acceptodds
Under review as a conference paper at ICLR 2027

SPS: Sufficient Patch Selection for Information-Preserving Time Series Forecasting

Abstract

Patch-based tokenization has become a standard design choice for time series forecasting. Most existing methods treat patching as a fixed tokenization step, generating local windows and passing all resulting patches to the forecasting model regardless of their predictive utility. We instead formulate patching as a two-stage process: candidate generation followed by candidate selection. Building on this view, we propose the **Sufficient Patch Selector (SPS)**, a plug-in module for information-preserving patch selection. SPS constructs spectrally guided multi-scale candidate intervals from dominant temporal periods, then learns to retain a compact subset that preserves the predictive distribution induced by the full lookback sequence. To train SPS, we introduce an approximate posterior-preservation objective that matches the selected-patch predictive distribution to a full lookback reference, encouraging retained patches to be sufficient for prediction. We further propose Information Preservation Sufficiency (IPS), a metric that quantifies how well a patching strategy preserves full lookback predictive information. Across standard multivariate forecasting benchmarks and multiple patch-based backbones, SPS improves information preservation and achieves strong forecasting accuracy while retaining fewer patches than existing patching baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Patching, Adaptive Patch Selection, Multi-Scale Patching, Information Preservation Sufficiency, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xrTBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Sufficient Patch Selector (SPS), a plug-in framework for long-term time series forecasting that reformulates patching as two stages: candidate generation and candidate selection. SPS constructs multi-scale patch candidates using dominant periods identified through spectral analysis, and learns to retain a compact subset under a fixed patch budget. It also introduce a posterior-preservation objective that minimizes the KL divergence between the selected-patch predictor and a full-history teacher.

Critical strengths

1.The paper provides a useful two-stage perspective that explicitly separates the construction of candidate patches from the selection of informative patches.

2.Meaningful empirical evaluation and useful controlled comparisons.

Critical weaknesses

1.The experiments do not directly compare the spectral generator against fixed-scale, uniformly sampled multi-scale, or random multi-scale generators under identical candidate and selection budgets.

2.A reduction in IPS demonstrates closer agreement with the chosen teacher, rather than directly establishing preservation of the true predictive information. The manuscript does not provide sufficient evidence that the IPS-based conclusions are insensitive to the teacher architecture or predictive uncertainty calibration.

3.The selector uses a global context vector derived from the entire historical input, and the selected representation is produced through mask-aware attention. This complicates the interpretation of the output as information retained solely within the selected patches.

Questions for authors

See critical weakness

6 comments

We thank the reviewer for the valuable and constructive feedback.

W1. The spectral generator is not compared with fixed-scale, uniformly sampled multi-scale, or random multi-scale generators under identical candidate and selection budgets.

[Experiment 2a] Candidate Generator Controls. Only the period list changes. Every variant has N=80 candidates and K=10 selected patches, with the same learned selector, encoder, pooling, KL and training recipe. SPSNet, 29 settings (seven datasets × four horizons, and Traffic at H=96). MSE and MAE are means of the eight dataset averages, and relative changes are averaged over settings:

Candidate familyDistinct lengthsMSEMAEMSE vs. spectralSpectral better inTraining time vs. spectral
Spectral (submission)4–100.3170.320––1.00
Fixed multi-scale {8, 16, 32, 64}40.3220.323+1.6%22/290.89–1.07
Random lengths (3 draws)6–120.3240.326+2.2%24/290.92–1.20
Uniformly sampled grid510.3250.326+2.5%26/291.74–2.19
Fixed single-scale, P=1610.3270.329+3.2%27/290.81–0.90
  • All three generators you name. The spectral period list has the lowest MSE and MAE. A fixed single scale is 3.2% worse (27/29), a uniformly sampled multi-scale grid 2.5% worse (26/29), and random multi-scale lengths 2.2% worse (24/29). We also add a fixed multi-scale list, which is 1.6% worse (22/29).
  • Multi-scale against single-scale. The single scale P=16 is the weakest family. A fixed multi-scale list closes half of its gap to the spectral list, and the spectral choice of scales closes the other half.
  • More scales are not better. The uniformly sampled grid has 51 distinct lengths, takes 1.74–2.19× the training time, and is worse than four fixed scales. The gain comes from which scales are used, not from how many.
  • Size of the effect. The differences are 0.005–0.010 in MSE. They are consistent across settings but modest, and the revision describes them in these terms.

Summary. Under identical candidate and selection budgets, the spectral generator is better than fixed-scale, uniformly sampled multi-scale and random multi-scale generators, by 1.6–3.2% in MSE and in 22–27 of 29 settings. The revision adds this table to the ablation study (Appendix J).

W2. Lower IPS shows closer agreement with the chosen teacher, not preservation of the true predictive information.

We agree, and [Experiment 1] tests both sensitivities: the values of IPS depend strongly on the variance calibration and on the teacher, while the orderings between methods are stable. We now report IPS as teacher-relative.

Proposition 1. For any predictive distribution qm(Y∣Z)q_m(Y\mid Z) with Z=S(X)Z=\mathcal{S}(X),

E[DKL(P(Y∣X) ∥ qm(Y∣Z))]≥I(Y;X∣Z),E[−log⁡qm(Y∣Z)]=H(Y∣X)+E[DKL(P(Y∣X) ∥ qm(Y∣Z))].\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big] \ge I(Y;X\mid Z), \qquad \mathbb{E}\big[-\log q_m(Y\mid Z)\big] = H(Y\mid X) + \mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big].

The gap in the inequality is E[DKL(P(Y∣Z) ∥ qm(Y∣Z))]\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid Z)\,\|\,q_m(Y\mid Z))\big], the approximation error of the predictor. IPS therefore bounds the information loss only if its reference is the true P(Y∣X)P(Y\mid X), and no such bound holds for a teacher. The difference in held-out NLL between two methods equals the difference in their exact-reference IPS, which gives a teacher-free test.

[Experiment 1a] Teacher Accuracy and Calibration. Mean over four horizons, with each model's own variance head:

DatasetTeacher MSE / MAESPSNet MSE / MAETeacher NLL / CRPS / 90% coverageSPSNet NLL / CRPS / 90% coverage
ETTh10.464 / 0.4510.434 / 0.4280.827 / 0.304 / 0.9480.786 / 0.298 / 0.924
ETTh20.385 / 0.4050.369 / 0.3940.895 / 0.287 / 0.9350.736 / 0.288 / 0.901
ETTm10.415 / 0.4130.377 / 0.3870.879 / 0.277 / 0.9440.697 / 0.267 / 0.888
ETTm20.285 / 0.3260.277 / 0.3190.769 / 0.231 / 0.9240.686 / 0.240 / 0.889
Weather0.270 / 0.2890.253 / 0.2730.705 / 0.202 / 0.9690.684 / 0.187 / 0.935
Solar0.228 / 0.2410.193 / 0.2230.640 / 0.160 / 0.9830.631 / 0.161 / 0.970
Electricity0.183 / 0.2730.181 / 0.2700.572 / 0.188 / 0.9290.537 / 0.189 / 0.911
Traffic0.430 / 0.2830.468 / 0.2840.490 / 0.201 / 0.9480.410 / 0.210 / 0.935
Mean0.333 / 0.3350.319 / 0.3220.722 / 0.231 / 0.9480.646 / 0.230 / 0.919
  • Accuracy. The teacher is not an upper bound: SPSNet has lower MSE in 25/32 settings and lower MAE in 28/32. The teacher is more accurate on Traffic. Its average error is close to that of the iTransformer column of Table 1 (0.329 / 0.336), so it is a standard full-history model.
  • Calibration. The teacher is over-dispersed. Its nominal 90% intervals cover 92.4–98.3% of the test targets, 94.8% on average. SPSNet's intervals cover 91.9% on average and are closer to 90% on all eight datasets, with slight under-coverage on ETTm1 and ETTm2 (88.8% and 88.9%). SPSNet also has the lower NLL on all eight datasets (0.646 against 0.722), while CRPS is tied (0.230 against 0.231).
  • Caveat. The NLL gap reflects the variances more than the means: the teacher has the lower CRPS on five datasets, and on Traffic it has the lower MSE and still the higher NLL. We do not claim that SPSNet is the better probabilistic forecaster. The table shows that the teacher is not a close stand-in for P(Y∣X)P(Y\mid X).

[Experiment 1b] Variance-Robust and Teacher-Free IPS. We recompute IPS on the 60 plug-in settings (three backbones × five datasets × four horizons) under four variance treatments: the submission's heads, the same heads calibrated on validation, all heads refitted with one NLL recipe, and one shared variance, where only forecast disagreement counts.

IPS computed with+SPS below its backbone+SPS below +KL+KL below its backboneMean rank of SPSNet among 10 models
submission heads60/6054/6060/601.00
calibrated heads60/6049/6060/601.00
refitted heads60/6051/6059/601.00
shared variance60/6052/6060/601.00
  • Sensitivity to the variance calibration: magnitudes. The variance term is 92–99% of every IPS value in the submission. The size of an IPS value, and the percentage reductions in Tables 6 and 7, therefore mostly reflect how each model's variance head was fitted. The revision reports orderings instead.
  • Sensitivity to the variance calibration: orderings. They are stable. +SPS is below its backbone in 60/60 settings under all four treatments and below +KL in 49–54/60. SPSNet has rank 1.00 among the 10 models (SPSNet, the three backbones, and their +KL and +SPS variants). With a shared variance, IPS reduces to the distance between the student's and the teacher's forecasts, and the conclusions are the same.
  • Teacher-free test. Held-out NLL needs no teacher, and by Proposition 1 it orders methods as exact-reference IPS would. In [Experiment 1a], SPSNet's NLL is below that of the full-history teacher on all eight datasets. In exact-reference terms the teacher is further from P(Y∣X)P(Y\mid X) than its student, which is one more reason not to read teacher-relative IPS as lost information.

[Experiment 1c] Sensitivity to the Teacher. We train SPSNet at H=96 with teachers stopped after 1, 3 and 10 epochs, with the full teacher, with no teacher, and with a teacher of SPSNet's own architecture without patch selection. IPS uses the submission's heads. Mean over six datasets, relative changes are averaged over datasets:

Training teacherTeacher MSESPSNet MSESPSNet MSE vs. full teacherIPS vs. its own teacherIPS vs. the full teacher
1 epoch0.3250.255+4.7%0.8360.776
3 epochs0.2890.252+3.2%0.4620.410
10 epochs0.2680.248+0.9%0.2980.252
full (submission)0.2590.247–0.2090.209
none (no KL)–0.252+2.9%–0.408
architecture-matched0.2490.247+0.1%0.2230.219
  • Teacher quality matters for accuracy. Removing the full teacher raises MSE by 2.9%, and a 1-epoch teacher is worse than none.
  • IPS depends on the teacher. The same student architecture has an IPS between 0.209 and 0.836 depending on its teacher. An IPS value describes a student–teacher pair and not a patching strategy alone.
  • A weak teacher is not copied. With the 1-epoch teacher (MSE 0.325), SPSNet reaches 0.255, and its IPS against that teacher is the highest in the table (0.836). The supervised loss keeps the student away from a poor teacher, so a low IPS does not arise from imitating one.
  • Sensitivity to the teacher architecture. Trained with a teacher of SPSNet's own architecture, SPSNet's MSE changes by 0.1%. Its IPS against the iTransformer teacher, which it never saw in training, is 0.219, against 0.209 when trained with that teacher and 0.408 with no teacher. The low IPS of SPSNet is therefore not tied to one teacher architecture. This experiment varies the teacher for SPSNet. It does not re-rank the baselines against a second teacher.

Summary. IPS shows agreement with the chosen teacher and not preservation of the true predictive information. The teacher is as accurate as a standard iTransformer but over-dispersed, and it is not an upper bound on SPSNet. The magnitudes of IPS are driven by the variance heads, while the orderings are unchanged under four variance treatments and under a teacher of a different architecture. The revision defines IPS as teacher-relative, adds Proposition 1 and [Experiment 1], and reports held-out NLL as the teacher-free comparison.

W3. The selector uses a global context vector, and the output is produced through mask-aware attention, which complicates reading it as information retained solely within the selected patches.

We agree that the output cannot be read as information retained solely within the selected patches. The revision spells out the mechanism in Section 3.4 and Appendix C:

  • Tokens are local, and g is used only for scoring. Each candidate token is computed from its own interval and position (Equation 8). The global context g enters the scorer (Equations 10–11) and is not passed downstream.
  • Pooling reads only the selected tokens. At inference the mask is binary, so unselected candidates receive exactly zero attention (Equation 28).
  • The index set depends on the full history. Which K candidates are kept is a function of the full history, and it can carry up to log⁡2(8010)≈40.6\log_2\binom{80}{10}\approx 40.6 bits per series.

Three measurements show how much each route matters.

[Experiment 2b] Selector Controls. Only the selection rule changes, on identical spectral candidates. "All pooled" attends over all 80 candidates into the same K output slots:

SelectorMSEMAEMSE vs. learnedLearned better in
Learned (submission)0.3170.320––
Learned, zeroed global context0.3180.320+0.3%18/29
All pooled0.3220.323+1.6%23/29
Recency (K latest-ending candidates)0.3230.323+1.9%23/29
Uniform (every N/K-th candidate)0.3240.324+2.2%25/29
Random K per series0.3250.325+2.5%27/29
  • Global context (first measurement). Zeroing g changes MSE by 0.3% and leaves MAE unchanged, and the learned selector is better in 18/29 settings. The scores rely mainly on the content of each candidate, and the global context adds little.
  • Index set (second measurement). Recency, uniform and random selection keep K candidates without looking at the content of the history. They are 1.9–2.5% worse than learned selection. This is the whole value of choosing the candidates as a function of the history, so it is an upper bound on what the index set contributes as a side channel.
  • Attention over the candidates. "All pooled" lets the K slots attend to all 80 candidates and is 1.6% worse than attending to the selected 10 (23/29). The accuracy of SPS does not rely on access to the unselected candidates.
Reviewer_zvKJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper decomposes time-series patching into candidate generation and candidate selection. SPS constructs multiscale intervals from dominant periods estimated on the training data and scores candidates using their content together with an encoding of the full history. Training uses soft masks, annealing, and an alignment loss. At inference, the top-K candidates are retained and aggregated into K tokens through masked attention. The training objective also includes the KL divergence to the diagonal Gaussian predictive distribution of a frozen iTransformer teacher that observes the full history. The authors motivate predictive sufficiency through a log-loss risk identity involving the true conditional distributions, and introduce IPS as a metric defined by the teacher–student KL divergence. Across eight forecasting datasets, SPSNet applies SPS to its seasonal branch and combines it with a trend branch and channel mixing. The paper also evaluates SPS as a plug-in for three patch-based backbones. The appendices include results over five seeds, KL-regularized baselines, ablations, a longer-lookback evaluation, and efficiency analysis. The main L=96 configuration retains 10 of 80 candidates; the supplementary L=336 experiments retain 10 of 320 candidates.

Critical strengths
  1. Useful forecasting results with compact representations and evidence across backbones. SPS combines data-dependent candidate scales with explicit selection, estimating spectral information only from the training split. Table 1 reports favorable MAE on most datasets, including average Solar MSE/MAE of 0.195/0.221. Appendix Table 10 further evaluates L=336 while retaining 10 tokens. The SRSNet, PatchMLP, and PatchTST plug-in experiments indicate that the benefits extend beyond the complete SPSNet architecture, although the cases of degraded MSE should also be acknowledged.
  2. The authors address several important confounds. Appendix Table 7 (p. 23) applies the same KL regularization to the baselines, helping isolate the gains beyond distillation. The paper reports that SPS further improves both MSE and MAE by approximately 2.55% on average over these KL-regularized baselines. The appendices also provide means and standard deviations over five seeds, ablations removing KL, alignment, annealing, and selection, and a shared-teacher protocol. These controls help distinguish the overall gains from the effect of adding distillation alone.
Critical weaknesses
  1. Teacher agreement does not establish predictive sufficiency. Theorem 1 uses true P(Y|X) and P(Y|Z), whereas IPS (Eq. 19) uses a single trained iTransformer teacher and Gaussian outputs. Matching a teacher that omits predictive information need not make I(Y;X|Z) close to zero. Diagonal-variance estimates can also change KL without changing information retention. Appendix O acknowledges teacher limitations, and Table 7’s matched-KL controls are valuable, but neither establishes calibration, ranking stability across teachers, or a link between IPS and information loss. In Table 6, adding SPS to PatchTST on Solar lowers IPS from 2.4266 to 1.4319 while increasing MSE from 0.224 to 0.231. Lower IPS therefore need not indicate better forecasts or establish preservation of predictive information.
  2. Selection is not isolated from the other added components. Comparisons jointly introduce multiscale encoding, full-history-conditioned scoring, attention pooling, and KL regularization. The “without selection” ablation (Table 8) changes K from 10 to 80, altering downstream processing. Missing controls include random or uniform selection with identical candidates, K=10, encoder, pooling, and KL objective, and fixed-scale candidates with matched N and K. SPSNet’s trend branch also observes the full history, so low IPS for the complete model cannot establish that seasonal-branch patches alone preserve its information. At L=96, token count falls only from 11 to 10 while the front end encodes 80 candidates. Token compression is greater at L=336 but does not establish an end-to-end speedup. Matched selection controls and complete cost accounting remain necessary to establish the module’s independent value.
Questions for authors
  1. Metric clarification. Please report the full-history teacher’s forecasting accuracy, NLL, and calibration. Are IPS rankings stable across teachers and with fixed or calibrated predictive variances? Otherwise, please define IPS as a proxy for agreement with a shared teacher and qualify the sufficiency claims accordingly.
  2. Selection controls. How does learned selection compare with random or uniform selection under identical N, K, encoding capacity, full-history information for scoring, pooling, and KL training conditions? Are fixed-scale candidates substantially worse under the same budget? Please distinguish complete SPSNet results from matched ablations retaining the same trend branch.
  3. Implementation and costs. Appendix E says the Gaussian head reads detached forecasts. Is only the variance-branch input detached, or also the mean used in the KL term? What gradient path connects KL to the backbone and selector? Please separate teacher-training, student-training, and candidate-generation/encoding costs, and identify those included in the per-iteration efficiency figure (p. 25).
5 comments

We thank the reviewer for the detailed and constructive feedback.

W1 / Q1. Teacher agreement does not establish predictive sufficiency; metric clarification.

We adopt your suggestion. IPS is defined as a proxy for agreement with a shared teacher, and the sufficiency claims are qualified accordingly: Theorem 1 states the ideal objective, and Proposition 1 states what a KL to a reference can bound.

Proposition 1. For any predictive distribution qm(Y∣Z)q_m(Y\mid Z) with Z=S(X)Z=\mathcal{S}(X),

E[DKL(P(Y∣X) ∥ qm(Y∣Z))]≥I(Y;X∣Z),E[−log⁡qm(Y∣Z)]=H(Y∣X)+E[DKL(P(Y∣X) ∥ qm(Y∣Z))].\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big] \ge I(Y;X\mid Z), \qquad \mathbb{E}\big[-\log q_m(Y\mid Z)\big] = H(Y\mid X) + \mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big].

The gap in the inequality is E[DKL(P(Y∣Z) ∥ qm(Y∣Z))]\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid Z)\,\|\,q_m(Y\mid Z))\big], the approximation error of the predictor. IPS therefore bounds the information loss only if its reference is the true P(Y∣X)P(Y\mid X), and no such bound holds for a teacher. As you say, matching a teacher that omits predictive information need not make I(Y;X∣Z)I(Y;X\mid Z) small. The difference in held-out NLL between two methods equals the difference in their exact-reference IPS, which gives a teacher-free test.

[Experiment 1a] Teacher Accuracy and Calibration (Q1). Mean over four horizons, with each model's own variance head:

DatasetTeacher MSE / MAESPSNet MSE / MAETeacher NLL / CRPS / 90% coverageSPSNet NLL / CRPS / 90% coverage
ETTh10.464 / 0.4510.434 / 0.4280.827 / 0.304 / 0.9480.786 / 0.298 / 0.924
ETTh20.385 / 0.4050.369 / 0.3940.895 / 0.287 / 0.9350.736 / 0.288 / 0.901
ETTm10.415 / 0.4130.377 / 0.3870.879 / 0.277 / 0.9440.697 / 0.267 / 0.888
ETTm20.285 / 0.3260.277 / 0.3190.769 / 0.231 / 0.9240.686 / 0.240 / 0.889
Weather0.270 / 0.2890.253 / 0.2730.705 / 0.202 / 0.9690.684 / 0.187 / 0.935
Solar0.228 / 0.2410.193 / 0.2230.640 / 0.160 / 0.9830.631 / 0.161 / 0.970
Electricity0.183 / 0.2730.181 / 0.2700.572 / 0.188 / 0.9290.537 / 0.189 / 0.911
Traffic0.430 / 0.2830.468 / 0.2840.490 / 0.201 / 0.9480.410 / 0.210 / 0.935
Mean0.333 / 0.3350.319 / 0.3220.722 / 0.231 / 0.9480.646 / 0.230 / 0.919
  • Accuracy. The teacher is not an upper bound: SPSNet has lower MSE in 25/32 settings and lower MAE in 28/32. The teacher is more accurate on Traffic. Its average error is close to that of the iTransformer column of Table 1 (0.329 / 0.336), so it is a standard full-history model.
  • Calibration. The teacher is over-dispersed. Its nominal 90% intervals cover 92.4–98.3% of the test targets, 94.8% on average. SPSNet's intervals cover 91.9% on average and are closer to 90% on all eight datasets, with slight under-coverage on ETTm1 and ETTm2 (88.8% and 88.9%). SPSNet also has the lower NLL on all eight datasets (0.646 against 0.722), while CRPS is tied (0.230 against 0.231).
  • Caveat. The NLL gap reflects the variances more than the means: the teacher has the lower CRPS on five datasets, and on Traffic it has the lower MSE and still the higher NLL. We do not claim that SPSNet is the better probabilistic forecaster. The table shows that the teacher is not a close stand-in for P(Y∣X)P(Y\mid X).

[Experiment 1b] Variance-Robust and Teacher-Free IPS. We recompute IPS on the 60 plug-in settings (three backbones × five datasets × four horizons) under four variance treatments: the submission's heads, the same heads calibrated on validation, all heads refitted with one NLL recipe, and one shared variance, where only forecast disagreement counts.

IPS computed with+SPS below its backbone+SPS below +KL+KL below its backboneMean rank of SPSNet among 10 models
submission heads60/6054/6060/601.00
calibrated heads60/6049/6060/601.00
refitted heads60/6051/6059/601.00
shared variance60/6052/6060/601.00
  • Magnitudes. The variance term is 92–99% of every IPS value in the submission. This is the effect you describe: diagonal-variance estimates change the KL without changing what is retained. The size of an IPS value, and the percentage reductions in Tables 6 and 7, mostly reflect how each model's variance head was fitted. The revision reports orderings instead.
  • Orderings with fixed or calibrated variances. They are stable. +SPS is below its backbone in 60/60 settings under all four treatments and below +KL in 49–54/60. SPSNet has rank 1.00 among the 10 models (SPSNet, the three backbones, and their +KL and +SPS variants). With a shared variance, IPS reduces to the distance between the student's and the teacher's forecasts, and the conclusions are the same.
  • Your Table 6 example. PatchTST+SPS on Solar lowers IPS from 2.4266 to 1.4319 while MSE rises from 0.224 to 0.231. Since the variance term dominates both values, this drop is mostly a change in the variance term and says little about the forecast. We agree that a lower IPS does not indicate a better forecast, and the revision says so. It also states in the text the four dataset averages of Table 6 where +SPS increases MSE (SRSNet on ETTh1 and ETTm1, PatchTST on ETTh1 and Solar).
  • Teacher-free test. Held-out NLL needs no teacher, and by Proposition 1 it orders methods as exact-reference IPS would. In [Experiment 1a], SPSNet's NLL is below that of the full-history teacher on all eight datasets. In exact-reference terms the teacher is further from P(Y∣X)P(Y\mid X) than its student, which is one more reason not to read teacher-relative IPS as lost information.

[Experiment 1c] Sensitivity to the Teacher. We train SPSNet at H=96 with teachers stopped after 1, 3 and 10 epochs, with the full teacher, with no teacher, and with a teacher of SPSNet's own architecture without patch selection. IPS uses the submission's heads. Mean over six datasets, relative changes are averaged over datasets:

Training teacherTeacher MSESPSNet MSESPSNet MSE vs. full teacherIPS vs. its own teacherIPS vs. the full teacher
1 epoch0.3250.255+4.7%0.8360.776
3 epochs0.2890.252+3.2%0.4620.410
10 epochs0.2680.248+0.9%0.2980.252
full (submission)0.2590.247–0.2090.209
none (no KL)–0.252+2.9%–0.408
architecture-matched0.2490.247+0.1%0.2230.219
  • Teacher quality matters for accuracy. Removing the full teacher raises MSE by 2.9%, and a 1-epoch teacher is worse than none.
  • IPS depends on the teacher. The same student architecture has an IPS between 0.209 and 0.836 depending on its teacher. An IPS value describes a student–teacher pair and not a patching strategy alone.
  • A weak teacher is not copied. With the 1-epoch teacher (MSE 0.325), SPSNet reaches 0.255, and its IPS against that teacher is the highest in the table (0.836). The supervised loss keeps the student away from a poor teacher, so a low IPS does not arise from imitating one.
  • Stability across teachers. The five variants trained with a teacher are ordered identically against their own teachers and against the full teacher. Trained with a teacher of SPSNet's own architecture, SPSNet's MSE changes by 0.1%, and its IPS against the iTransformer teacher, which it never saw in training, is 0.219, against 0.209 when trained with that teacher and 0.408 with no teacher. This experiment varies the teacher for SPSNet. It does not re-rank the baselines against a second teacher.

Summary. IPS is a proxy for agreement with a shared teacher. The link to information loss holds only for the true reference (Proposition 1), and the teacher is not that reference: it is less accurate than SPSNet in most settings and over-dispersed. The magnitudes of IPS are driven by the variance heads, while the orderings are unchanged with calibrated, refitted and shared variances and with a teacher of a different architecture. The revision reports the teacher's accuracy, NLL and calibration, variance-robust IPS, and held-out NLL as the teacher-free comparison.

W2 / Q2. Selection is not isolated from the other added components, selection controls.

We agree with each point. In [Experiment 2a] and [Experiment 2b], all arms share N=80, K=10, the encoder, pooling, KL and training recipe, and they all keep the same trend branch. The SPSNet experiments use 29 settings (seven datasets × four horizons, and Traffic at H=96). Relative changes are averaged over settings and are positive when MSE is higher than in the reference row.

[Experiment 2a] Candidate Generator Controls. Only the period list changes. MSE and MAE are means of the eight dataset averages:

Candidate familyDistinct lengthsMSEMAEMSE vs. spectralSpectral better inTraining time vs. spectral
Spectral (submission)4–100.3170.320––1.00
Fixed multi-scale {8, 16, 32, 64}40.3220.323+1.6%22/290.89–1.07
Random lengths (3 draws)6–120.3240.326+2.2%24/290.92–1.20
Uniformly sampled grid510.3250.326+2.5%26/291.74–2.19
Fixed single-scale, P=1610.3270.329+3.2%27/290.81–0.90
  • Fixed-scale candidates at matched N and K. A single fixed scale is 3.2% worse than the spectral list (27/29), and four fixed scales are 1.6% worse (22/29). Fixed-scale candidates are consistently worse, but not substantially: the differences are 0.005–0.010 in MSE, and the revision describes them in these terms.
  • Other multi-scale families. Random lengths are 2.2% worse and a uniformly sampled grid 2.5% worse. The grid has 51 distinct lengths and takes 1.74–2.19× the training time, so the gain comes from which scales are used, not from how many.

[Experiment 2b] Selector Controls. Only the selection rule changes, on identical spectral candidates. "All pooled" attends over all 80 candidates into the same K output slots:

SelectorMSEMAEMSE vs. learnedLearned better in
Learned (submission)0.3170.320––
Learned, zeroed global context0.3180.320+0.3%18/29
All pooled0.3220.323+1.6%23/29
Recency (K latest-ending candidates)0.3230.323+1.9%23/29
Uniform (every N/K-th candidate)0.3240.324+2.2%25/29
Random K per series0.3250.325+2.5%27/29
  • Random and uniform selection. With identical candidates, K=10, encoder, pooling and KL, learned selection is better than random (+2.5%, 27/29), uniform (+2.2%, 25/29) and recency selection (+1.9%, 23/29).
  • Full-history information for scoring. Zeroing the global context of the scorer changes MSE by 0.3% and leaves MAE unchanged. The learned selector without it (0.318) is still better than random or uniform selection (0.325, 0.324), so the gain does not come from the full-history input of the scorer.
  • Selection at an unchanged K. As you point out, the "w/o Candidate Selection" row of Table 8 sets K=N=80 and changes the downstream processing. "All pooled" removes the selection and keeps the K=10 output slots. It is 1.6% worse than learned selection (23/29), so discarding candidates improves accuracy and does not only compress.

[Experiment 2c] Gain Attribution (Q2). Starting from each backbone's original script, we add one component at a time, and the last two columns compare each row with the row above it. MSE / MAE, mean over 20 settings per backbone (ETTh1, ETTh2, ETTm1, ETTm2 and Weather × four horizons):

StepSRSNetPatchMLPPatchTSTMSE change of the step, pooledBetter in
Original script0.350 / 0.3680.363 / 0.3780.347 / 0.368
+ KL to the shared teacher0.345 / 0.3640.347 / 0.3660.341 / 0.360−2.5%44/60
+ SPS candidates, all pooled0.344 / 0.3590.338 / 0.3570.338 / 0.358−1.3%48/60
+ random selection of K=100.347 / 0.3610.342 / 0.3580.341 / 0.359+1.0%17/60
+ learned selection (= +SPS)0.343 / 0.3570.334 / 0.3540.337 / 0.356−1.6%52/60
+SPS vs +KL−1.8%52/60
  • KL. Adding the KL to the shared teacher lowers MSE by 2.5% (44/60). It is the largest single step, and it is not specific to SPS.
  • Candidates. Replacing the backbone's patching with the spectral candidates, all pooled into K=10 slots, lowers MSE by a further 1.3% (48/60).
  • Selection. Random selection of K=10 gives this gain back (+1.0%, and it is better than all pooled in only 17/60 settings). Learned selection is 1.6% better than random selection (52/60), which is about 0.6% better than all pooled.
  • Answer to Q2. +SPS is about 1.8% below +KL in MSE (0.338 against 0.344 on average) and better in 52/60 settings. About two thirds of this gain comes from the candidate set (−1.3%) and one third from learned selection (about −0.6% against pooling all candidates).

The revision defines IPS as teacher-relative, so we attribute the gain in MSE and MAE and not in IPS.

Reviewer_TibVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper studies patch selection for long-term time series forecasting. It proposes SPS, which generates spectrally guided multiscale intervals and learns a compact selection using soft-to-hard masking and predictive-distribution matching to a full-history teacher. Experiments on eight benchmarks, several plug-in backbones, two lookback lengths, and ablation and efficiency studies report competitive forecasting accuracy and lower teacher-relative IPS with fewer downstream tokens.

Critical strengths
  • Separating candidate generation from candidate retention is a useful way to reason about forecasting representations. SPS realizes that strategy through spectrally guided interval generation, length-aware encoding, budgeted selection, and teacher-distribution matching. Theorem 1 supplies an ideal predictive-information objective, while the frozen teacher provides an implementable surrogate.

  • The eight-dataset study and plug-in backbone comparisons demonstrate practical potential beyond SPSNet.

Critical weaknesses
  • Theorem 1 concerns true conditional distributions and Bayes predictors, but IPS measures agreement between learned diagonal Gaussians and one iTransformer teacher (Sections 3.6-3.7 and Appendix E). The same teacher is used for training and evaluation, so lower IPS can follow from successful distillation even if the teacher is biased or poorly calibrated. Table 7's KL-matched controls are helpful, but they do not validate IPS as a measure of actual retained predictive information. Teacher accuracy, calibration, and sensitivity to a different evaluation teacher are essential here.

  • The contributions of spectral generation and learned selection are not independently isolated. Appendix J removes losses, annealing, or selection, but does not replace the spectral period list with an equally sized fixed multiscale or random-length candidate family. It also lacks equal-budget random or simple recency selection on the same candidates.

Questions for authors
  • What are the teacher's forecasting error, NLL, and predictive calibration relative to the students?
  • Which component accounts for the gain beyond KL-only training?
6 comments

We thank the reviewer for the thoughtful and constructive feedback.

W1 / Q1. Theorem 1 concerns the true conditionals, while IPS measures agreement with one teacher that is also used for training; the teacher's error, NLL and calibration.

We agree, and [Experiment 1] confirms your concern. We now report IPS as agreement with a stated teacher and not as evidence of sufficiency.

Proposition 1. For any predictive distribution qm(Y∣Z)q_m(Y\mid Z) with Z=S(X)Z=\mathcal{S}(X),

E[DKL(P(Y∣X) ∥ qm(Y∣Z))]≥I(Y;X∣Z),E[−log⁡qm(Y∣Z)]=H(Y∣X)+E[DKL(P(Y∣X) ∥ qm(Y∣Z))].\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big] \ge I(Y;X\mid Z), \qquad \mathbb{E}\big[-\log q_m(Y\mid Z)\big] = H(Y\mid X) + \mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid X)\,\|\,q_m(Y\mid Z))\big].

The gap in the inequality is E[DKL(P(Y∣Z) ∥ qm(Y∣Z))]\mathbb{E}\big[D_{\mathrm{KL}}(P(Y\mid Z)\,\|\,q_m(Y\mid Z))\big], the approximation error of the predictor. IPS therefore bounds the information loss only if its reference is the true P(Y∣X)P(Y\mid X), and no such bound holds for a teacher. The difference in held-out NLL between two methods equals the difference in their exact-reference IPS, which gives a teacher-free test.

[Experiment 1a] Teacher Accuracy and Calibration (Q1). Mean over four horizons, with each model's own variance head:

DatasetTeacher MSE / MAESPSNet MSE / MAETeacher NLL / CRPS / 90% coverageSPSNet NLL / CRPS / 90% coverage
ETTh10.464 / 0.4510.434 / 0.4280.827 / 0.304 / 0.9480.786 / 0.298 / 0.924
ETTh20.385 / 0.4050.369 / 0.3940.895 / 0.287 / 0.9350.736 / 0.288 / 0.901
ETTm10.415 / 0.4130.377 / 0.3870.879 / 0.277 / 0.9440.697 / 0.267 / 0.888
ETTm20.285 / 0.3260.277 / 0.3190.769 / 0.231 / 0.9240.686 / 0.240 / 0.889
Weather0.270 / 0.2890.253 / 0.2730.705 / 0.202 / 0.9690.684 / 0.187 / 0.935
Solar0.228 / 0.2410.193 / 0.2230.640 / 0.160 / 0.9830.631 / 0.161 / 0.970
Electricity0.183 / 0.2730.181 / 0.2700.572 / 0.188 / 0.9290.537 / 0.189 / 0.911
Traffic0.430 / 0.2830.468 / 0.2840.490 / 0.201 / 0.9480.410 / 0.210 / 0.935
Mean0.333 / 0.3350.319 / 0.3220.722 / 0.231 / 0.9480.646 / 0.230 / 0.919
  • Accuracy. The teacher is not an upper bound: SPSNet has lower MSE in 25/32 settings and lower MAE in 28/32. The teacher is more accurate on Traffic. Its average error is close to that of the iTransformer column of Table 1 (0.329 / 0.336), so it is a standard full-history model.
  • Calibration. The teacher is over-dispersed. Its nominal 90% intervals cover 92.4–98.3% of the test targets, 94.8% on average. SPSNet's intervals cover 91.9% on average and are closer to 90% on all eight datasets, with slight under-coverage on ETTm1 and ETTm2 (88.8% and 88.9%). SPSNet also has the lower NLL on all eight datasets (0.646 against 0.722), while CRPS is tied (0.230 against 0.231).
  • Caveat. The NLL gap reflects the variances more than the means: the teacher has the lower CRPS on five datasets, and on Traffic it has the lower MSE and still the higher NLL. We do not claim that SPSNet is the better probabilistic forecaster. The table shows that the teacher is not a close stand-in for P(Y∣X)P(Y\mid X), which is your point.

[Experiment 1b] Variance-Robust and Teacher-Free IPS. We recompute IPS on the 60 plug-in settings (three backbones × five datasets × four horizons) under four variance treatments: the submission's heads, the same heads calibrated on validation, all heads refitted with one NLL recipe, and one shared variance, where only forecast disagreement counts.

IPS computed with+SPS below its backbone+SPS below +KL+KL below its backboneMean rank of SPSNet among 10 models
submission heads60/6054/6060/601.00
calibrated heads60/6049/6060/601.00
refitted heads60/6051/6059/601.00
shared variance60/6052/6060/601.00
  • Magnitudes. The variance term is 92–99% of every IPS value in the submission. The size of an IPS value, and the percentage reductions in Tables 6 and 7, therefore mostly reflect how each model's variance head was fitted. The revision reports orderings instead.
  • Distillation alone lowers IPS. +KL is below its backbone in 59–60 of 60 settings. As you note, a lower IPS can follow from distillation, so the comparisons that isolate SPS are those at matched KL.
  • Orderings. They do not depend on the variance treatment. +SPS is below its backbone in 60/60 settings under all four treatments and below +KL in 49–54/60. SPSNet has rank 1.00 among the 10 models (SPSNet, the three backbones, and their +KL and +SPS variants). With a shared variance, IPS reduces to the distance between the student's and the teacher's forecasts, and the conclusions are the same.
  • Teacher-free test. Held-out NLL needs no teacher, and by Proposition 1 it orders methods as exact-reference IPS would. In [Experiment 1a], SPSNet's NLL is below that of the full-history teacher on all eight datasets. In exact-reference terms the teacher is further from P(Y∣X)P(Y\mid X) than its student, which is one more reason not to read teacher-relative IPS as lost information.

[Experiment 1c] Sensitivity to the Teacher. We train SPSNet at H=96 with teachers stopped after 1, 3 and 10 epochs, with the full teacher, with no teacher, and with a teacher of SPSNet's own architecture without patch selection. IPS uses the submission's heads. Mean over six datasets, relative changes are averaged over datasets:

Training teacherTeacher MSESPSNet MSESPSNet MSE vs. full teacherIPS vs. its own teacherIPS vs. the full teacher
1 epoch0.3250.255+4.7%0.8360.776
3 epochs0.2890.252+3.2%0.4620.410
10 epochs0.2680.248+0.9%0.2980.252
full (submission)0.2590.247–0.2090.209
none (no KL)–0.252+2.9%–0.408
architecture-matched0.2490.247+0.1%0.2230.219
  • Teacher quality matters for accuracy. Removing the full teacher raises MSE by 2.9%, and a 1-epoch teacher is worse than none.
  • IPS depends on the teacher. The same student architecture has an IPS between 0.209 and 0.836 depending on its teacher. An IPS value describes a student–teacher pair and not a patching strategy alone.
  • A weak teacher is not copied. With the 1-epoch teacher (MSE 0.325), SPSNet reaches 0.255, and its IPS against that teacher is the highest in the table (0.836). The supervised loss keeps the student away from a poor teacher, so a low IPS does not arise from imitating one.
  • A different teacher architecture. Trained with a teacher of SPSNet's own architecture, SPSNet's MSE changes by 0.1%. Its IPS against the iTransformer teacher, which it never saw in training, is 0.219, against 0.209 when trained with that teacher and 0.408 with no teacher. The low IPS of SPSNet is therefore not tied to using one teacher for both training and evaluation. This experiment varies the teacher for SPSNet. It does not re-rank the baselines against a second teacher.

Summary. IPS measures agreement with a stated teacher and is not evidence of sufficiency. The teacher is as accurate as a standard iTransformer but over-dispersed, and it is not an upper bound on SPSNet. The magnitudes of IPS are driven by the variance heads, while the orderings are unchanged under four variance treatments and under a teacher of a different architecture. The revision defines IPS accordingly, adds Proposition 1 and [Experiment 1], and reports held-out NLL as the teacher-free comparison.

W2 / Q2. The contributions of spectral generation and learned selection are not independently isolated, which component accounts for the gain beyond KL-only training.

We ran both controls and a component-by-component attribution. In [Experiment 2a] and [Experiment 2b], all arms share N=80, K=10, the encoder, pooling, KL and training recipe. The SPSNet experiments use 29 settings (seven datasets × four horizons, and Traffic at H=96). Relative changes are averaged over settings and are positive when MSE is higher than in the reference row.

[Experiment 2a] Candidate Generator Controls. Only the period list changes. MSE and MAE are means of the eight dataset averages:

Candidate familyDistinct lengthsMSEMAEMSE vs. spectralSpectral better inTraining time vs. spectral
Spectral (submission)4–100.3170.320––1.00
Fixed multi-scale {8, 16, 32, 64}40.3220.323+1.6%22/290.89–1.07
Random lengths (3 draws)6–120.3240.326+2.2%24/290.92–1.20
Uniformly sampled grid510.3250.326+2.5%26/291.74–2.19
Fixed single-scale, P=1610.3270.329+3.2%27/290.81–0.90
  • Equally sized multi-scale families. With the candidate and selection budgets fixed, the spectral period list has the lowest MSE and MAE. A fixed multi-scale list is 1.6% worse, random lengths 2.2% worse, and a uniformly sampled grid 2.5% worse. The spectral list is better in 22, 24 and 26 of the 29 settings.
  • Multi-scale against single-scale. The single scale P=16 is the weakest family (+3.2%, 27/29). A fixed multi-scale list closes half of that gap, and the spectral choice of scales closes the other half.
  • More scales are not better. The uniformly sampled grid has 51 distinct lengths, takes 1.74–2.19× the training time, and is worse than four fixed scales. The gain comes from which scales are used, not from how many.
  • Size of the effect. The differences are 0.005–0.010 in MSE. They are consistent across settings but modest.

[Experiment 2b] Selector Controls. Only the selection rule changes, on identical spectral candidates. "All pooled" attends over all 80 candidates into the same K output slots:

SelectorMSEMAEMSE vs. learnedLearned better in
Learned (submission)0.3170.320––
Learned, zeroed global context0.3180.320+0.3%18/29
All pooled0.3220.323+1.6%23/29
Recency (K latest-ending candidates)0.3230.323+1.9%23/29
Uniform (every N/K-th candidate)0.3240.324+2.2%25/29
Random K per series0.3250.325+2.5%27/29
  • Learned against simple rules. On the same candidates and at the same budget, learned selection is better than random (+2.5%, 27/29), uniform (+2.2%, 25/29) and recency selection (+1.9%, 23/29).
  • Selection against no selection. Pooling all 80 candidates into the same K slots is 1.6% worse (23/29). Discarding candidates therefore improves accuracy and does not only compress.
  • Global context. Zeroing the full-history context of the scorer changes MSE by 0.3% and leaves MAE unchanged. The selector relies mainly on the content of each candidate.

[Experiment 2c] Gain Attribution (Q2). Starting from each backbone's original script, we add one component at a time, and the last two columns compare each row with the row above it. MSE / MAE, mean over 20 settings per backbone (ETTh1, ETTh2, ETTm1, ETTm2 and Weather × four horizons):

StepSRSNetPatchMLPPatchTSTMSE change of the step, pooledBetter in
Original script0.350 / 0.3680.363 / 0.3780.347 / 0.368
+ KL to the shared teacher0.345 / 0.3640.347 / 0.3660.341 / 0.360−2.5%44/60
+ SPS candidates, all pooled0.344 / 0.3590.338 / 0.3570.338 / 0.358−1.3%48/60
+ random selection of K=100.347 / 0.3610.342 / 0.3580.341 / 0.359+1.0%17/60
+ learned selection (= +SPS)0.343 / 0.3570.334 / 0.3540.337 / 0.356−1.6%52/60
+SPS vs +KL−1.8%52/60

Comments on the paper

2 comments

We thank all Reviewers for their careful and constructive reviews.

To make cross-referencing easy, we use a single global numbering for the new experiments across all responses:

1. What IPS measures

[Experiment 1a] Teacher Accuracy and Calibration (TibV-W1/Q1, zvKJ-W1/Q1, xrTB-W2): the iTransformer teacher averages MSE 0.333 / MAE 0.335 against 0.319 / 0.322 for SPSNet, which is more accurate than its teacher in 25/32 settings. The teacher is over-dispersed: its 90% intervals cover 94.8% of the test targets (SPSNet: 91.9%), and its NLL is higher on all eight datasets (0.722 against 0.646). It is a reference, not an upper bound.

[Experiment 1b] Variance-Robust and Teacher-Free IPS (TibV-W1, zvKJ-W1/Q1, xrTB-W2): the variance term is 92–99% of every IPS value in the submission. The magnitudes of IPS therefore reflect the variance heads, but the orderings do not: with calibrated, refitted or shared variances, +SPS stays below its backbone in 60/60 settings and below +KL in 49–52/60, and SPSNet stays first among the 10 models.

[Experiment 1c] Sensitivity to the Teacher (TibV-W1, zvKJ-W1/Q1, xrTB-W2): SPSNet's IPS ranges from 0.209 to 0.836 with the quality of its teacher, and a 1-epoch teacher is worse for accuracy than no teacher. With a teacher of a different architecture, MSE changes by 0.1%, and the IPS against the iTransformer teacher, which is then unseen in training, is 0.219 against 0.209.

2. Which component gives the gain

[Experiment 2a] Candidate Generator Controls (TibV-W2, zvKJ-W2/Q2, xrTB-W1): on SPSNet at N=80, K=10, replacing the spectral period list raises MSE by 1.6% (fixed multi-scale), 2.2% (random lengths), 2.5% (uniformly sampled grid) and 3.2% (fixed single-scale). The spectral list is better in 22–27 of 29 settings.

[Experiment 2b] Selector Controls (TibV-W2/Q2, zvKJ-W2/Q2, xrTB-W3): on identical candidates in SPSNet, learned selection is better than random (+2.5% MSE, 27/29), uniform (+2.2%, 25/29) and recency selection (+1.9%, 23/29), and better than pooling all 80 candidates into the same K slots (+1.6%, 23/29). Zeroing the global context changes MSE by 0.3%.

[Experiment 2c] Gain Attribution (TibV-Q2, zvKJ-W2): on three backbones, the training recipe changes MSE by −2.5% (KL to the shared teacher), the SPS candidates by a further −1.3%, and learned selection by −1.6% against random selection of the same K=10. +SPS is better than +KL in 52/60 settings, and about two thirds of the gain comes from the candidates and one third from learned selection.

3. Costs

[Experiment 3] Cost Accounting (zvKJ-W2/Q3): candidate generation is a one-time FFT of under 7 ms per dataset, and teacher training is a one-time cost of 19–54% of the teacher + SPSNet total. In a training iteration, the frozen-teacher forward takes 9–18% of the time and the SPS front-end 19–21%. As a plug-in at L=96, SPS is 1.30–1.93× slower at inference and 1.58–1.75× slower in training than its backbone, so we do not claim an end-to-end speedup.

4. What the selected representation contains

[Experiment 4] Information Paths (zvKJ-W2/Q2, xrTB-W3): the trend branch alone is within 2.3% of the complete SPSNet in MSE (IPS 0.313 against 0.210), so complete-model results do not isolate the selected patches. With the trend branch kept, K=10 selected patches match a seasonal branch that reads the whole seasonal series (MSE 0.304 against 0.306). Without the decomposition, selecting 10 candidates equals pooling all 80 (MSE 0.338 for both), and both are behind the whole series (0.322).

All new results will be integrated into the revised paper. We remain available for further experiments and clarifications during the discussion period.

Area_Chair_ogyMOpenReview ↗

Discussion reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.