acceptodds
Under review as a conference paper at ICLR 2027

ECHO: Correcting Time Series Foundation Model Forecasts with Historical Analogues and Errors

Abstract

Time Series Foundation Models (TSFMs) provide general-purpose forecasting capabilities, yet their forecasts can remain misaligned with the temporal characteristics of a target domain. This motivates lightweight forecast correction even when only black-box access to the TSFM is available. Our key observation is that two forms of historical information provide complementary correction signals: historical analogues reveal how similar contexts evolved, while past forecast errors reveal a TSFM's characteristic deviations in the target domain. Based on this observation, we propose ECHO, a black-box forecast correction framework that learns from both sources. ECHO uses AnalogE to assess retrieved historical analogues together with the current context and base forecast, learning transferable corrections through cross-domain pretraining. In parallel, RecentE learns target-specific residual corrections from past forecast errors and refines them as causally available feedback accumulates. ECHO dynamically combines the two corrections according to their observed forecasting performance. Across 5 representative TSFMs and 9 real-world datasets, ECHO reduces MSE in all 45 backbone–dataset settings, yielding an average relative reduction of 9.26% in macro-average MSE. Its gains also extend to long-horizon forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Time Series Foundation Models, Retrieval-Augmented Forecasting, Online Learning, Black-Box Adaptation

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

ECHO is a black-box forecast-correction framework for frozen time-series foundation models. It combines two sources of historical evidence: AnalogE retrieves context–future analogues and learns a gated correction toward a historical template, while RecentE learns residual corrections from past forecasting errors and updates them as fully observed feedback becomes available. TempRA combines the two branches using their historical performance over multiple temporal views. The evaluation covers five foundation models and nine datasets, alongside long-horizon forecasting and component analyses. The principal strength is the consistency of the MSE improvements under a carefully specified evaluation protocol. The main questions concern the incremental value of the framework’s complexity and the scope of its benefits across error metrics.

Critical strengths
  • Broad and consistent empirical gains. ECHO improves MSE across all 45 backbone–dataset settings, supported by long-horizon and multi-seed evaluations.
  • Practical design and rigorous evaluation. Black-box access makes the framework broadly applicable, while fully matured feedback and matched-input baselines strengthen the credibility of the results.
Critical weaknesses
  • Complexity exceeds the demonstrated marginal gains. Dynamic aggregation and online updates contribute relatively small improvements, leaving it unclear whether a simpler correction pipeline could achieve most of the benefit.
  • Benefits are less consistent beyond MSE. MAE improvements are smaller and sometimes negative, while RecentE alone can degrade forecasts, limiting claims of uniformly reliable correction.
Questions for authors
  1. How much of the overall gain remains with fixed fusion and a plain residual regressor, without online updates or robust reweighting?
  2. Can the headline results present MAE alongside MSE and explicitly distinguish point-forecast improvements from probabilistic forecasting quality?
  3. Does TempRA reliably down-weight RecentE in settings where that branch underperforms the backbone?
  4. What are the complete backbone-specific pretraining and deployment costs, including upstream forecasts, retrieval encoding, and memory requirements?
  5. Can you compare full ECHO against strong end-to-end black-box adaptation baselines under matched inputs, tuning budgets, and causal feedback availability, rather than primarily comparing individual branches?
  6. Can you extend these competing-method comparisons across additional foundation-model backbones and datasets to determine whether ECHO’s advantage generalizes beyond the current comparison subset?
3 comments

Thank you for reading our paper carefully and for your constructive feedback. We address your questions below.

Q1. How much gain remains after simplification?

On TimesFM 2.5, fixed fusion with plain residual regression, without RecentE online updates or robust reweighting, produces different results across datasets.

On ETTh1, the simplified version has slightly lower MSE than ECHO (.342 vs .346). On Exchange, it retains only 33.4% of the MSE gain, while MAE rises from .164 to .170, exceeding the frozen backbone's .168. On PEMS08, it retains 91.9% of the MSE gain but only 66.2% of the MAE gain.

These results indicate that the dynamic model helps preserve gains on datasets sensitive to fusion weights, such as Exchange and PEMS08, where simplification can sacrifice a substantial portion of the benefit.

Q2. Point versus probabilistic forecasting

ECHO is designed to improve point forecasts. For quantile-based backbones such as Chronos-2, we use the median (q50) as the point forecast. Both MSE and MAE therefore measure point-forecast accuracy.

Q3. Does TempRA down-weight RecentE when it underperforms?

We define the recent loss difference as ΔL=LRecentE−Lbackbone\Delta L=L_{\mathrm{RecentE}}-L_{\mathrm{backbone}}. RecentE is classified as recently worse when ΔL>0\Delta L>0 and no worse when ΔL≤0\Delta L\leq0. The table reports its average weight in each state. Weights lie in [0,1], with lower values indicating smaller contributions to the final forecast.

DatasetRecently worse (ΔL>0\Delta L>0)No worse (ΔL≤0\Delta L\leq0)
ETTh1.2755.3847
ETTh2.3952.4794
ETTm1.2723.3548
ETTm2.2951.3709
Exchange.0828.1836
Weather.3235.5069
Electricity.4381.5604
PEMS08.5590.6136
Solar.0685.1546

Across all nine datasets, RecentE receives a lower average weight when its recent performance is worse.

Q4. Training and deployment costs

Overall, ECHO maintains a low computational overhead across both offline training and downstream deployment, with one-time backbone-specific training and efficient inference-time adaptation. We measured TimesFM 2.5 and Chronos-2 on the same H200. AnalogE is trained once per backbone on 50k samples for three epochs and reused across downstream datasets. The retrieval encoder and candidate bank can be shared across backbones.

Training uses batch size 32. We timed 30 steps after five warm-up steps, including upstream frozen-backbone forecasts, AnalogE forward/backward passes, and optimizer updates.

Offline costTimesFM 2.5Chronos-2
AnalogE training, 30 steps2.10 s1.65 s
ETTh1 Train forecasts + RecentE initialization38.01 s23.89 s

Deployment uses ETTh1, L=512, H=64, and K=20. Latency includes the backbone forecast, query encoding, exact retrieval, case loading, AnalogE, RecentE, and TempRA. Matured-feedback processing and refitting add approximately 0.15 ms per variate per origin, amortized.

Deployment costTimesFM 2.5Chronos-2
Single-variate latency with live retrieval47.50 ms31.63 ms
Backbone, retriever, and AnalogE weight storage1.237 GiB0.820 GiB

ETTh1 candidate vectors occupy 165.40 MiB, and retained RecentE training features occupy approximately 849.37 MiB. These total approximately 0.991 GiB, separate from model weight storage.

Q5. Complete-method comparisons

We compared complete ELF, ORCA, and ECHO on all nine datasets using Chronos-2 and TimesFM 2.5 (L=512, H=64). ECHO results come from the main evaluation. Macro is the unweighted nine-dataset mean. Wins counts first places using unrounded results.

BackboneMethodMacro MSE/MAEWins MSE/MAE
Chronos-2Frozen.200/.2420/2
Chronos-2ELF.190/.2420/2
Chronos-2ORCA.187/.2432/1
Chronos-2ECHO.181/.2397/4
TimesFM 2.5Frozen.185/.2440/1
TimesFM 2.5ELF.182/.2461/1
TimesFM 2.5ORCA.178/.2450/1
TimesFM 2.5ECHO.172/.2418/6

ECHO achieves the lowest macro MSE and MAE on both backbones, ranking first in MSE in 15 of 18 backbone–dataset settings and in MAE in 10 of 18.

Q6. Extension to additional backbones

We evaluated ECHO on Toto-2 across six datasets. Each cell reports MSE / MAE. The lower value for each metric is bolded based on the unrounded results.

DatasetFrozen Toto-2ECHO
ETTh1.371 / .367.346 / .367
ETTh2.254 / .297.237 / .301
ETTm1.311 / .322.270 / .316
ETTm2.153 / .227.140 / .225
Exchange Rate.060 / .168.057 / .165
Weather.144 / .163.126 / .162

ECHO reduces MSE on all six datasets and MAE on four. MAE increases slightly on ETTh1 and ETTh2.

We hope these additions address your concerns and would appreciate your reconsideration of the score. Thank you for your constructive feedback.

Thank you for this thorough rebuttal and the substantial additional experiments. The complete method comparison against ELF and ORCA on all nine datasets with two backbones was exactly what was missing, and ECHO's lowest macro MSE and MAE there meaningfully strengthens the framework-level claim. I also appreciate the concrete cost accounting and the Toto-2 extension, which answer the practical-deployment questions I raised.

But two concerns are not fully resolved by the new evidence. First, on complexity versus marginal benefit: your own simplification answer shows that on ETTh1 a fixed fusion with a plain ridge regressor is essentially as good as full ECHO (0.342 versus 0.346), so the elaborate dynamic aggregation and online-update machinery is only clearly justified on weight-sensitive datasets such as Exchange and PEMS08. This leaves my concern that a much simpler pipeline recovers most of the gain, and I would ask that the simplified baseline be reported as a full row in the main comparison rather than for a few datasets. Second, on the MSE–MAE asymmetry: the long-horizon macro MAE improvements are only -0.63% to -1.33%, and Solar's MAE worsens at every horizon. Clarifying that ECHO targets point forecasts does not address this; it suggests the gains are concentrated in suppressing large errors rather than uniformly improving point accuracy, which is an important limitation for a correction method and should be stated explicitly, ideally with MAE reported as a first-class headline metric.

On balance, the rebuttal resolves my missing-comparison and cost questions well, but my two central weaknesses stand. I maintain my rating of weak reject, and would only reconsider if you can further substantiate that the aggregation and online machinery are necessary beyond a few datasets and that the benefit is not specific to MSE. Thank you again for the careful and constructive response.

Thank you for your careful review and further feedback. We acknowledge that the benefits of dynamic aggregation and online updates are dataset-dependent, and the current evidence does not establish their necessity across all datasets. We also recognize that ECHO’s strengths lie primarily in reducing large forecast errors. These remain limitations of ECHO. In future work, we will further investigate when dynamic fusion can deliver consistent gains and how to achieve more consistent improvements in MAE.

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

ECHO corrects frozen time series foundation models by combining predictors using historically similar values and a historical feedback stream. This improves standalone forecast models in all evaluated cases and also outperforms retrieval and feedback models in isolation.

Critical strengths
  • Results show clear improvements over the backbones in all evaluated cases.
  • Extensive experiments for ablations and comparisons to models with retrieval and feedback streams are given.
Critical weaknesses
  • There is no comparison to a baseline which also performs retrieval and error correction from feedback. This is relevant to understand if the main contribution lies in the combination of both approaches or also in the single methods.
  • Some effects are rather small: According to Table 3(b), RecentE is outperformed in 2 out of 4 datasets, according to Table 10, the effect of TempRA is limited to less than 1% in MSE.
  • RecentE results compared to ELF and Ada are not consistently superior, as per Table 3(b) and 9. Table 9 misses comparisons to Electricity, PEMS08 and Solar.
  • The problem statement does not make very clear which additional information is provided w.r.t. traditional time series forecasting. It is unclear what HT\mathcal{H}^T exactly contains (is this the historical analog bank in Figure 1?), and the availability of observed outcomes (Equation (8)) is not mentioned at all.
  • Formal notation to compute Δi\Delta_i and ptbasep_t^{base} would be good.
  • The description of RecentE lacks intuition.
  • No code provided for review.

Minor comments:

  • Embed Figure 1 as PDF/SVG.
  • Briefly describe the retrieval strategy in the main paper (not just "see Appendix A.1")
  • Equation (5): It is not clear how a single value (xtlastx_t^{last}) is added to the vectors, I assume componentwise addition.
  • Highlight best performing methods in Table 9 in bold.
Questions for authors
  • Q1: Can you precisely define Δi\Delta_i and ptbasep_t^{base}?
  • Q2: Can you provide more intuition regarding the results of Figure 2 and 3 (e.g., why PEMS08 does not benefit from AnalogE)?
  • Q3: Can you give extended results for Table 9 for Electricity, PEMS08 and Solar?
2 comments

Thank you for the careful reading and suggestions. We address your questions and other concerns below.

Q1. Clarification of notation

Let xilastx_i^{\mathrm{last}} and xtlastx_t^{\mathrm{last}} be the last observations of historical context XiX_i and current context XtX_t. We define

p_t^{\mathrm{base}}=Y_t^{\mathrm{base}}-x_t^{\mathrm{last}}\mathbf{1}_H.$$ Here $\mathbf{1}_H\in\mathbb{R}^H$ is the all-ones vector. Both transitions are relative to their respective context endpoints. In Eq. (5), adding $x_t^{\mathrm{last}}$ is componentwise, i.e., adding $x_t^{\mathrm{last}}\mathbf{1}_H$. Computations use normalized quantities, and outputs are restored to the original scale. We will clarify these definitions. **Q2. Intuition for Figures 2 and 3** We added nine-dataset diagnostics on TimesFM 2.5. AnalogE uses historical futures as references, while RecentE learns target-domain backbone residuals. They outperform each other on five and four datasets, respectively. |Dataset|AnalogE MSE reduction|RecentE|ECHO| |---|---|---|---| |Solar|13.14%|10.59%|13.74%| |Exchange|5.06%|−18.37%|5.54%| |Weather|3.28%|9.65%|8.60%| |PEMS08|−6.60%|7.73%|8.48%| AnalogE benefits when historical futures provide useful references, as on Solar and Exchange. On PEMS08, context similarity may provide limited support for future prediction. With the same Top-20 candidates, uniform and distance-weighted templates, using validation-selected correction strengths, each improve test MSE by only 1.27%, consistent with this explanation. RecentE mainly corrects forecast-level bias. On Weather and PEMS08, it reduces MSE by 9.65% and 7.73%, with 74% and 85% of the net gains coming from reduced window-level bias. ECHO improves MSE on all nine datasets and beats both branches on seven, supporting their complementary roles. **Q3. Extended Table 9** We extended Table 9 to nine datasets. Below are the requested additions on TimesFM 2.5 (L=512, H=64), without AnalogE or TempRA. Cells report MSE/MAE. Minima are bolded, including ties at displayed precision. |Method|Electricity|PEMS08|Solar| |---|---|---|---| |ELF|.119/.213|.097/.199|.206/**.217**| |ELF + Base Input|.119/**.212**|.096/.198|.206/**.217**| |Ada-Y|.121/.214|.096/.195|.199/**.217**| |Ada-Y + Context|.121/.214|.093/**.190**|.197/.218| |RecentE|**.118**/.214|**.089**/.193|**.186**/.246| RecentE achieves the lowest MSE on six of nine datasets, including all three additions. **Other concerns** **Combined baselines and module contributions.** Combining retrieval and feedback-based correction is a central ECHO design. To assess both the framework and its modules, we added complete-method comparisons with ELF/ORCA and branch replacements retaining the other branch and TempRA. The nine-dataset summary below reports unweighted macro MSE/MAE and first-place counts based on unrounded values. |Backbone|Method|Macro MSE/MAE|Wins MSE/MAE| |---|---|---|---| |Chronos-2|Frozen|.200/.242|0/2| |Chronos-2|ELF|.190/.242|0/2| |Chronos-2|ORCA|.187/.243|2/1| |Chronos-2|ECHO|**.181/.239**|**7/4**| |TimesFM 2.5|Frozen|.185/.244|0/1| |TimesFM 2.5|ELF|.182/.246|1/1| |TimesFM 2.5|ORCA|.178/.245|0/1| |TimesFM 2.5|ECHO|**.172/.241**|**8/6**| ECHO has the lowest macro MSE and MAE on both backbones. Branch replacements on TimesFM 2.5 yield the following six-dataset macro results. |Replaced branch|Expert|Macro MSE/MAE| |---|---|---| |AnalogE|AnalogE|**.195**/.258| |AnalogE|BB-TS-RAG|.198/.260| |AnalogE|Naive RAF|.196/**.257**| |RecentE|RecentE|**.195/.258**| |RecentE|ELF|.198/.260| |RecentE|ELF + Base Input|.198/.260| |RecentE|Ada-Y|.200/.260| |RecentE|Ada-Y + Context|.200/.260| RecentE is not uniformly best alone, but retaining it gives the lowest macro MSE/MAE here. AnalogE gives the lowest macro MSE among historical replacements. **TempRA.** Against validation-selected fixed fusion, six-dataset macro MSE drops by 0.81%/0.84% and MAE by 1.48%/2.78% on TimesFM 2.5/Chronos-2. Gains are larger on Exchange: |Backbone|Fixed MSE/MAE|TempRA MSE/MAE| |---|---|---| |TimesFM 2.5|.064/.179|**.056/.164**| |Chronos-2|.084/.183|**.058/.166**| This supports feedback-based weighting in weight-sensitive settings such as Exchange. **Historical information.** $\mathcal H_t$ includes the two gray components in Figure 1: the fixed training-set context–future bank and past backbone forecasts with observed outcomes. Feedback from origin $j$ is available only when $j+H\leq t$. We will clarify this in the problem statement. **Code.** Code was included in the supplementary materials at submission. We will clarify its location. We hope these additions address your concerns. Thank you again for your constructive feedback.

I appreciate the focus on my questions and the conciseness of the reply and the focused revisions. Specifically due to the new results in Table 2 and since code was indeed delivered, I will increase my score.

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

The paper looks at correcting the forecasts of a frozen time series foundation model (TSFM) when only its inputs and outputs are accessible. It argues that two kinds of history give complementary signals: past contexts that resemble the current one (with their observed futures), and the backbone's own past forecast errors in the target domain. The proposed method ECHO has three parts: First, the AnalogE (pretrained on 50K Chronos corpus examples, about 42K parameters, frozen downstream) scores K=20 retrieved cases and builds a transition template. A bounded gate moves the base forecast toward that template. Second, the RecentE fits a complex linear map from the spectra of the context and base forecast to the residual spectrum. It uses Huber-weighted ridge regression and refits periodically on fully observed feedback. Finally, the TempRA blends the two corrected forecasts using weights from long, short and recent views of past errors. The method is evaluated on a large benchmark (ETT, Exchange, Weather, Electricity, PEMS08, Solar) with 5 backbones (TimesFM2.5, Chronos-2, Chronos-Bolt, Moirai 2.0, Sundial). The authors claimed that ECHO lowers MSE in all 45 backbone and dataset pairs, with a mean 9.26% reduction in macro MSE.

Critical strengths
  1. The gains are consistent, for instance, the MSE drops for every backbone and dataset (Table 1), and this holds across three seeds (Table 11). Additionally, the retrieval only uses the training split, and online updates only use errors whose full horizon has been observed. So the "plug-in, black-box" claim is a strong factor, and the extra cost is small (about 5.6 ms per query).

  2. The ablations back up the main idea, which the Table 7 shows that AnalogE is better on some datasets and RecentE on others, and Fig. 4 shows both work better as corrections than as standalone predictors, showing that this is a good support for combining the two sources.

Critical weaknesses

1.The full ECHO is never compared to another full correction method. So, the baselines in Table 3 are branch-level only, on one backbone and four to six datasets. Even there, ELF alone is close to or better than full ECHO on ETTh1 (0.342 vs 0.346). ORCA, which the paper says uses the same inputs as RecentE, is not compared at all. As it stands, it is unclear whether the gains come from the new modules or just from combining any two reasonable correctors.

  1. The MSE gains mostly do not show up in MAE, in the long horizons, MSE improves by 7 to 8.5% while MAE improves by only about 1% (Table 6). On Solar, MAE actually gets worse at every horizon even though MSE drops a lot. Several ETT settings also have worse MAE in Table 5. This looks more like diminishing large errors than making the forecasts generally better, and the paper does not discuss it.
Questions for authors
  1. How does full ECHO compare with ELF, and ORCA as complete methods on all datasets and at least two backbones? And what happens if RecentE is swapped for ELF, or AnalogE for Naive RAF?
  2. Why does MAE barely change, or get worse, when MSE improves a lot (e.g. Solar in Table 6)?
  3. Does the Chronos pretraining data used for AnalogE overlap with Electricity, Solar, or Weather, especially their test periods?
  4. ECHO uses the training and validation splits, but the backbone does not. How does ECHO compare with a backbone that is lightly fine-tuned on the same data?
1 comment

Thank you for the careful reading and concrete suggestions. We appreciate your recognition of the consistent MSE gains and the practical value of black-box correction. We address your questions with additional experiments below.

Q1. Complete-method comparisons and component replacements

We compared ECHO with complete ELF and ORCA on all nine datasets using Chronos-2 and TimesFM 2.5 (L=512, H=64). ECHO results come from the main evaluation.

Table 1. Nine-dataset summary. Macro is the unweighted mean. Entries report MSE/MAE; Wins counts first places across all four methods using unrounded results.

BackboneMethodMacro MSE/MAEWins MSE/MAE
Chronos-2Frozen.200/.2420/2
Chronos-2ELF.190/.2420/2
Chronos-2ORCA.187/.2432/1
Chronos-2ECHO.181/.2397/4
TimesFM 2.5Frozen.185/.2440/1
TimesFM 2.5ELF.182/.2461/1
TimesFM 2.5ORCA.178/.2450/1
TimesFM 2.5ECHO.172/.2418/6

ECHO has the lowest macro MSE and MAE on both backbones, winning 15/18 settings in MSE and 10/18 in MAE. ELF remains better on TimesFM ETTh1.

We also replaced individual branches on six datasets with TimesFM 2.5, retaining the other branch and TempRA.

Table 2. Component replacements, six-dataset macro MSE/MAE. Bold indicates the best metric within each replacement group, based on unrounded values.

Replaced branchExpert usedMacro MSE/MAE
AnalogEAnalogE (ECHO).195/.258
AnalogEBB-TS-RAG.198/.260
AnalogENaive RAF.196/.257
RecentERecentE (ECHO).195/.258
RecentEELF.198/.260
RecentEELF + Base Input.198/.260
RecentEAda-Y.200/.260
RecentEAda-Y + Context.200/.260

AnalogE gives the lowest macro MSE, while Naive RAF gives the lowest macro MAE. RecentE gives the lowest macro MSE and MAE among its replacements and beats ELF on both metrics in five of six datasets.

Q2. Why are MAE gains smaller or negative?

We analyzed TimesFM 2.5 on ETTh1, ETTm1, and Weather (L=512, H=64). For each dataset, we used the validation backbone absolute-error P90 as a threshold. Test coordinates meeting or exceeding it form a common high-error subset for all methods. This is post-hoc analysis only.

Table 3. High-error subset: MSE/MAE reductions (%) relative to Frozen. Share is the percentage of test coordinates in this subset.

DatasetShareAnalogERecentEECHO
ETTh16.38%4.78/2.1813.52/8.818.45/4.28
ETTm17.80%10.26/5.5911.68/7.4711.13/5.98
Weather5.48%6.02/5.1614.52/16.9411.32/11.36

Both branches reduce large forecast deviations. Gain decomposition shows that these subsets contribute the main squared-error reductions. Squared loss gives greater weight to large deviations, consistent with the larger overall MSE gains. On ETTh1, absolute-error reductions in this subset are outweighed by increases elsewhere, yielding slightly worse overall MAE. Overall MAE thus depends on the balance across error regions.

Q3. Pretraining data overlap

We conducted a full-window audit of the 50,000 training samples, model-selection samples, and retrieval candidates actually used by AnalogE. Allowing for rescaling and shifting, we found no matches to contiguous windows within the test periods of Electricity, Solar, or Weather.

Q4. Comparison with light fine-tuning

We compared ECHO with per-domain LoRA on TimesFM 2.5 . LoRA uses Train for updates and Validation for checkpoint selection, then remains fixed on Test.

Table 4. Test MSE/MAE. Bold marks minima using unrounded values, including ties.

DatasetFrozenLoRAECHO
ETTh1.355/.369.355/.369.346/.371
ETTh2.257/.315.265/.320.242/.310
ETTm1.280/.316.265/.311.263/.314
ETTm2.148/.232.137/.222.136/.225
Exchange.059/.168.059/.168.056/.164
Weather.135/.171.121/.159.124/.164
Wins MSE/MAE0/11/45/2

ECHO wins MSE on five of six datasets; LoRA wins MAE on four, including a tie. LoRA requires white-box access and trains 2.77M parameters, versus approximately 42K in AnalogE.

We hope these additions address your concerns and would appreciate your reconsideration of the score. Thank you for your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.