acceptodds
Withdrawn from ICLR 2027

Correction-space Cross-variate Interaction for Test-time Adaptation in Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time adaptation (TTA) is a promising paradigm for handling distribution shift in time-series forecasting (TSF), where models adapt at inference time, often leveraging delayed observed data to refine predictions. In the multivariate setting, distribution shifts often exhibit cross-variate dependencies, yet existing TSF-TTA methods adapt each variate independently and ignore this cross-variate structure. Exploiting such structure motivates cross-variate interaction, but coupling variates through backbone predictions introduces direct pathways for mixing uncorrected errors across variates, a concern under the delayed supervision of TSF-TTA. We identify the interaction space as a key design choice, and show that acting on adapter corrections that refine backbone outputs, the correction space, rather than on the predictions themselves, avoids directly propagating backbone errors across variates. We build on this to propose CoRe (Correction-space Interaction Refinement), realizing correction-space interaction through (i) Shared-anchor Correction Refinement (SCR), which combines each variate's correction with a shared anchor through a parameter-efficient bottleneck, and (ii) input-conditioned spectral gating, which adaptively modulates the refinement from the current input window. Across seven backbones, six datasets, and four prediction horizons, CoRe reduces MSE by 25.82% on average over backbones and 10.57% over the state-of-the-art TSF-TTA method, with stronger gains at medium-to-long horizons and modest computational overhead. Data and code are available at: anonymous.4open.science/r/CoRe-TTA-CD98.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, test-time adaptation, distribution shift, multivariate time series, online learning

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 4 reviews, above 27% of the others.
Reviewer_GcBaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-variate interaction for test-time adaptation in multivariate time-series forecasting under delayed supervision. CoRe models interactions among corrections produced by an existing output adapter, rather than directly mixing predictions from the frozen backbone. It combines a cross-variate mean anchor with shared-bottleneck refinement, modulated by a spectral gate computed from the input history. The evaluation compares correction-space and prediction-space interaction across forecasting backbones, datasets, forecast horizons, and additional adaptation methods.

Critical strengths
  1. A clear design choice with strong comparative evidence. CoRe addresses a limitation of per-variate adaptation and compares its correction-space approach with COSA, its closest predecessor. Across the 168 configurations in Table 1, the reported mean per-setting relative MSE improvement is 10.57%, with 162/168 favorable comparisons. Both figures agree with the tabulated values and support the comparison.

  2. Broad evaluation and a useful parameter-efficiency trade-off. The experiments cover multiple backbone families, additional adapters, multiple random seeds, interaction-space controls, and ablations of the anchor and gate. The detailed comparison with dense refinement shows similar aggregate accuracy, while the shared bottleneck requires substantially fewer additional refinement parameters. The resulting parameter-efficiency advantage supports the practical value of the architecture.

Critical weaknesses
  1. Mechanism attribution and supporting result summaries need clarification. In Table 4, refinement without the shared anchor already improves performance by 7.1%, compared with 9.7% for the full method. The COSA+MLP control in Table 2 yields 0.97%, but the differences between this control and the no-anchor variant remain unclear. In addition, the aggregate improvements in Table 14 are not readily recoverable from the 168 detailed entries in Table 15. Averaging per-setting relative improvements over frozen backbones gives approximately 26.85% and 26.72% for the default and dense variants, respectively. These estimates differ from the reported improvements of 25.82% and 23.09%. These discrepancies complicate the interpretation and reproduction of the supporting comparisons.

  2. The update timeline needs to be reconciled with the delayed-label protocol. The paper states that forecast targets become eligible for adaptation only after their labels have matured. In the supplied snapshot, the sequential path appears to update on the current batch's complete forecast targets before proceeding to adjacent windows. This could expose the adapter to information unavailable at the next forecast time. A small source-level test confirms this dependency in the supplied path, but does not reproduce the reported benchmark runs. The supplied COSA path also follows a similar update pattern. The authors therefore need to identify the code version and update timeline used for the reported results.

Questions for authors
  1. Could the authors specify how the COSA+MLP, no-anchor, and prediction-space controls differ in input features, parameter sharing, gating, and optimization? Please also define the aggregation used in Tables 14–15 and identify the underlying runs. These details would clarify both the parameter-efficiency comparison and the contribution of the interaction space.

  2. Could the authors provide a short time-indexed example or execution trace showing forecast issuance, label availability, and updates permitted before each subsequent forecast? Was the supplied adapt_simple() path used to produce Table 1? If so, please provide the main CoRe–COSA comparison with label maturation explicitly enforced.

1 comment

Thank you for the source-level test and for recomputing our appendix tables; the trace you asked for is what exposed the leak. The code we adopted does not follow our stated protocol, and we did not verify it. Details and corrected results are in the General Response.

Q2. Yes, the supplied adapt_simple path (or a loop of the same structure) produced Table 1, and your trace is right.

Q1. Tables 2, 14 and 15 come from the leaky runs; your recomputation of Tables 14 and 15 is correct, and the mismatch is an error from compiling results.

Reviewer_MY3XOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies output-space test-time adaptation of a frozen multivariate forecaster, where an adapter learns per-variate corrections as ground truth arrives with delay. It argues that cross-variate interaction should act on these corrections, not on the backbone predictions. The module, SCR, sits on an existing adapter (COSA by default). It averages the per-variate corrections into a mean anchor, passes each correction together with the anchor through a shared low-rank bottleneck, and adds the result back through a spectral gate. Proposition 1 states that the correction-space operator does not take backbone error as an argument. Experiments cover 7 backbones, 6 datasets and 4 horizons over 10 seeds, with a reported 10.57% average MSE reduction over COSA.

Critical strengths

S1. A clearly posed design question, answered with a cheap and portable module. Choosing between interaction on predictions and interaction on corrections is a real design decision, and the paper frames it cleanly. The intuition for the correction space is sensible, and to my knowledge this is the first test-time adaptation method for forecasting that places cross-variate interaction there. SCR adds about 9.9k parameters and raises per-step time only from 6.28 to 6.76 ms. On top of TAFAS, run under TAFAS's own delayed queue, it gives an average gain of +4.49% on DLinear (Table 7). This is the most protocol-transparent evidence in the paper, though it also shows a few long-horizon regressions.

S2. Broad evaluation grid with seeds, and released code that makes the pipeline checkable. Table 1 covers 7 channel-independent and channel-dependent backbones on ETT, Exchange and Weather at four horizons over 10 seeds. The appendices add further datasets, backbones and published TSF-TTA baselines such as DynaTTA. The anonymous repository made the protocol question in W1 answerable from the code. This coverage meets or exceeds what is usual in this subarea.

Critical weaknesses

W1. The released code updates the adapter on targets that, under the stated protocol, have not yet been revealed. The headline gains are therefore not measured under the leakage-free setting the paper describes. The paper states that targets arrive with an H-step delay and that the adapter uses only revealed targets (Sec. 3.1, Alg. 1, App. B.2). In tta/core.py (adapt_simple), each batch is forecast and scored, and then the adapter takes up to 20 gradient steps on that batch's full H-step ground truth before the next batch, only 25 windows later, is forecast. At H = 720, the next batch is therefore predicted by an adapter just fit on targets covering about 695 of its 720 steps. The TAFAS path in the same repository (tta/tafas.py) instead uses a horizon-keyed delayed queue. The numbers fit this reading: DLinear on Exchange-720 drops from 0.8383 to 0.1239 with CoRe, while SCR inside TAFAS's delayed queue moves only from 0.8281 to 0.8217 (Table 7). Most of the drop already comes from COSA, so the problem lies in the shared protocol rather than in SCR. It still invalidates the headline comparisons against TAFAS and the frozen backbone, and these results cannot be relied on until they are re-run under a matured-target queue.

W2. The ablations attribute only a minority of the gain to cross-variate interaction, which is the claimed contribution. In Table 4, COSA averages 0.3050 and CoRe 0.2754 (-9.7%). A per-variate bottleneck with no shared anchor already reaches 0.2833 (-7.1%), so about three quarters of the gain comes without cross-variate mixing. The parameter-matched COSA+MLP control in Table 2 is single-seed and unspecified, so it does not settle this. The motivation also fits the mechanism loosely, since Fig. 2 shows pairwise dependencies while SCR uses only an unweighted mean of corrections. What remains novel beyond a per-variate refinement head is a modest increment.

W3. Proposition 1 does not support the accuracy claims it is used for, and backbone error still crosses variates through the anchor. Proposition 1 correctly states that backbone error is not an explicit argument of the correction-space operator, and the paper concedes that it implies no accuracy ordering (l.202-204). The main text still uses it to support the empirical advantage (l.781, l.891). Because the base correction depends on the backbone prediction, the mean anchor carries every variate's backbone-dependent correction into every variate's refinement. The separation is one step removed, not absent. The placement argument therefore rests entirely on the empirical comparison, which W1 undermines.

Minor points, listed separately from the weaknesses above and not weighed in the assessment:

  • Table 2 uses a single seed (42), while the main tables use 10.
  • The sentence at l.390 ("reported separately due and 10 seeds") is garbled.
  • The main setup should state that Exchange Rate reuses the ETTm2 configuration.
  • The reference to a "leakage-free objective in Section 3.6" (l.784) should point to a concrete definition of the batch B_t.
  • At H = 96, CoRe is worse than the frozen backbone in a few cells, such as ETTm1 DLinear (0.3715 to 0.3881), which deserves a brief discussion.

Further requests, not decisive for the assessment:

  • Please repeat the Fig. 8 singular-value analysis on column-centered corrections, since a rank-one mode is already expected from per-variate offsets.
  • Please measure the cross-variate path of backbone error through the mean anchor for CoRe versus the prediction-space variant.
  • Raw PETSA numbers in Table 1 and a comparison with delta-Adapter (ICLR 2026) or LIFT (ICLR 2024) would be useful additions.
  • A brief positioning against delayed-feedback online forecasters such as PROCEED and OneNet would help readers.
  • On Electricity, Traffic and TTFBench, please compare beyond COSA and the no-SCR variant.
  • Please explain why the Table 15 dense-variant iTransformer column matches the DLinear column in 9 of 24 cells.
  • Please explain the jump in Table 16 from +20.24% at r = 32 to +25.45% at r = C, given that C is 7 or 8 on five of six datasets.
  • Please describe how r, b_g and the adaptation steps would be chosen without test-stream results, ideally on a held-out validation stream.
Questions for authors

Q1 (W1). Please confirm whether adapt_simple in tta/core.py produced Table 1, and report for each adaptation step when each target it uses matures and when it is consumed. Please re-run Table 1 for a representative subset of backbones, with Fig. 5/6 and Table 4, under one matured-target queue applied identically to all methods. Without this re-run, the headline gains cannot be taken as evidence for the method. If the gains over TAFAS and the frozen backbone largely survive, W1 is resolved; if they shrink toward the Table 7 level, the headline claims must be restated.

Q2 (W2). Please specify the COSA+MLP architecture in Table 2, run it over 10 seeds next to the no-anchor SCR variant, and clarify what the "full CxC mixing" row in Table 4 implements. A clear multi-seed gain from the anchor beyond a capacity-matched per-variate head would resolve W2. If the anchor still adds only about 2.6 of the 9.7 points, this confirms that most of the improvement comes from per-variate capacity.

1 comment

Thank you for going through our code thoroughly and helping us find the problem. Your diagnosis in W1 is correct. The code we adopted does not follow our stated protocol, and we did not verify it. Details and corrected results are in the General Response.

Q1. Yes, adapt_simple (or a loop of the same structure) produced Table 1. We re-ran all methods as you suggested, and the gains do not survive.

Q2. After the fix, the shared anchor is not effective.

We also agree with W3. Of your further requests, the closest test we ran found that an error injected into one variate spreads further through corrections than through normalised predictions. The Table 15 and 16 issues are errors from compiling results.

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

The paper studies test-time adaptation for multivariate time-series forecasting (TSF-TTA), where ground truth arrives only after the horizon. Existing methods adapt each variate independently. The paper argues that cross-variate interaction should act on the adapter's corrections rather than on backbone predictions. The stated reason is that backbone errors are then not mixed across variates before delayed feedback arrives. The method, CoRe, sits on top of a base adapter and it averages the per-variate corrections into a shared anchor and concatenates it to each variate's correction, then a shared two-layer bottleneck with rank r=C maps this to a refinement and finally a gate computed from four spectral statistics of the input window (entropy and three band ratios) scales the refinement. CoRe reports 25.82% lower MSE than the frozen backbones and 10.57% lower than COSA. Further results cover the prediction-space comparison and other TSA-TTA baselines with overall good improvements.

Critical strengths
  1. CoRe improves on COSA across 168 settings with 10 seeds (Table 1, Table 6). The gains hold on a modern backbone as TimeXer. They also hold when SCR is added to other adapters (TAFAS +4.49%, standalone MLP +3.70% with 156 of 168 wins, Tables 7 and 8). The overhead is small (about 10k parameters, +7.7% per step). This is important because it shows the refinement is not tied to one adapter or backbone, which makes a good practical contribution.

  2. The work provide some good insights and interesting experimental design, for instance, Figures 5 and 6 hold the refinement architecture fixed and change only the interaction space. They show a horizon-dependent gap in favor of correction space on different backbones. The result is a bit empirical but the work is useful for the community.

Critical weaknesses
  1. The gain is credited to correction space and cross-variate interaction, but the evidence does not isolate either. This matters because the title and framing depend on this attribution.
  • Proposition 1 only shows that a function of Y+e depends on e. The paper itself says Δ depends on the backbone prediction, and Δ is trained to cancel the backbone error. So mixing Δ across variates also mixes backbone error, just indirectly.
  • Fig. 5 changes more than the "space", becaue the prediction-space input has large, non-stationary levels, and Δ is a small residual and the advantage may come from input scale and centering, not due to the error mixing. Therefore, it requires a centered or normalized prediction-space variant to test this.
  • Most of the gain does not come from cross-variate interaction. In Table 4, a per-variate bottleneck with no anchor already gives −7.1% vs COSA. Full CoRe gives −9.7%, so the shared anchor adds about 2.6 points. The spectral gate adds about 0.5 points over a fixed gate (−9.2%).
  1. The experimental protocol favors CoRe, and some results look inconsistent.
  • Table 16 looks odd. On Exchange Rate (C=8), r=8 and r=C should be the same setting. Across the table, r=C beats r=8 by about 6.6 points (25.45% vs 18.84%), and r=C is below 16 for most datasets. Rank alone cannot explain this.
  • In Table 15, several dense-variant entries look identical across different backbones (for example ETTh1 H=96: 0.5133 for iTransformer and DLinear). This may be a table error.
  • The headline averages are dominated by Exchange Rate and long horizons, where the frozen backbones are very poor (Exchange H=720: 0.91 to 0.19), and COSA already accounts for much of that gain. At H=96, CoRe often does not beat TAFAS or even the frozen backbone (for example ETTh1 with DLinear: 0.4695 base vs 0.4916 CoRe). On Traffic the gains are marginal (0.6352 to 0.6325).
  • The reliability analysis uses only 3 backbones. Seed variance is much larger for CoRe than for COSA in several Table 12 rows (for example ±0.0049 vs ±0.0001), which limits the average gains and the sensitivity conclusions.
Questions for authors
  1. Does the advantage of correction space remain when the prediction-space input is centered or normalized (for example, the same bottleneck applied to Ŷbase minus its window mean, or to a normalized Ŷbase)? If the gap in Fig. 5 and Fig. 6 shrinks, the "error mixing" explanation would not hold. If it stays, it would support the paper's claim.

  2. Can error mixing be tested directly? For example, inject a controlled error into one variate's backbone prediction and measure how much it spreads to the other variates' final forecasts in each space.

  3. In Table 4, a per-variate bottleneck with no anchor gives −7.1% vs COSA, and full CoRe gives −9.7%. Can the paper report the gain of the shared anchor alone (with seeds or confidence intervals, per dataset), and revise the claim that cross-variate interaction is the main source of improvement?

  4. In Table 15, several dense-variant entries are identical across different backbones (for example ETTh1 H=96: 0.5133 for iTransformer and DLinear). Are these correct, or is this a table error? Please check the whole table.

  5. Seed variance is much larger for CoRe than for COSA in several Table 12 rows (for example ±0.0049 vs ±0.0001). What causes this?

1 comment

Thank you for your detailed and constructive review, and for your positive assessment of the work; we are sorry that it rested on results that turned out to be invalid. The code we adopted does not follow our stated protocol, and we did not verify it. Details and corrected results are in the General Response. Your points on the headline averages and on the seed variance in Table 12 (Q5) follow from this leak. The controls you proposed were a great suggestion, and we ran them with the leak fixed.

Q1 and Q3. You are right on both. With the leak fixed, neither the correction space nor the shared anchor gives an advantage.

Q2. Yes, we ran this test. An error injected into one variate’s forecast (noise, a ramp or a random walk) spreads further to the other variates through corrections than through normalised predictions; for a constant offset there is no significant difference. Proposition 1 holds, but neither this test nor experiments after the fix support the advantage we attributed to it.

Q4. These are table errors from compiling results.

Reviewer_1GboOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper asks where cross-variate interaction should happen in test-time adaptation (TTA) for multivariate forecasting under delayed ground truth. The method, CoRe, is implemented on top of a base adapter (which was COSA by default). CoRe proposed to perform cross-variate interaction in the correction space, using a shared correction anchor, a learnable bottleneck refinement module, and input-conditioned spectral gating. The method is evaluated using seven forecasting backbones, six standard datasets, and four prediction horizons, with additional experiments on high-dimensional datasets and alternative adaptation modules. The authors report an average MSE reduction of 10.57% over COSA and 25.82% over frozen backbones across the main experimental settings.

Critical strengths
  1. The paper is easy to follow and the problem is well motivated. The authors identifies an important design question of exploiting cross variable information without directly mixing backbone predictions.

  2. The main experiments cover seven forecasting backbones, six datasets, and four prediction horizons. The evaluation across different backbone families helps demonstrate that the proposed approach is not restricted to one forecasting architecture.

  3. The authors performed component ablations.They also evaluate alternative base adapters, including TAFAS and a standalone MLP, providing some evidence that the proposed refinement is not tied exclusively to COSA.

  4. SCR adds about 10K parameters on top of COSA’s roughly 1.63M (Tables 2 and 14). It also works with three different base adapters (COSA, TAFAS, and a two-layer MLP), and the paper uses the same hyperparameters across datasets and backbones.

Critical weaknesses

Major concerns:

CoRe builds on an existing forecasting backbone and a pre-existing test-time adaptation method, using COSA as the default base adapter. Its main contribution is an additional correction-refinement mechanism comprising an averaged correction anchor, a shared bottleneck, and a spectral gate. Overall, the novelty feels somewhat incremental. Although the paper compares correction-space and prediction-space interaction, The paper could do a better job of explaining how CoRe differs from previous methods that use cross-variable attention or shared components, and why the proposed design is a meaningful improvement over them.

The major inconsistency in the paper is that COSA’s errors are very different from those in COSA’s own paper. For example, in Table 1 (Exchange Rate rows, DLinear columns), COSA’s errors are reported as 0.0816, 0.0844, 0.0867 and 0.1469 for 96, 192, 336 and 720 days ahead. COSA’s paper reports 0.0812, 0.1459, 0.2039 and 0.3494 for the same backbone and dataset (Table 2 from COSA's paper). While the H=96 value is about the same, the H=336 and H=720 values are about 2.4 times lower than in COSA’s paper. This is inconsistent because the starting model’s errors are identical in both papers for the first three horizons (0.0913, 0.1827, 0.3277) and close at H=720 (0.8383 here vs. 0.8873). Also, the reported COSA results are nearly flat across horizons. The error grows only from 0.0816 to 0.1469, while the starting model’s error grows about nine times (0.0913 to 0.8383) and COSA’s own paper shows a clear rise (0.0812 to 0.3494). Since CoRe is built on top of COSA, this flatness likely carries over to CoRe (0.0706 to 0.1239), so CoRe’s flat errors and large long-horizon gains may come from the COSA setup rather than from CoRe itself. Notably, the paper runs COSA with a batch size of 25 and 20 adaptation steps, whereas COSA’s own paper reports using 48 and 3. The authors should clarify which configuration they followed and whether this difference explains the gap.

The paper's main contribution is letting different variables share their corrections with each other to make better forecast. However, in the ablation studies (table 4), without sharing, it already lowered the error by 7.1% compared with COSA that is around 7.1/9.7 or around 73.2%. the experiments suggest most of the benefit comes from simply adding another correction step, and the theory doesn’t prove that sharing corrections between variables is what makes the forecasts better.

The paper states that the model can only learn from a forecast once its whole future period has passed (page 4, Sec. 3.1). For Exchange Rate at H=720, the test stream is short compared with this 720-day delay (Fig. 9 shows only a few dozen adaptation steps), so most forecasts are made before the first full answer even arrives. This is also why Wang et al. (2026) leave this setting out of their strict evaluation. As a result, most forecasts should come from a model that has barely learned anything, and the overall error should stay fairly close to the original model’s. Yet Table 1 reports CoRe’s error as about 85% lower than the original model’s (0.1239 vs. 0.8383 for DLinear), which is hard to explain under the stated rule. The authors are encouraged to explain this inconsistency.

The paper claims consistent improvements over existing TSF-TTA methods (contribution 3, page 3) and only mentions a few underperforming cases at H=96 (Sec. 4.2, page 7). But at H=96 in Table 1, CoRe is worse than the original model in 20 of 42 settings, all on the ETT datasets, while TAFAS is worse than the original model in only 2 of those. TAFAS also beats CoRe in 26 of the 42 settings at H=96 (29 of 168 overall), so the gains are not consistent at short horizons.

few minor Issues:

  1. The finding that the leading singular mode captures about 73% of the correction energy is repeated many times in the paper (Sec. 1, page 2; Sec. 4.4, page 8; App. E.1 and Figure 8, page 23; App. G.1, page 28; App. H.1, page 29).
  2. Table numbering is out of order (table 15 appears after table 24).
  3. Table 3's caption says all methods use r = 16 but only CoRe has a rank.

References: J. Im and H.-Y. Kwon. COSA: Context-aware output-space adapter for test-time adaptation in time series forecasting. In The Fourteenth International Conference on Learning Representations (ICLR), 2026. Wang, Haochun, et al. "Towards Principled Test-Time Adaptation for Time Series Forecasting." arXiv preprint arXiv:2605.17250 (2026).

Questions for authors

1.Why are COSA’s errors here so different from COSA’s own paper (mentioned in the weakness), and which settings did you use (wasn't clear in the paper) ? 2. Under the rule in Sec. 3.1, how many Exchange H=720 forecasts are made before the first full answer arrives, and how does CoRe still reach such a low error there? 3. Table 4 shows that most of the gain remains without sharing, so isn’t the extra correction step, rather than sharing, the main source of improvement?

1 comment

Thank you for your careful reading. The problem you noticed is real. The code we adopted does not follow our stated protocol, and we did not verify it. Details and corrected results are in the General Response. Your questions on our COSA numbers (Q1), the Exchange H = 720 results (Q2) and the H = 96 inconsistencies all follow from this leak.

Q3. We tested this, and the new results show that after the fix, CoRe is 1.3% worse than COSA, and the shared anchor makes it a further 0.2% worse. You are right that the gain did not come from sharing.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all four reviewers. Reviewers MY3X and GcBa inspected the released code and found a problem that invalidates our main results. We confirm it, and we will withdraw the submission. Below we describe the problem and report the corrected results and observations.

The problem

We adopted code from the COSA release (March 2026) without checking for later revisions or verifying it against its protocol, which we follow.

As Reviewers MY3X and GcBa found, the adaptation loop trained on each batch’s full targets before forecasting the next batch, so on average 85–98% of each forecast’s targets had already been used. Wang et al. (2026, Sec. 3.2.2), cited by Reviewer 1Gbo, also note that under this streaming update scheme the targets used to update one batch can overlap with the prediction spans of later batches.

Corrected main experiments

We re-ran all methods following the partial-label convention of TAFAS (Kim et al., 2025): gradient updates use only the target steps already observed, and full labels are used only once a batch's last target has been observed. The re-run covers 8 backbones × 6 datasets × 4 horizons with 5 seeds (192 settings; TAFAS with one seed).

  • CoRe vs COSA: CoRe is 1.3% worse (worse in 133 of 192 settings).
  • Against no adaptation: TAFAS improves MSE by about 4%, COSA by 1.4% and CoRe by about 0% (excluding Informer, where adaptation results are unstable).
  • Effect of the leak: in the paper, CoRe reduced MSE by 25.82% on average over the frozen backbones; after the fix, the reduction is 0.55%.
  • SCR on TAFAS: adding SCR to TAFAS changes MSE by +0.1%.
  • COSA: our COSA results after the fix differ substantially from those reported in the COSA paper. This may reflect differences in setup or the target overlap described above.

Ablations after the fix

The shared anchor makes results 0.2% worse. As Reviewer YicS suggested, we applied the same refinement to normalised predictions instead of corrections, which is 1.6% better. Proposition 1 holds, but our experiments after the fix do not support the advantage we attributed to it.

We are sincerely grateful to all four reviewers. Reviewers MY3X and GcBa took the time to read our released code and trace the update timeline, which is how the leak was found; without that effort, the error would have stayed in the literature. Reviewers 1Gbo and YicS identified the anomalies and suggested the controls (centred prediction space, injected errors, the anchor ablation) that help us better analyze the model. We have learned a great deal from this review and apologise for the time spent on results that were not valid.

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

Related papers

Discussion (0)

Sign in to comment.