acceptodds
Under review as a conference paper at ICLR 2027

Adaptation as Memory: Learning to Correct Time Series Foundation Models

Abstract

Time series foundation models (TSFMs) generalize across domains, but adapting them to target-domain dynamics remains challenging. Existing adaptation methods use diverse mechanisms, yet lack a unified framework for understanding how target-domain knowledge is retained, accessed, updated, and applied. We address this gap by viewing TSFM adaptation as a memory process of construction, writing, reading, and execution. We instantiate this view with TS-CorM, a structured fast–slow corrective memory for frozen TSFMs. TS-CorM represents forecast corrections with compact coefficients, addresses them using joint history–forecast states, and retains their associations in explicit memory matrices. Slow Memory consolidates offline knowledge, while Fast Memory incorporates delayed feedback through recursive online updates without backpropagation. Across eight datasets and five TSFM backbones, TS-CorM improves forecasting in all 40 backbone–dataset pairs, with MSE reductions of up to 22.28%; compared with the best adaptation baseline, it reduces average MSE by 3.78% with 98.32% fewer parameters and 10.23% lower inference latency.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Time Series Foundation Model, Model Adaptation, Adaptation as Memory, Associative Memory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UXFNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents TS-CorM, a memory-based framework for adapting frozen time-series foundation models through forecast correction. The method represents forecast residuals using compact coefficients over a structured dictionary and constructs memory keys from joint history–forecast representations. Slow Memory learns offline correction mappings through ridge regression, while Fast Memory incorporates delayed feedback through recursive least squares without backpropagation. Experiments across eight datasets and five TSFM backbones report improvements across all 40 backbone–dataset combinations. Compared with the strongest adaptation baseline, the offline method reduces average MSE by 3.78% with 98.32% fewer adaptation parameters. Online updates provide an additional average MSE reduction of 0.565% in the evaluated settings.

Critical strengths

The paper focuses on how to achieve target domain adaptation of TSFM without modifying pre trained backbone parameters, or it can be understood as post training, which has practical application significance.

From the experimental results in Table 1, it does indeed exceed the fine-tuning method of the compared TSFM, although only ChronosBolt and we have some doubts about the subsequent experimental results.

Critical weaknesses

We have concerns about the transparency and reproducibility of the experimental evaluation. While we respect the authors' decision not to release supplementary materials or source code, the anonymous repository linked in the abstract contains only an essentially empty README file. This is particularly concerning given the existing issues with the fairness and comprehensiveness of the experimental comparisons, the insufficiently documented experimental settings, and the mismatch between the claimed contributions of Slow and Fast Memory and their empirical validation. The lack of detailed experimental configurations and implementation information in the appendix further undermines our confidence in the reliability and reproducibility of the reported results.

W1. Limited benchmark diversity and generalization evidence. Although the paper evaluates eight datasets, four belong to the ETT family, leaving relatively limited diversity in domains and temporal characteristics. TiRex-2 is also evaluated only at horizons 96 and 192. Considering the paper's claim of general TSFM adaptation, I recommend extending the evaluation to broader benchmarks such as GIFT-Eval and fev-bench, using appropriate target-domain adaptation protocols.

W2. Insufficient and potentially unfair comparisons under supervised adaptation. The paper compares supervised TS-CorM adaptation with the original zero-shot TSFMs, but these results do not establish its advantages under comparable training conditions. Since TS-CorM requires labeled target-domain data, it would be important to evaluate how the original TSFMs perform after fine-tuning with the same supervision budget, beyond the reported LoRA baseline. Moreover, once downstream training data are available, conventional long-term forecasting models, such as DLinear, PatchTST, and iTransformer, become relevant alternatives. These models may achieve competitive or better performance without requiring a pretrained TSFM or an additional memory module. Although TS-CorM introduces relatively few adaptation parameters, this alone does not establish its efficiency or practical advantages over lightweight forecasting models trained directly on the target domain. The paper also does not sufficiently clarify whether all adaptation baselines use identical supervised data budgets and training protocols.

W3. Limited contribution and validation of Fast Memory. Although the paper emphasizes a fast–slow memory framework, the main performance comparisons and ablations rely exclusively on the offline Slow Memory. Fast Memory is evaluated separately on only six datasets with one backbone, providing an additional average MSE reduction of just 0.565% and a negligible MAE improvement of 0.061%. Since both memories are essentially linear correction mappings learned through ridge regression or its recursive counterpart, the experiments do not sufficiently establish the necessity of the fast–slow decomposition or its advantages over simpler offline and online residual regression baselines.

W4. Numerical inconsistencies across experimental results. Several results appear inconsistent across the main experiments and appendix. For example, the average Weather MSE in Tables 1 and 2 is approximately 0.228, whereas Table 6 reports 0.174 for the full model. Similarly, the offline MSE for Exchange at horizon 336 differs between Tables 1 and 7 (0.307 vs. 0.332). Without clarification of the underlying configurations, the reliability of these experimental conclusions remains uncertain.

Questions for authors

Q1. Could the authors clarify the numerical inconsistencies across Tables 1, 2, 6, and 7, particularly the different offline results reported for identical dataset–horizon settings?

 Q2. What distinguishes the proposed memory mechanism from conventional ridge regression and RLS? What are the specific advantages of separating Slow Memory and Fast Memory?

 Q3. Figure 2 and Eq. (8) appear to use different residual definitions for Fast Memory. Which definition is used in the actual implementation?

Q4. Can the author provide more experimental settings and code, as many of the experiments in this paper require detailed verification, and providing an empty repository is irresponsible behavior.

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer UXFN (1/3)

W1 Limited benchmark diversity and generalization evidence.

Response: We thank the reviewer for pointing out the limited dataset diversity in our original evaluation. We extended our experiments to five datasets from GIFT-Eval. US Births records daily birth counts, Saugeen measures river flow, and M_DENSE contains urban traffic data. BizITObs Application and BizITObs-L2C cover application monitoring and business-process monitoring, respectively.

Using Chronos-Bolt as the backbone, we evaluated offline target-domain adaptation at horizons of 96, 192, 336, and 720 steps. TS-CorM reduced both MSE and MAE relative to unadapted Chronos-Bolt in all 20 settings. Compared with TS-Memory, TS-CorM achieved lower MSE in 18/20 settings and lower MAE in 18/20, with both metrics lower in 17/20. These results show that the gains from TS-CorM extend beyond the ETT family to all five additional datasets.

Each cell reports MSE / MAE. Bold indicates the best.

DatasetHorizonChronos-Bolt+TS-Memory+TS-CorM (offline)
US Births960.2689 / 0.35000.2159 / 0.32770.1870 / 0.2778
US Births1920.2865 / 0.37190.2572 / 0.36010.1890 / 0.2869
US Births3360.2771 / 0.36560.2493 / 0.34590.1992 / 0.2993
US Births7200.3657 / 0.43920.3161 / 0.40060.2675 / 0.3655
BizITObs Application960.7306 / 0.45080.4321 / 0.40300.0906 / 0.1801
BizITObs Application1920.8437 / 0.51450.6971 / 0.49960.1056 / 0.2136
BizITObs Application3360.9066 / 0.57740.8029 / 0.54630.1273 / 0.2514
BizITObs Application7201.0429 / 0.66850.7968 / 0.60450.1735 / 0.3048
Saugeen961.0787 / 0.46841.0046 / 0.47350.8877 / 0.4318
Saugeen1921.1637 / 0.51161.0890 / 0.50300.8919 / 0.4473
Saugeen3361.1665 / 0.52881.1230 / 0.53050.8876 / 0.4506
Saugeen7201.2803 / 0.59751.1303 / 0.52500.9052 / 0.4642
BizITObs-L2C960.4449 / 0.32990.3197 / 0.32460.3429 / 0.2977
BizITObs-L2C1920.9692 / 0.54470.7080 / 0.50770.6708 / 0.4656
BizITObs-L2C3361.4003 / 0.72900.9661 / 0.70150.9026 / 0.7080
BizITObs-L2C7202.4208 / 1.06382.1273 / 0.99392.0714 / 0.9614
M_DENSE960.2176 / 0.25050.2137 / 0.24380.1884 / 0.2293
M_DENSE1920.2269 / 0.26160.1891 / 0.23140.1902 / 0.2414
M_DENSE3360.2384 / 0.27620.2092 / 0.25980.1985 / 0.2036
M_DENSE7200.3150 / 0.33830.2516 / 0.29910.2330 / 0.2781
Average—0.7822 / 0.48190.6550 / 0.45410.4855 / 0.3679

Q1 & W4. Numerical inconsistencies across experimental results.

Response: Thank you for pointing this out. We sincerely apologize for this oversight. The discrepancy between the offline MSE for Weather in Table 2 (0.228) and Table 6 (0.174) arose because we averaged only the results for H=96H=96 and H=192H=192 in Table 6, inadvertently omitting H=336H=336 and H=720H=720. We have recalculated the results across all four horizons. The corrected offline MSE is 0.228, consistent with Table 2. The corrected Weather results for all configurations are shown below.

ConfigurationWeather
Full0.228 / 0.255
w/o dictionary (direct residual)0.234 / 0.260
w/o historical directions0.236 / 0.259
w/o DCT directions0.232 / 0.257
w/o forecast embedding0.229 / 0.255
w/o history embedding0.232 / 0.259
w/o local memory0.231 / 0.257
w/o shared memory0.230 / 0.257

The discrepancy between Tables 1 and 7 has a different source. Table 1 presents our main experimental results and reports offline performance averaged across multiple random seeds. For the Fast Memory experiment in Table 7, we used a saved checkpoint from a single seed for both the offline and online runs. Consequently, some offline results in Table 7 differ slightly from those in Table 1. Using the same checkpoint for both runs in Table 7 allows us to isolate the effect of enabling Fast Memory. We will clarify how the results were aggregated in the captions of both tables. We apologize again for this oversight.

AuthorsOpenReview ↗

Author Response to Reviewer UXFN (2/3)

W2 Insufficient and potentially unfair comparisons under supervised adaptation.

Response: We thank the reviewer for raising this point. To evaluate the benefit of TS-CorM when labeled target-domain data are available, we added full fine-tuning of Chronos-Bolt alongside the existing LoRA baseline. We also evaluated DLinear, PatchTST, and iTransformer trained directly on each target dataset. The table below reports results on six datasets at four forecast horizons (96, 192, 336, and 720 steps). It also includes FW-FC, EC-FC, and MEMTS, the additional baselines requested by Reviewer VUFa in W1 & Q1.

Across the 24 settings, offline TS-CorM achieved the lowest average MSE/MAE (0.326/0.350), compared with 0.338/0.358 for fully fine-tuned Chronos-Bolt and 0.359/0.391 for DLinear, the strongest of the three directly trained models. TS-CorM achieved the lowest MSE in 20/24 settings and the lowest MAE in 22/24 settings. We will include these additional experiments in the revised manuscript.

DatasetHorizonTS-CorM(offline)FW-FCEC-FCMEMTSChronos-Bolt(full FT)DLinearPatchTSTiTransformer
ETTh1960.359/0.3720.364/0.3820.362/0.3810.394/0.3920.372/0.3750.372/0.3980.390/0.4220.397/0.420
ETTh11920.405/0.4030.414/0.4160.413/0.4140.460/0.4400.425/0.4060.407/0.4210.464/0.4700.430/0.443
ETTh13360.440/0.4220.456/0.4430.465/0.4470.479/0.4390.461/0.4280.443/0.4470.511/0.4960.454/0.458
ETTh17200.445/0.4420.479/0.4790.484/0.4790.610/0.5410.475/0.4530.474/0.4960.587/0.5290.465/0.482
ETTh2960.279/0.3230.381/0.3850.370/0.3790.289/0.3270.275/0.3180.272/0.3410.328/0.3740.326/0.374
ETTh21920.335/0.3630.604/0.4650.604/0.4650.372/0.4030.360/0.3720.352/0.3980.413/0.4300.384/0.412
ETTh23360.369/0.3930.706/0.5090.522/0.4730.393/0.4050.396/0.4010.406/0.4420.436/0.4490.413/0.443
ETTh27200.381/0.4110.551/0.5070.555/0.5150.423/0.4550.398/0.4190.716/0.6000.429/0.4600.437/0.465
ETTm1960.293/0.3210.301/0.3340.294/0.3350.332/0.3390.296/0.3270.307/0.3510.333/0.3680.330/0.371
ETTm11920.336/0.3490.343/0.3630.339/0.3610.355/0.3730.341/0.3500.339/0.3720.377/0.3930.370/0.394
ETTm13360.385/0.3760.381/0.3880.370/0.3860.412/0.4100.393/0.3940.387/0.4060.372/0.4000.381/0.406
ETTm17200.450/0.4130.445/0.4280.429/0.4280.478/0.4490.444/0.4200.451/0.4510.430/0.4330.447/0.444
ETTm2960.166/0.2390.175/0.2550.252/0.3210.174/0.2470.169/0.2440.169/0.2590.180/0.2700.189/0.273
ETTm21920.229/0.2840.249/0.3100.245/0.3070.259/0.3170.232/0.2900.230/0.3110.234/0.3040.266/0.323
ETTm23360.283/0.3210.312/0.3560.307/0.3520.317/0.3360.289/0.3380.284/0.3470.295/0.3500.336/0.367
ETTm27200.362/0.3740.404/0.4190.396/0.4150.423/0.4390.380/0.3770.407/0.4270.375/0.3980.400/0.410
Exchange960.090/0.2091.091/0.5271.102/0.5280.092/0.2150.090/0.2100.083/0.2070.205/0.3260.139/0.275
Exchange1920.174/0.2971.510/0.6671.537/0.6650.179/0.2990.177/0.3010.178/0.3060.196/0.3220.254/0.371
Exchange3360.307/0.4012.962/0.8743.026/0.8920.342/0.4240.352/0.4220.529/0.5360.771/0.7010.388/0.454
Exchange7200.821/0.6802.898/1.0261.402/0.8521.001/0.7600.828/0.6970.834/0.6901.240/0.8721.200/0.854
Weather960.153/0.1920.168/0.2200.166/0.2190.183/0.2120.163/0.1920.177/0.2430.158/0.2100.163/0.218
Weather1920.195/0.2330.212/0.2620.212/0.2620.225/0.2730.201/0.2490.213/0.2700.207/0.2540.213/0.264
Weather3360.246/0.2720.270/0.3040.269/0.3030.277/0.2960.251/0.2750.257/0.3070.265/0.2990.261/0.300
Weather7200.316/0.3210.349/0.3560.348/0.3550.367/0.3980.344/0.3330.327/0.3680.339/0.3500.329/0.341
Average—0.326/0.3500.668/0.4450.603/0.4390.368/0.3830.338/0.3580.359/0.3910.397/0.4120.374/0.398
AuthorsOpenReview ↗

Author Response to Reviewer UXFN (3/3)

Q2 & W3 Limited contribution and validation of Fast Memory.

Response:

1. Extending the test sequences. The modest 0.061% improvement in average MAE for Fast Memory in the paper is mainly due to the fact that Fast Memory often has a negative effect at longer forecast horizons. This is because, for a forecast horizon of HH, the model must wait HH steps to receive complete feedback. Therefore, for a test sequence of fixed length, the longer the horizon, the less feedback is available for updates. We adjusted the proportions of the training, validation, and test sets to extend the test sequences. Under the new split, enabling Fast Memory reduced both MSE and MAE on all six datasets compared with using frozen Slow Memory alone. Average MSE and MAE decreased by 3.23% and 2.31%, respectively, compared with reductions of 0.565% and 0.061% under the original split. This result shows that the online gains persist and are more pronounced on longer test sequences.

DatasetTS-CorM (offline)TS-CorM (online)ΔMSEΔMAE
ETTm10.537722 / 0.4486330.529754 / 0.447623−1.482%−0.225%
ETTm20.382896 / 0.3877740.379025 / 0.386001−1.011%−0.457%
Weather0.392253 / 0.4026310.390255 / 0.401152−0.509%−0.367%
ETTh10.662415 / 0.6923470.601727 / 0.643529−9.162%−7.051%
ETTh20.432258 / 0.4273570.402891 / 0.409825−6.794%−4.102%
Exchange0.857721 / 0.7002410.856021 / 0.700184−0.198%−0.008%
Average0.544211 / 0.5098310.526612 / 0.498052−3.23%−2.31%

2. Contribution of separating Slow Memory and Fast Memory and its relationship to ridge regression and RLS. Slow Memory and Fast Memory both use established regression methods. They differ in what they learn and how they correct the forecast. We represent TSFM forecast residuals as correction coefficients. Slow Memory learns an offline mapping from the forecast state to these coefficients and uses them to correct the backbone forecast. Once the ground truth for the full forecast horizon is available, Fast Memory uses RLS to learn an online correction for the remaining residual. This design preserves the offline mapping and allows us to adjust the strength of the online correction independently.

To test the benefit of separating Slow and Fast Memory, we added a direct-update baseline that updates Slow Memory as feedback becomes available. The table reports average results across six datasets and four forecast horizons using Chronos-Bolt.

VariantMean MSE ↓Mean MAE ↓
A. Frozen Slow Memory0.3308710.353754
B. Direct online updates to Slow Memory0.3295640.353695
C. Slow Memory + Fast Memory0.3290020.353538

Direct updates improved performance over Slow Memory. Separating Slow and Fast Memory yielded further improvement over direct updates, with lower MSE in 20/24 settings, lower MAE in 20/24. The relative reductions averaged 0.171% and 0.044% for MSE and MAE. We will include this comparison in the revised manuscript.

Q3 Figure 2 and Eq. (8) appear to use different residual definitions for Fast Memory. Which definition is used in the actual implementation?

Response: Thank you for pointing this out. The residual formula in Figure 2(c) is incorrect and should match Eq. (8). We will correct it to �2��2�, where �3��3� is the forecast corrected by Slow Memory.

Q4. Experimental settings and code availability

Response: Thank you for raising this concern. We apologize for providing an empty anonymous repository. We originally intended to release the code at submission. Since this work was conducted during a company internship, public release requires the company’s confidentiality approval. We should have made this clear when we shared the link.

We are committed to releasing the code for all experiments as soon as approval is granted. In the meantime, we will expand the appendix with the data splits, fitting procedures, online update protocol, and aggregation rules used for each table to make the experimental protocol explicit and support reproducibility.

AuthorsOpenReview ↗

To Reviewer UXFN

Dear Reviewer UXFN,

Thank you for your careful review and constructive feedback. We appreciate your recognition of the practical value of adapting time series foundation models while keeping their backbones frozen.

In our rebuttal, we expanded the evaluation to include GIFT-Eval and added supervised comparisons with full fine-tuning of Chronos-Bolt and with DLinear, PatchTST, and iTransformer trained directly on the target datasets. We also evaluated Fast Memory on longer test sequences, explained why we separate Slow and Fast Memory, and clarified the residual definition used for online updates. Finally, we addressed the numerical discrepancies across tables, provided more details on the experimental settings, and explained the current status of the code release.

We would appreciate your feedback on whether our responses address your concerns. If you find them satisfactory, we would be grateful for your support. Please let us know if you have any further questions. Thank you again for your time and consideration.

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

The authors propose TS-CorM, a system for adapting TSFM to a targeted domain. TS-CorM modifies the predicted forecasts directly by adding a correction to the model’s zero-shot prediction rather than modifying the FM’s latent space. This correction is broken into a slow, offline memory and a fast online memory. Both memories work using a key (forecast state) and a value (correction coefficients) where the queried correction coefficients weight a fixed correction dictionary. They show that TS-CorM outperforms existing parametric adaptation/memory and RAG-style adaptation methods across 8 datasets with both faster inference time and less parameters than TS-Memory.

Critical strengths
  • The methodology is clearly explained and easy to follow with clean contributions.
  • The ablations are appreciated and slow memory results are positive.

The paper was relatively easy to follow and the ablations were complete. Assuming correct benchmarks (see weaknesses) this paper furthers progress in TSFM domain adaptation.

Critical weaknesses
  • Missing citations and baselines
    • The paper lacks relevant prior work citations like the following (but not limited to):
      • Frequency-Aware Calibration (FAC) which also calibrates forecasts using a frequency correction component (similar to their DCT dictionary) [1]
      • Context-aware Output-Space Adapter (COSA) a gated test time adaptation to correct a TSFM forecast [2]
    • The paper never compares their online fast memory adapter to an external baseline. A strong baseline to compare against would be ELF [3]
  • Slow memory adaptation comparison to external baselines was only reported for a single Chronos-Bolt backbone and not other TSFM backbones. It would be good to have Table 1 style results for additional backbones - a table showing average rank per backbone would suffice.

The paper clearly missed a sizable set of prior work which on its own is not strong enough for rejection, but this in combination with limited baseline comparisons may indicate cherry picked results. If this is not addressed during rebuttal I may consider reducing my score.

[1] Wang, Haochun, et al. "Towards Principled Test-Time Adaptation for Time Series Forecasting." arXiv preprint arXiv:2605.17250 (2026).

[2] Im, Jeonghwan, and Hyuk-Yoon Kwon. "Cosa: Context-aware output-space adapter for test-time adaptation in time series forecasting." The Fourteenth International Conference on Learning Representations. 2026.

[3] Lee, Thomas L., et al. "Lightweight online adaption for time series foundation model forecasts." arXiv preprint arXiv:2502.12920 (2025).

Questions for authors
  • Chronos-Bolt is used as a backbone rather than the improved Chronos-2 model with native multi-variate support. Chronos-2 model should be included in your benchmarks.
  • In your inference latency comparison adapter only inference is minor compared to backbone latency. However, since you’re using ChronosBolt as the backbone you can reuse its encoder in the adapter. If you are using a different backbone then you must run the Chronos encoder separately increasing latency. Is this interpretation correct and if so what are the Table 3 results for TimesFM-3 as the backbone?
  • Table 4 seems misleading as TS-memory parameter count is constant regardless of channel count, so Electricity and Traffic result in a higher parameter count than TS-memory.
4 comments
AuthorsOpenReview ↗

Author Response to Reviewer VnTi (1/2)

Q1 Additional experiments on Chronos-2.

Thank you for suggesting that we evaluate our method using Chronos-2 as the backbone. We used the same dataset splits as in Table 1 and evaluated all five methods on eight datasets at four forecast horizons (96, 192, 336, and 720), for a total of 32 settings. Each cell reports MSE / MAE. Bold indicates the best.

DatasetHorizonChronos-2+TS-Memory+RAFT-style+TS-RAG-style+TS-Corm (offline)
ETTh1960.3978 / 0.38090.3786 / 0.37780.3955 / 0.39120.4063 / 0.40780.3528 / 0.3746
ETTh11920.4407 / 0.40840.4119 / 0.40780.4371 / 0.42480.4594 / 0.44930.4035 / 0.4062
ETTh13360.4712 / 0.42440.4367 / 0.42730.4892 / 0.47160.5084 / 0.48480.4390 / 0.4214
ETTh17200.4672 / 0.43550.4377 / 0.45040.6330 / 0.58500.6400 / 0.58490.4363 / 0.4284
ETTh2960.3023 / 0.32730.2891 / 0.32910.3023 / 0.32730.3023 / 0.32730.2814 / 0.3222
ETTh21920.3805 / 0.37960.3615 / 0.37530.3805 / 0.37960.3805 / 0.37960.3430 / 0.3689
ETTh23360.4038 / 0.40440.3812 / 0.40060.3974 / 0.41680.4008 / 0.41560.3653 / 0.3935
ETTh27200.4048 / 0.41660.3825 / 0.41340.4095 / 0.44650.4002 / 0.42560.3784 / 0.4113
ETTm1960.3347 / 0.32640.2811 / 0.31870.3236 / 0.33370.3244 / 0.33590.2927 / 0.3173
ETTm11920.4013 / 0.36390.3503 / 0.36370.3851 / 0.37660.3912 / 0.38280.3401 / 0.3477
ETTm13360.4533 / 0.39340.3906 / 0.38940.4332 / 0.41190.4459 / 0.42340.3758 / 0.3709
ETTm17200.5419 / 0.43840.4816 / 0.42960.5160 / 0.47330.5480 / 0.49620.4311 / 0.4062
ETTm2960.1797 / 0.24160.1675 / 0.24250.1797 / 0.24160.1797 / 0.24160.1653 / 0.2367
ETTm21920.2550 / 0.29150.2301 / 0.28990.2550 / 0.29150.2550 / 0.29150.2287 / 0.2815
ETTm23360.3196 / 0.33310.2843 / 0.32780.3196 / 0.33310.3144 / 0.34320.2788 / 0.3180
ETTm27200.4264 / 0.39480.3888 / 0.39420.4158 / 0.40620.4161 / 0.40410.3598 / 0.3719
Exchange960.0913 / 0.20850.0881 / 0.20530.0913 / 0.20850.0913 / 0.20850.0838 / 0.2006
Exchange1920.1967 / 0.31400.1920 / 0.31540.1761 / 0.30340.1770 / 0.30480.1718 / 0.2960
Exchange3360.3627 / 0.43310.4019 / 0.46170.3627 / 0.43310.3627 / 0.43310.3219 / 0.4119
Exchange7200.9435 / 0.72650.9402 / 0.73510.9455 / 0.73450.9439 / 0.73160.9328 / 0.7267
Weather960.1711 / 0.19040.1553 / 0.20390.4908 / 0.42630.4121 / 0.36080.1551 / 0.1860
Weather1920.2282 / 0.24130.1974 / 0.24280.4203 / 0.39490.4007 / 0.36110.1820 / 0.2322
Weather3360.3012 / 0.28870.2479 / 0.28530.4253 / 0.39740.4007 / 0.36920.2470 / 0.2736
Weather7200.4054 / 0.34760.3863 / 0.36520.4647 / 0.41130.4471 / 0.38910.3379 / 0.3262
Electricity960.1274 / 0.21100.1271 / 0.21010.1265 / 0.21320.1267 / 0.21310.1251 / 0.2097
Electricity1920.1469 / 0.22970.1445 / 0.22800.1454 / 0.23120.1458 / 0.23120.1437 / 0.2278
Electricity3360.1667 / 0.24960.1629 / 0.24760.1645 / 0.25490.1646 / 0.25030.1622 / 0.2470
Electricity7200.2099 / 0.28650.2051 / 0.28520.2043 / 0.29340.2048 / 0.28870.2033 / 0.2825
Traffic960.3563 / 0.21850.3533 / 0.21850.3543 / 0.22050.3542 / 0.22140.3502 / 0.2185
Traffic1920.3825 / 0.23040.3804 / 0.22620.3793 / 0.23190.3792 / 0.23290.3729 / 0.2200
Traffic3360.3979 / 0.23810.3904 / 0.23630.3941 / 0.23950.3941 / 0.24050.3925 / 0.2372
Traffic7200.4441 / 0.26280.4398 / 0.26350.4395 / 0.26380.4393 / 0.26460.4370 / 0.2622
Average—0.3473 / 0.33240.3271 / 0.33340.3705 / 0.36150.3693 / 0.35920.3154 / 0.3230

With Chronos-2 as the backbone, TS-Corm achieved the lowest average MSE (0.3154) and MAE (0.3230) across 32 settings, improving on Chronos-2 alone by 9.2% and 2.8%, respectively. Together with the Chronos-Bolt results in Table 1, these gains demonstrate the effectiveness of our method across different backbones.

AuthorsOpenReview ↗

Author Response to Reviewer VnTi (2/2)

Q2 Inference latency with TimesFM-3

Response: Thank you for raising this point. We apologize for not making the timing protocol clear in Section 4.6. In Table 3, TS-CorM used a separate frozen Chronos-Bolt encoder to embed the history and the forecast. We did not reuse encoder activations from the forecasting backbone. The reported end-to-end latency therefore already includes both encoder forward passes. This timing protocol applies across backbones because it does not rely on encoder sharing.

We benchmarked TimesFM-3 on an NVIDIA H100 (FP32, batch size 1, context length 512), timing 20 fixed requests five times each per setting. The table below reports latency in milliseconds per query.

& \textbf{End-to-end inference} & \textbf{Adapter-only inference} \\\\ \hline \begin{array}{c} \text{Horizon} \\\\ 96 \\\\ 192 \\\\ 336 \\\\ 720 \\\\ \hline \textbf{Overall} \end{array} & \begin{array}{rrr} \text{TimesFM-3} & \text{+ TS-CorM} & \text{+ TS-Memory} \\\\ 114.215 & 120.155 & 119.591 \\\\ 116.512 & 120.093 & 121.672 \\\\ 127.752 & 133.440 & 134.162 \\\\ 130.154 & 135.282 & 136.267 \\\\ \hline \mathbf{122.158} & \mathbf{127.243} & \mathbf{127.923} \end{array} & \begin{array}{rr} \text{TS-CorM} & \text{TS-Memory} \\\\ 5.940 & 5.376 \\\\ 3.581 & 5.160 \\\\ 5.688 & 6.410 \\\\ 5.128 & 6.113 \\\\ \hline \mathbf{5.084} & \mathbf{5.765} \end{array} \end{array}$$ Averaged across the four horizons, TS-CorM adds **5.084 ms per query** to TimesFM-3 inference (**4.16%**), including both encoder passes. Its adapter-only latency is **0.681 ms (11.8%) lower** than the estimated **5.765 ms** for TS-Memory. > ### **Q3 Adaptation parameter count.** **Response:** Thank you for pointing this out. TS-Memory shares its 5.47M parameters across all channels, and its total parameter count is indeed lower than ours on datasets with **a large number of channels**, such as **Electricity (321 channels)** and **Traffic (862 channels)**. Table 4 was intended to show that forecasting a single channel with TS-CorM uses only 92.16K fitted parameters (**98.32% fewer than TS-Memory**), with 46.08K in the shared memory and 46.08K in that channel’s local memory. This design inevitably causes the total parameter count to increase as the number of channels grows. For a dataset with **C** channels, the total is **46,080 × (C + 1)**. Even when counting all channels, TS-CorM uses only **0.37M–1.01M** parameters on six of the eight datasets in Table 1 and achieves **the lowest MSE and MAE among all baselines.** We will revise Table 4 to report both single-channel and dataset-level parameter counts. > ### **W1 Missing citations and baselines.** **Response:** Thank you for highlighting these relevant works and for suggesting ELF as an external baseline for Fast Memory. We will add citations to FAC [1], COSA [2], and ELF [3] in the revision. And we agree that a comparison with ELF would strengthen the online evaluation. To date, **we have been unable to find a public implementation from the authors.** We implemented ELF based on the paper, but obtained poor results in our experiments. To avoid underestimating ELF’s performance, we did not include these results in the formal comparison. > ### **W2 Limited baseline comparisons across TSFM backbones** Thank you for this suggestion. We extended the baseline comparison to **Chronos-2** across the same eight datasets and four horizons as Table 1. TS-CorM achieves the lowest average MSE and MAE among the evaluated methods. Given the limited rebuttal period, we have completed the additional comparison on Chronos-2 so far. We will evaluate the remaining backbones and include these results in the final version if accepted.
Reviewer_VnTiOpenReview ↗

Review Response

Thank you. This has resolved all my concerns and I will be keeping my score as weak accept (3).

In your comparison table RAFT/TS-RAG Style seem to do worse than the baseline. I assume this is a similar finding as to ELF where your own implementation underperforms?

Thank you for your feedback and for taking the time to review our additional experiments. We appreciate your support for our work and your final decision. We will continue to refine and tune our implementations of the RAFT-style and TS-RAG-style baselines to better assess their performance. Thank you again for your careful reading and constructive feedback.

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

The paper frames TSFM adaptation as a memory process with four operations — construction, writing, reading, and execution — and proposes TS-CorM, a structured fast–slow corrective memory for frozen TSFMs. Corrections are stored as compact coefficients over a DCT + history-derived dictionary, keyed by joint history–forecast embeddings: slow memory is fitted offline via ridge regression, while fast memory updates online via RLS, requiring no backpropagation and no inference-time retrieval. Across five backbones and eight datasets, TS-CorM improves all 40 backbone–dataset pairs with fewer parameters and lower adaptation latency. The main weaknesses are: 1) limited technical novelty — the framework mostly re-labels existing methods, and the components (dictionary expansion, ridge regression, RLS) are standard; 2) marginal gains from fast memory (−0.565% avg MSE) with degradation in several settings; and 3) insufficient evaluation — no significance tests, no ablation of the Chronos-encoder dependency for keys, and online updates untested on high-channel datasets.

Critical strengths
  1. The structured correction dictionary provides a sensible inductive bias. Ablations support its effectiveness, particularly the contribution of history-derived directions.

  2. The method is simple to fit, requires no inference-time retrieval, and supports online updates without backpropagation. The feedback protocol explicitly waits for fully observed target horizons.

  3. Offline improvements are reported across all 40 backbone–dataset combinations after averaging over horizons. The main baseline comparison disables Fast Memory, separating offline gains from test-time feedback.

Critical weaknesses
  1. Novelty is limited mainly to the correction representation and combination of established components. Historical directions effectively enable adaptive combinations of historical reference and backbone forecasts, but comparisons with simpler forecast-combination models are missing. The closely related MEMTS is also not evaluated, leaving the incremental contribution insufficiently established.

  2. Fast Memory improves average MSE by 0.565% and MAE by 0.061%, without uncertainty estimates or online baselines. Its cumulative update lacks explicit temporal forgetting, and no controlled drift experiment establishes its responsiveness to changing dynamics.

  3. Unexplained result differences weaken confidence in the evaluation: Weather's offline MSE is 0.228 in Table 2 versus 0.174 in Table 6. Efficiency claims also need qualification: the 10.23% latency reduction is adapter-only, parameter counts exclude online state, and inclusion of all encoding costs is unclear.

Questions for authors
  1. Do the gains persist against simple forecast-combination models using the same historical references and embeddings, and against MEMTS?

  2. Are the online gains robust, and how does Fast Memory compare with forgetting-factor RLS or sliding-window regression under matched feedback delays and controlled drift?

  3. Could the authors reconcile the results across tables and clarify the fitting protocols, full encoding latency, and online time and memory costs?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer VUFa (1/3)

W1 & Q1. Do the gains persist against simple forecast-combination models using the same historical references and embeddings, and against MEMTS?

Response: We thank the reviewer for suggesting a direct comparison with simple forecast-combination models and MEMTS. To test whether reweighting historical references is sufficient to match Slow Memory, we evaluated two baselines that use the same Chronos-Bolt forecasts and historical reference sequences as TS-CorM.

  • Fixed-Weight Forecast Combination (FW-FC) learns one set of reference weights per channel from the target-domain training data.
  • Embedding-Conditioned Forecast Combination (EC-FC) uses the same state representation as our Slow Memory to generate reference weights for each forecast. It learns this mapping by directly minimizing the final forecast error.

Together, these two baselines test whether adapting the combination weights to the current forecast state is sufficient to match the accuracy of our Slow Memory. We also evaluated a local reproduction of MEMTS based on its official KPM + AF architecture. To assess performance under supervised adaptation, we compared Slow Memory with full-parameter fine-tuning of Chronos-Bolt and with DLinear, PatchTST, and iTransformer trained on the target-domain data. All methods were evaluated on the test splits and four forecast horizons used in Table 1. Each cell reports MSE / MAE. Bold indicates the lowest value.

Across the 24 settings, Slow Memory achieved lower MSE than each forecast-combination baseline in 22 settings and lower MAE in all 24. It also achieved lower MSE and MAE than MEMTS in every setting, with average MSE/MAE of 0.326/0.350 for Slow Memory and 0.368/0.383 for MEMTS. Slow Memory retained the lowest average error on both metrics when the comparison was extended to the supervised forecasting models. These results support the effectiveness of the Slow Memory design.

DatasetHorizonTS-CorM(offline)FW-FCEC-FCMEMTSChronos-Bolt(full FT)DLinearPatchTSTiTransformer
ETTh1960.359/0.3720.364/0.3820.362/0.3810.394/0.3920.372/0.3750.372/0.3980.390/0.4220.397/0.420
ETTh11920.405/0.4030.414/0.4160.413/0.4140.460/0.4400.425/0.4060.407/0.4210.464/0.4700.430/0.443
ETTh13360.440/0.4220.456/0.4430.465/0.4470.479/0.4390.461/0.4280.443/0.4470.511/0.4960.454/0.458
ETTh17200.445/0.4420.479/0.4790.484/0.4790.610/0.5410.475/0.4530.474/0.4960.587/0.5290.465/0.482
ETTh2960.279/0.3230.381/0.3850.370/0.3790.289/0.3270.275/0.3180.272/0.3410.328/0.3740.326/0.374
ETTh21920.335/0.3630.604/0.4650.604/0.4650.372/0.4030.360/0.3720.352/0.3980.413/0.4300.384/0.412
ETTh23360.369/0.3930.706/0.5090.522/0.4730.393/0.4050.396/0.4010.406/0.4420.436/0.4490.413/0.443
ETTh27200.381/0.4110.551/0.5070.555/0.5150.423/0.4550.398/0.4190.716/0.6000.429/0.4600.437/0.465
ETTm1960.293/0.3210.301/0.3340.294/0.3350.332/0.3390.296/0.3270.307/0.3510.333/0.3680.330/0.371
ETTm11920.336/0.3490.343/0.3630.339/0.3610.355/0.3730.341/0.3500.339/0.3720.377/0.3930.370/0.394
ETTm13360.385/0.3760.381/0.3880.370/0.3860.412/0.4100.393/0.3940.387/0.4060.372/0.4000.381/0.406
ETTm17200.450/0.4130.445/0.4280.429/0.4280.478/0.4490.444/0.4200.451/0.4510.430/0.4330.447/0.444
ETTm2960.166/0.2390.175/0.2550.252/0.3210.174/0.2470.169/0.2440.169/0.2590.180/0.2700.189/0.273
ETTm21920.229/0.2840.249/0.3100.245/0.3070.259/0.3170.232/0.2900.230/0.3110.234/0.3040.266/0.323
ETTm23360.283/0.3210.312/0.3560.307/0.3520.317/0.3360.289/0.3380.284/0.3470.295/0.3500.336/0.367
ETTm27200.362/0.3740.404/0.4190.396/0.4150.423/0.4390.380/0.3770.407/0.4270.375/0.3980.400/0.410
Exchange960.090/0.2091.091/0.5271.102/0.5280.092/0.2150.090/0.2100.083/0.2070.205/0.3260.139/0.275
Exchange1920.174/0.2971.510/0.6671.537/0.6650.179/0.2990.177/0.3010.178/0.3060.196/0.3220.254/0.371
Exchange3360.307/0.4012.962/0.8743.026/0.8920.342/0.4240.352/0.4220.529/0.5360.771/0.7010.388/0.454
Exchange7200.821/0.6802.898/1.0261.402/0.8521.001/0.7600.828/0.6970.834/0.6901.240/0.8721.200/0.854
Weather960.153/0.1920.168/0.2200.166/0.2190.183/0.2120.163/0.1920.177/0.2430.158/0.2100.163/0.218
Weather1920.195/0.2330.212/0.2620.212/0.2620.225/0.2730.201/0.2490.213/0.2700.207/0.2540.213/0.264
Weather3360.246/0.2720.270/0.3040.269/0.3030.277/0.2960.251/0.2750.257/0.3070.265/0.2990.261/0.300
Weather7200.316/0.3210.349/0.3560.348/0.3550.367/0.3980.344/0.3330.327/0.3680.339/0.3500.329/0.341
Average—0.326/0.3500.668/0.4450.603/0.4390.368/0.3830.338/0.3580.359/0.3910.397/0.4120.374/0.398
AuthorsOpenReview ↗

Author Response to Reviewer VUFa (2/3)

Q2 Are the online gains robust, and how does Fast Memory compare with forgetting-factor RLS or sliding-window regression under matched feedback delays and controlled drift?

Response: We thank the reviewer for raising this question. We address it in three parts below.

1. Extending the test sequences. The modest 0.061% improvement in average MAE reported for Fast Memory in the paper is mainly due to the fact that Fast Memory often has a negative effect at longer forecast horizons. This is because, for a forecast horizon of HH, the model must wait HH steps to receive complete feedback. Therefore, for a test sequence of fixed length, the longer the horizon, the less feedback is available for updates. We adjusted the proportions of the training, validation, and test sets to extend the test sequences. Under the new split, enabling Fast Memory reduced both MSE and MAE on all six datasets compared with using frozen Slow Memory alone. Average MSE and MAE decreased by 3.23% and 2.31%, respectively, compared with reductions of 0.565% and 0.061% under the original split. This result shows that the online gains persist and are more pronounced on longer test sequences.

DatasetTS-CorM (offline)TS-CorM (online)ΔMSEΔMAE
ETTm10.537722 / 0.4486330.529754 / 0.447623−1.482%−0.225%
ETTm20.382896 / 0.3877740.379025 / 0.386001−1.011%−0.457%
Weather0.392253 / 0.4026310.390255 / 0.401152−0.509%−0.367%
ETTh10.662415 / 0.6923470.601727 / 0.643529−9.162%−7.051%
ETTh20.432258 / 0.4273570.402891 / 0.409825−6.794%−4.102%
Exchange0.857721 / 0.7002410.856021 / 0.700184−0.198%−0.008%
Average0.544211 / 0.5098310.526612 / 0.498052−3.23%−2.31%

2. Comparison with online baselines. We compared our method with forgetting-factor RLS and sliding-window ridge regression across 24 settings on the original splits. All three methods shared the same frozen Slow Memory, residual targets, and feedback delay of HH steps. The forgetting half-life and sliding-window length were selected on the validation set as 32 and 128 updates. Fast Memory achieved the lowest average MSE/MAE (0.329002/0.353538), outperforming forgetting-factor RLS (0.330962/0.353811) and sliding-window ridge regression (0.331103/0.353928). Compared with frozen Slow Memory, Fast Memory reduced MSE in 21/24 settings and MAE in 16/24 settings.

MethodAverage MSE ↓Average MAE ↓ΔMSE (%)ΔMAE (%)Settings with both metrics lower than Frozen Slow
Slow Memory0.3308710.353754———
Fast Memory0.3290020.353538−0.565%−0.061%16/24
Forgetting-factor RLS0.3309620.353811+0.028%+0.016%9/24
Sliding-window ridge regression0.3311030.353928+0.070%+0.049%9/24

3. Controlled drift experiment. We tested whether Fast Memory can adapt when a time series begins to evolve more slowly. From a chosen point in the test stream, we slowed the series to 75% of its original speed. For example, a pattern that repeats every 24 steps would now repeat every 32 steps, so a model following the old rhythm would predict peaks too early. We introduced this slowdown either abruptly or gradually over 2H2H steps.

We tested three datasets (ETTh2, ETTm1, and Weather), two forecast horizons (H=96,192H=96,192), and five drift locations per test stream, giving 60 settings across the two drift types. All methods forecast the same modified series and received feedback after the same HH-step delay. We measured post-drift error once a complete forecast window at the new speed had been observed. To isolate the value of continued updates, we also froze a copy of Fast Memory at drift onset, preserving everything it had learned up to that point.

MethodPost-drift MSE ↓Post-drift MAE ↓
Frozen Slow Memory0.3863260.362664
Fast Memory frozen at drift onset0.3851590.362985
Forgetting-factor RLS0.3858300.362505
Sliding-window ridge regression0.3854280.362474
Fast Memory0.3798870.360884

Fast Memory achieved the lowest average post-drift MSE and MAE. Its MSE was lower than that of forgetting-factor RLS in 58/60 settings and sliding-window ridge in 43/60, with average reductions of 1.54% and 1.44%, respectively. Compared with freezing Fast Memory at drift onset, continuing to update reduced average MSE by 1.37% and MAE by 0.58%. Across individual settings, MSE gains were more consistent than MAE gains.

AuthorsOpenReview ↗

Author Response to Reviewer VUFa (3/3)

W3 & Q3 Inconsistencies across tables, fitting protocols, end-to-end latency including all encoding costs, and online time and memory costs.

Response:

1. Inconsistencies across tables: Thank you for pointing this out. The discrepancy between the offline MSE for Weather in Table 2 (0.228) and Table 6 (0.174) was due to an error in our calculation for Table 6. We averaged the results for H = 96 and H = 192 only, inadvertently omitting H = 336 and H = 720. We have recalculated the results across all four horizons. The corrected Weather results for Table 6 are shown below.

ConfigurationETTh1ETTh2ETTm1ETTm2ExchangeWeather
Full0.412 / 0.4110.342 / 0.3730.364 / 0.3640.261 / 0.3050.348 / 0.3970.228 / 0.255
w/o dictionary (direct residual)0.464 / 0.4410.361 / 0.3870.388 / 0.3750.272 / 0.3120.407 / 0.4170.234 / 0.260
w/o historical directions0.455 / 0.4250.355 / 0.3780.398 / 0.3750.276 / 0.3120.371 / 0.4040.236 / 0.259
w/o DCT directions0.418 / 0.4160.344 / 0.3740.365 / 0.3650.262 / 0.3060.355 / 0.4020.232 / 0.257
w/o forecast embedding0.417 / 0.4170.343 / 0.3730.364 / 0.3650.262 / 0.3070.353 / 0.4000.229 / 0.255
w/o history embedding0.421 / 0.4140.351 / 0.3760.365 / 0.3650.265 / 0.3070.363 / 0.4050.232 / 0.259
w/o local memory0.416 / 0.4170.346 / 0.3740.364 / 0.3640.265 / 0.3070.352 / 0.3970.231 / 0.257
w/o shared memory0.416 / 0.4110.346 / 0.3730.369 / 0.3660.264 / 0.3070.362 / 0.4050.230 / 0.257

2. Full encoding latency: We apologize for not making the timing protocol clear in Section 4.6. In Table 3, TS-CorM used a separate frozen Chronos-Bolt encoder to encode the history and the backbone forecast, with one forward pass for each. The end-to-end measurements include backbone inference and both encoder passes. TS-CorM did not reuse encoder activations from the forecasting backbone.

Because backbone inference accounted for most of the total latency, we also reported adapter-only latency to quantify the cost of adaptation. TS-CorM took 2.343 ms per query, compared with 2.610 ms for TS-Memory. The reported 10.23% reduction refers to adapter-only latency. End-to-end latency decreased from 87.490 ms to 87.223 ms, a 0.31% reduction. Section 4.6 compares offline inference for both methods because TS-Memory has no online update mechanism.

In addition, at the request of Reviewer VnTi's Q2, we measured inference latency with TimesFM-3 as backbone. The results are shown below for reference.

& \textbf{End-to-end inference} & \textbf{Adapter-only inference} \\\\ \hline \begin{array}{c} \text{Horizon} \\\\ 96 \\\\ 192 \\\\ 336 \\\\ 720 \\\\ \hline \textbf{Overall} \end{array} & \begin{array}{rrr} \text{TimesFM-3} & \text{+ TS-CorM} & \text{+ TS-Memory} \\\\ 114.215 & 120.155 & 119.591 \\\\ 116.512 & 120.093 & 121.672 \\\\ 127.752 & 133.440 & 134.162 \\\\ 130.154 & 135.282 & 136.267 \\\\ \hline \mathbf{122.158} & \mathbf{127.243} & \mathbf{127.923} \end{array} & \begin{array}{rr} \text{TS-CorM} & \text{TS-Memory} \\\\ 5.940 & 5.376 \\\\ 3.581 & 5.160 \\\\ 5.688 & 6.410 \\\\ 5.128 & 6.113 \\\\ \hline \mathbf{5.084} & \mathbf{5.765} \end{array} \end{array}$$ **3. Online time and memory costs:** In the Fast Memory configuration, we concatenate a 768-dimensional history embedding, a 768-dimensional forecast embedding, and an intercept term to form a key vector of dimension $�0�$. The correction value has dimension $�1�$. Each channel also maintains a Fast weight matrix $�2�$ with 46,110 coefficients. Its size scales linearly with the number of channels. We measured the incremental cost of Fast Memory on an H100. Across forecast horizons of 96, 192, 336, and 720 steps, Fast Memory averaged **0.047 ms per read** and **0.489 ms per write**. Each write occurs once the full forecast target is observed and includes residual encoding and the memory update. The read is inexpensive because Fast Memory **reuses the key and correction dictionary already computed by Slow Memory for the same forecast**. | Horizon | Mean read latency (ms) | Mean write latency (ms) | |---:|---:|---:| | 96 | 0.0476 | 0.4943 | | 192 | 0.0464 | 0.4842 | | 336 | 0.0480 | 0.4891 | | 720 | 0.0477 | 0.4898 |
AuthorsOpenReview ↗

To Reviewer VUFa

Dear Reviewer VUFa,

Thank you for your careful review and constructive feedback. We greatly appreciate your recognition of the value of the structured correction dictionary and your positive assessment of our systematic analysis and empirical results.

In our rebuttal, we added comparisons with simpler forecast-combination baselines and MEMTS, and evaluated Fast Memory on longer test sequences. We also addressed your questions about comparing Fast Memory with forgetting-factor RLS and sliding-window regression under matched feedback delays and controlled drift. Finally, we provided further details on the results reported across tables, end-to-end inference latency, and the computational and storage costs of online updates.

We would appreciate your feedback on whether our responses address your concerns. If you find them satisfactory, we would be grateful for your support. Please let us know if you have any further questions or need any clarification. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.