4

We compared HEADS and our Density-Aware Module as plug-ins for five backbones. Averaged equally across all 20 backbone–dataset pairs, HEADS reduced MSE by 3.00% and our module by 5.75% relative to the original backbones. Our module improved MSE over the original backbone in every pair, supporting its use as a plug-in across architectures. MSE is reported in units of 10−210^{-2}.

BackboneVariantHumanActivityUSHCNPhysioNetMIMIC-IIIImprovement
PrimeNetOriginal425.1±0.4149.30±0.1579.53±0.0090.73±0.01–
+HEADS431.0±0.1129.49±0.2275.26±0.0196.72±0.019.39%
+Density-Aware Module419.2±0.4835.49±0.2775.39±0.1685.29±0.2510.15%
GRU-DOriginal18.93±6.2720.97±4.9334.19±0.2947.59±1.00–
+HEADS18.91±0.3020.94±3.8833.64±0.1345.82±0.421.39%
+Density-Aware Module18.72±2.0417.09±2.0132.48±0.2444.64±0.687.70%
RaindropOriginal9.161±0.7220.35±3.3634.78±0.1967.54±18.3–
+HEADS9.149±0.4018.74±2.5434.35±0.1866.92±13.42.55%
+Density-Aware Module9.088±0.4418.31±1.9733.08±0.1663.42±8.735.45%
GraFITiOriginal4.370±0.0516.91±0.9331.75±0.1543.59±4.55–
+HEADS4.378±0.0815.38±0.1930.65±0.0941.69±1.724.17%
+Density-Aware Module4.344±0.0515.22±0.6830.87±0.1242.84±2.193.77%
TFMixerOriginal4.153±0.0415.18±0.5131.18±0.2341.82±0.37–
+HEADS4.409±0.0715.12±1.0931.91±0.5342.64±0.50-2.52%
+Density-Aware Module4.144±0.0414.26±0.4531.13±0.1941.71±0.321.68%

Appendix A.4 reports the hardware and main training settings (AdamW, learning rate 10−310^{-3}, weight decay 0.01, up to 200 epochs). Appendices A.1–A.4 describe synthetic data generation, splitting, normalization, and validation-based selection.

For the real-world experiments, we follow TFMixer’s preprocessing and evaluation. We reserve 10% of records for testing and split the remainder 90/10 for training and validation (approximately 81% / 9% / 10% overall). We split HumanActivity records before constructing windows. All methods use the same splits, forecasting targets, and evaluation protocol. We select checkpoints by validation MAE and compute test MSE and MAE only on observed future targets, excluding missing and padded positions. Reproduced and additional experiments report means and standard deviations across seeds 2024, 2025, and 2026.

DatasetHistory windowForecasting horizon
HumanActivity3000 msObservations in the following 300 ms
USHCN3 yearsNext three observation timestamps
PhysioNet 201236 hoursNext three observation timestamps
MIMIC-III36 hoursNext three observation timestamps
Reviewer_pXAfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes D2ARIF for irregular multivariate time-series forecasting. Its Delay-Aware Module maintains a separate delta-rule memory for each source variable, decays states according to elapsed time, and lets each target retrieve a weighted combination of source histories. Its Density-Aware Module estimates per-variable, per-patch timestamp distributions with locally adaptive kernel density estimation and encodes them for forecasting. The two representations are fused with future query-time embeddings. On HumanActivity, USHCN, PhysioNet, and MIMIC, D2ARIF ranks first on six of eight MSE/MAE metrics and achieves the lowest MAE on all four datasets. Ablations replace the delay-aware branch with self-attention and remove the density branch; an additional study adds the density module to five backbones and improves MSE in all 20 model-dataset combinations. Synthetic experiments vary source-target delay and within-patch timestamp distributions.

Critical strengths
  1. The model addresses two meaningful problems: useful information may arrive from another variable after a delay, and observation timing can itself help prediction. The source-specific memory and density branch match these goals.
  2. The reported results are competitive: D2ARIF has the lowest MAE on all four datasets, and adding the density branch lowers MSE in all 20 tested model-dataset combinations.
  3. The controlled studies go beyond standard benchmarks. They compare several memory methods and change timestamp patterns while keeping observation counts fixed.
Critical weaknesses
  1. The paper explains its adaptations of existing memory methods, but their individual benefits remain unclear. Appendix Table 6 includes Delta and Gated Delta comparisons; it still does not separately test source-specific memory, elapsed-time decay, and target-specific retrieval.
  2. Selecting a separate density bandwidth for each variable and patch adds processing steps, yet runtime and memory costs are not reported. Without these measurements, the practical cost of the accuracy gains is unclear.
  3. The controlled experiments report one result per model and condition without variation across seeds. The delay study also trains a separate model for each delay, so it does not show whether the model handles delays that change after training.
Questions for authors

See Weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer pXAf-W1/W2

Weakness1

Thank you for suggesting these ablations. We tested each design independently by sharing memory across sources at matched total memory-state capacity, removing elapsed-time decay, or making retrieval target-agnostic. Relative to the corresponding ablations, source-specific memory reduces MSE by 6.12% on HumanActivity and 36.09% on USHCN. Elapsed-time decay reduces MSE by 28.32% on USHCN and 6.49% on MIMIC-III. Target-specific retrieval reduces MAE by 6.17% on PhysioNet and MSE by 5.32% on MIMIC-III. These results indicate that each design contributes to forecasting accuracy, with its impact varying across datasets.

Each cell reports MSE / MAE in units of 10−210^{-2}.

MethodHumanActivityUSHCNPhysioNetMIMIC-III
Shared-state memory4.412±0.03 / 11.16±0.0422.14±0.79 / 24.21±0.4130.97±0.25 / 37.56±0.0943.58±0.36 / 39.26±0.14
No temporal decay4.243±0.04 / 11.12±0.0619.74±1.74 / 24.84±0.8032.11±0.05 / 37.50±0.0445.45±1.00 / 40.05±0.55
Target-agnostic retrieval4.265±0.04 / 11.15±0.0514.46±0.63 / 21.66±0.4731.11±0.27 / 37.62±0.0444.89±0.96 / 39.37±0.51
D²ARIF4.142±0.04 / 11.10±0.0614.15±0.57 / 20.97±0.4130.69±0.09 / 35.30±0.0942.50±0.28 / 38.61±0.69

Weakness2

We measured total test-set inference time and peak allocated GPU memory on all four datasets. D²ARIF improves forecasting accuracy with inference times on the same order of magnitude as GraFITi and TFMixer. Its peak GPU memory use is higher than GraFITi’s and approximately 78.5% lower than TFMixer’s on average. Overall, D²ARIF improves forecasting accuracy while maintaining inference efficiency comparable to existing leading methods and manageable resource requirements.

Each cell reports total test-set inference time (seconds) / peak allocated GPU memory (MiB).

MethodUSHCNHumanActivityP12MIMIC-III
GraFITi0.0395 / 271.10.0551 / 75.20.3825 / 95.60.9193 / 263.6
TFMixer0.0477 / 1870.50.0401 / 849.40.3098 / 997.61.7026 / 4667.8
D²ARIF0.0618 / 395.70.0770 / 188.90.3669 / 215.71.3887 / 989.0
AuthorsOpenReview ↗

Response to Reviewer pXAf-W3

Weakness 3

Both protocols use a history length of 96 and a forecast length of 24. We evaluate each trained model with fixed weights at the same seven delays and report mean ± SD for MSE / MAE across three independent runs with seeds 2024, 2025, and 2026.

D²ARIF achieves the lowest mean MSE and MAE at all seven test delays in both protocols, showing that its relative advantage extends to unseen delays with the same trained weights. Under fixed-delay training, absolute error increases as the test delay moves farther from 84. These experiments show that the performance advantage of D²ARIF does not depend on training a separate model for each delay and demonstrate its ability to adapt to different delay conditions.

Fixed-delay training. We trained and validated at delay 84.

Each cell reports MSE / MAE.

DelaySelf-attentionDeltaGated DeltaD²ARIF
603.487±0.677 / 1.482±0.1663.428±0.679 / 1.468±0.1693.368±0.657 / 1.461±0.1603.262±0.679 / 1.432±0.167
663.274±0.574 / 1.444±0.1483.240±0.587 / 1.433±0.1503.148±0.564 / 1.410±0.1423.003±0.585 / 1.381±0.150
722.851±0.444 / 1.345±0.1202.736±0.452 / 1.325±0.1252.743±0.431 / 1.331±0.1152.713±0.455 / 1.317±0.124
781.759±0.072 / 1.056±0.0261.678±0.044 / 1.032±0.0151.688±0.086 / 1.037±0.0241.643±0.088 / 1.010±0.024
840.636±0.063 / 0.623±0.0310.614±0.041 / 0.607±0.0230.648±0.040 / 0.631±0.0230.591±0.029 / 0.597±0.017
901.757±0.160 / 1.053±0.0501.798±0.054 / 1.071±0.0151.691±0.168 / 1.029±0.0551.564±0.090 / 0.993±0.027
962.804±0.444 / 1.342±0.1182.899±0.432 / 1.370±0.1162.840±0.440 / 1.349±0.1162.549±0.440 / 1.277±0.117

Mixed-delay training. We assigned each training trajectory one fixed delay of 72, 84, or 96, balancing assignments across delays and withholding the delay labels from the model.

Each cell reports MSE / MAE.

DelaySelf-attentionDeltaGated DeltaD²ARIF
602.520±0.307 / 1.249±0.0772.333±0.127 / 1.196±0.0302.490±0.248 / 1.246±0.0642.121±0.244 / 1.138±0.062
661.905±0.085 / 1.082±0.0271.894±0.041 / 1.070±0.0081.984±0.101 / 1.102±0.0341.790±0.067 / 1.043±0.024
721.398±0.184 / 0.928±0.0661.479±0.047 / 0.955±0.0171.504±0.194 / 0.964±0.0661.380±0.157 / 0.923±0.055
781.319±0.222 / 0.901±0.0781.309±0.044 / 0.906±0.0171.372±0.237 / 0.926±0.0811.205±0.204 / 0.868±0.068
841.343±0.252 / 0.907±0.0831.293±0.050 / 0.897±0.0161.314±0.268 / 0.906±0.0861.210±0.254 / 0.869±0.083
901.326±0.246 / 0.904±0.0841.312±0.057 / 0.898±0.0241.458±0.231 / 0.949±0.0811.309±0.252 / 0.892±0.086
961.496±0.181 / 0.950±0.0671.540±0.065 / 0.970±0.0261.458±0.159 / 0.949±0.0631.402±0.192 / 0.930±0.068

I thank the authors for their detailed responses. Reporting the computational cost is helpful.

The new ablation study on the memory design shows that the current design yields the largest benefit. However, performance varies across designs and datasets, and there is no discussion of why. The role of each component remains unclear, and in some cases the improvement is negligible.

The new synthetic experiments show empirically that the method’s relative advantage extends to unseen delays with the same trained weights. However, performance and gains vary across different types of delays, and this variation is not analyzed.

I therefore maintain my original scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.