acceptodds
Under review as a conference paper at ICLR 2027

Diversified Scaling Inference in Time Series Foundation Models

12 views

Abstract

The advancement of Time Series Foundation Models (TSFMs) has been driven primarily by large-scale pre-training, but inference-time compute potential remains largely untapped. This work systematically investigates two questions: how do TSFMs behave under standard sampling-based inference scaling, and can controlled sampling diversity enhance performance? We first examine the properties of TSFMs under standard sampling, revealing that they often fail to adhere to scaling laws due to insufficient exploration of the solution space. Building on this, we then delve into diversified inference scaling via tailored time series perturbations to expand the generative distribution's support. We theoretically analyze the diversity-fidelity trade-off and derive a critical sample threshold for diversified sampling to outperform standard sampling. Extensive experiments across various TSFMs and datasets show that proper diversified inference scaling yields substantial performance gains without model parameter updates, establishing inference scaling as a critical, highly parallelizable paradigm for TSFM optimization. As an application, we propose RobustMSE, a rigorous metric to quantify TSFM headroom under a fixed budget. Overall, our findings clarify these factor interactions, enabling reliable performance via diverse large-scale time series inference in parallel environments without re-training TSFMs.

Primary area:
learning on time series and dynamical systems
Keywords:
time series foundation models, inference-time scaling, diversified sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_arKvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies inference-time sampling and input diversification in frozen time-series foundation models, reporting useful empirical gains but leaving gaps between oracle analysis, practical inference, and the proposed evaluation metric.

Critical strengths
  1. The paper systematically examines how sampling budget, context length, model size, and input perturbations interact across several TSFM architectures.
  2. Input diversification offers a simple way to improve frozen models without retraining or changing their architecture.
  3. Reporting both successful and harmful perturbations provides useful evidence that increased diversity alone does not guarantee better forecasts.
Critical weaknesses
  1. The theory concerns oracle best-of-N selection using future targets, so it does not establish the benefits of practical median aggregation or label-free strategy selection.
  2. The paper omits SPRINT (Xu et al., 2026), leaving its positioning and comparative value against existing training-free inference enhancement insufficiently established.
  3. RobustMSE evaluates a configured inference pipeline rather than an intrinsic model property, while configuration-search costs and unequal candidate-generation costs complicate its claimed fairness.

Xu, L., et al. (2026). See More, Forecast Better and Faster: Enhancing Time Series Foundation Models via Inference-Time Plug-and-Play Downsampling. ICML 2026.

Questions for authors
  1. How does the method compare with SPRINT (Xu et al., 2026) under matched historical information and end-to-end compute budgets?
  2. Which reported gains remain when perturbation selection uses only validation data and forecasting requires no access to future targets?
  3. Can the theory explain median-aggregation gains when the original predictive distribution already has full support?
  4. How are configurations selected for RobustMSE, and does its budget include configuration search and auxiliary reconstruction calls?
  5. How much improvement comes from input diversification rather than ordinary multi-sample aggregation when sampling budgets and aggregation rules are matched?
5 comments
AuthorsOpenReview ↗

Response to Reviewer arKv

Dear Reviewer arKv,

We sincerely appreciate your careful assessment and recognition of our systematic empirical study. Please find our point-by-point responses below.

W1 & Q2. Distinguishing Oracle Analysis from Label-Free Inference

Response. The manuscript explicitly separates oracle analysis from deployable aggregation. Section 4.1 scopes the theory to best-of-NN (lines 238–241), while Section 3 distinguishes oracle EM from inference-time median aggregation (lines 145–159). The theory characterizes candidate-generation headroom; practical MV benefits are evaluated directly.

The relevant gains are the MV results reported throughout the paper, rather than only the oracle EM results. (Q2) For example, Table 11 (lines 1422–1426) fixes both N=128N=128 and MV aggregation. Gaussian diversification reduces Chronos-T5-Tiny’s MSE from 0.1564 to 0.1406 (10.1%), and improves Chronos-T5-Large by 7.0%. These gains require no oracle output selection. Moreover, these are individually reported perturbation results, not merely oracle-selected winners. Our practical claim concerns benefits from specified diversification policies; it does not require solving optimal label-free strategy selection. The default noise policy and optional oracle selector are explicitly distinguished in lines 426–430. Our practical claim is that specified diversification policies can improve median forecasts; it does not require a theorem guaranteeing label-free identification of the optimal policy.

W2 & Q1. Positioning and Comparison with SPRINT

Response. We have conducted additional comparisons with SPRINT and will incorporate its positioning and the new results into the manuscript. (Section 5.3) SPRINT improves inference through trend downsampling, seasonal pattern replication, and forecast reconstruction. Our method instead diversifies the observed context and aggregates predictions from the frozen backbone. Both are training-free, but they modify different components of inference.

Fair comparison. The primary comparison uses the same 512-point contexts, forecast origins, normalization, backbone checkpoints, prediction horizon H=96, and sliding-window stride 32. SPRINT is reimplemented within our evaluation framework. We additionally evaluate valid-edge decomposition, Moirai patch size 16, and a 720-point SPRINT look-back; the latter is a separate sensitivity analysis that gives SPRINT more history.

The new comparisons establish substantial practical value beyond the conceptual distinction. Across Chronos-Base, Moirai-Base, and TimesFM, DSI achieves lower error in all 36 comparisons. This remains true against the best tested SPRINT variants. Representative MV results are (see more results in Section 5.3 of the revised manuscript):

Dataset / backboneDSISPRINTBest tested SPRINT variant
Traffic / TimesFM0.04390.11420.0916
Traffic / Moirai-Base0.04310.18440.1414
Electricity / Chronos-Base0.20880.40120.3195

SPRINT requires 4.8–13.2 times fewer inference FLOPs per pass, so equal sample counts are not treated as equal compute. We calculate its affordable sample allocation from the measured FLOPs ratio: Chronos-Base's SPRINT MV errors change little from 16 to 128 samples: 0.1493→0.1477 on ETTh1, 0.1480→0.1484 on Traffic, and 0.4041→0.4014 on Electricity. We also evaluated the cost of the perturbed input of our method in Table 12. The conclusion is that all the cost caused by these perturbed inputs only adds a modest overhead compared to the prediction time.

W3 & Q4. RobustMSE Configurations and Computational Fairness

Response. Evaluating a configured inference pipeline is the intended scope of RobustMSE. The manuscript explicitly states that it measures performance under evaluated inference strategies rather than an intrinsic, strategy-independent model property. Its rankings should therefore be interpreted as protocol-conditioned comparisons. Thus, it can reflect the headroom performance of models under the same protocol.

RobustMSE uses the perturbation settings specified in Table 3 and the empirically screened strategy set described in Section 5.1, which excludes Suffix and Missing perturbations. Task-Agnostic and Task-Specific strategies are evaluated separately. Within each category, the evaluation samples 64 forecasts from the corresponding candidate pool and repeats the aggregation over five trials. Configuration search and auxiliary reconstruction calls are additional costs, but using RobustMSE is fair under the same configuration. Our profiling explicitly measures this difference: Task-Agnostic perturbations add 0.092–0.491 ms/window, whereas reconstruction adds 149.8 ms/window in the current implementation. Equal candidate counts should therefore be interpreted as a controlled sampling protocol, rather than identical computational expenditure across perturbation types.

Q3. MV Gains under Full Predictive Support

Response. Thank you for raising this point. Full support does not preclude median-aggregation gains: support describes what is possible, whereas the median depends on how probability mass is distributed. The mixture in Eq. (4) can redistribute probability without enlarging support. Section 5.1 explicitly discusses increasing probability mass near the target (lines 346–359).

Let mstdm_{\mathrm{std}} and mdivm_{\mathrm{div}} denote the coordinate-wise population medians. Under i.i.d. sampling and unique, regular medians, the sample medians converge to these respective values. Consequently, if L(mdiv,Y)<L(mstd,Y),\mathcal L(m_{\mathrm{div}},Y)<\mathcal L(m_{\mathrm{std}},Y), diversified sampling achieves lower asymptotic median-forecast loss even when both distributions have identical full support. This provides a conditional explanation through probability redistribution.

Q5. Marginal Gains from Input Diversification

Table 11 has already demonstrated this contribution by fixing both the sampling budget (N=128) and the aggregation rule (MV). The “None” baseline already performs ordinary multi-sample aggregation; Gaussian diversification changes the inputs while retaining the same number of forecasts and median aggregator. They show meaningful additional gains in several settings while transparently showing that diversification is not uniformly beneficial.

Chronos-T5Standard sampling + MVGaussian diversification + MVAdditional MSE reduction
Tiny0.15640.140610.1%
Mini0.15230.1548−1.6%
Small0.14450.14410.3%
Base0.14530.14311.5%
Large0.14810.13777.0%

To further address your concerns, We have added a standard-sampling baseline at N=64N=64 (Std@64) alongside the diversified inference results in the main RobustMSE table (Table 3). Std@64 applies MV aggregation to 64 forecasts generated from the unperturbed input, whereas MV@64 applies the same aggregation rule to 64 forecasts generated with input diversification. This comparison fixes both the candidate budget and aggregation rule, directly measuring the marginal benefit of diversification beyond standard multi-sample inference. For a concise comparison, the table below summarizes the dataset-averaged MSE across ETTh1, Traffic, and Electricity. Full dataset-level results and standard deviations are provided in Table 3. The conclusion is that diversified MV achieves lower dataset-averaged MSE in seven of the eight model–strategy comparisons.

ModelStrategyMV@64Std@64
Chronos-T5-BaseTask-Agnostic0.13260.1382
Chronos-T5-BaseTask-Specific0.13620.1382
Moirai-1.1-R-BaseTask-Agnostic0.12640.1272
Moirai-1.1-R-BaseTask-Specific0.12530.1272
TimesFM-2.5-200MTask-Agnostic0.12150.1246
TimesFM-2.5-200MTask-Specific0.12470.1246
TimeMoE-200MTask-Agnostic0.11300.1156
TimeMoE-200MTask-Specific0.11310.1156

Dear Reviewer arKv,

Thank you for the time and effort you devoted to reviewing our paper. We would greatly appreciate your feedback on our responses and additional experiments, and we would be happy to clarify any remaining concerns before the discussion period closes.

Best regards,

The Authors

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

The paper studies inference-time compute scaling for fixed (non-retrained) TSFMs. It first varies model size, context length, and temperature, finding TSFMs do not follow clean log-linear scaling laws. It then proposes "diversified sampling": drawing N candidates from perturbed inputs (9 task-agnostic/specific perturbation families) and aggregating via Exact-Match best-of-N (oracle) or Majority Voting (median). Two theoretical results are given — an asymptotic support-expansion lower bound (Prop 4.2) and a finite-sample critical threshold N* (Prop 4.3) — plus RobustMSE, a metric reporting a model's best achievable MSE under N=64 diversified samples. Validation: four TSFMs (Chronos, TimesFM, Moirai, Time-MoE), four–six datasets; claims peak improvements "up to ~50%" without parameter updates.

Critical strengths
  1. The empirical sweep is genuinely comprehensive (models 8M–1.2B, context 32–1024, 9 perturbation types, 2 aggregators, 4–6 datasets, 4 architectures) and yields concrete, deployable guidance, honestly accompanied by a failure analysis (Table 1, flagging Suffix/Missing as harmful) and a latency report (Appendix I, ~22% latency increase at 32× compute). This is consequential because such a practitioner-oriented study with honest failure reporting saves real engineering effort and is the paper's primary durable contribution regardless of the headline framing.
  2. The fidelity–diversity characterization (Figure 3) — a non-monotonic relation between perturbation relevance and MSE peaking around cosine 0.95–1.00, plus the finding that diversified sampling helps most where models are weak / OOD — is a clean empirical finding that explains when the method pays off. It is consequential because it provides the actionable "operating band" that distinguishes this work from a naïve best-of-N recipe.
Critical weaknesses
  1. Model comparisons are confounded by pre-training data leakage: TimesFM was pretrained on Electricity and Traffic (acknowledged in Appendix E.1), yet it is reported as the best model ("lowest RobustMSE in 5/6 scenarios, as it was pre-trained on some datasets"). The "1st Count" row thus measures pre-training overlap, not model quality; and RobustMSE is explicitly acknowledged (Limitations) to be strategy-dependent, conflating model quality with perturbation-tuning effort. This is consequential because a core use case of the paper — benchmarking TSFMs with RobustMSE — inherits this confound and cannot be trusted as written.
Questions for authors
  1. Please re-report headline gains as marginal improvements over standard sampling at the same N (e.g., add a standard-sampling at N=64 baseline alongside RobustMSE in the main table) and move Exact-Match to a clearly-labeled oracle/upper-bound reference with Majority Voting foregrounded. What are the matched-N MV gains?
  2. For TimesFM, please segregate in-distribution (Electricity/Traffic, seen in pretraining) vs. out-of-distribution results. Does TimesFM remain best on non-leaked datasets?
1 comment
AuthorsOpenReview ↗

Response to Reviewer edko

Dear Reviewer edko,

We sincerely appreciate your careful assessment of the model comparisons and the interpretation of RobustMSE. We address them point by point below.

W1 & Q2. Potential pre-training overlap and the interpretation of RobustMSE

Response. We acknowledge that TimesFM's pre-training on Electricity and Traffic may affect its results on these two datasets. We explicitly disclosed the overlap (line 1128). We therefore conducted an additional analysis excluding TimesFM's results on Electricity and Traffic from the model comparison. The recalculated first-place counts remain 2/2/2 under EM, MV, and MSE, respectively. This indicates that the overall comparison is not solely driven by TimesFM's pre-training overlap. We will report the exclusion analysis and clarify the limitation in the revised manuscript.

Regarding RobustMSE, pretraining overlap affects ordinary MSE comparisons as well and is not introduced by RobustMSE. As defined in the paper (Lines 464–481), RobustMSE reflects a TSFM's repeatable best achievable performance under a fixed inference budget NN, an aggregator (EM or MV), and a set of validated diversified sampling configurations. It is therefore an inference-time performance measure that captures how effectively a model can exploit diversified sampling, rather than a replacement for conventional forecasting metrics. Conventional MSE evaluates standard forecasting performance, whereas RobustMSE characterizes the additional performance that may be achieved through inference-time diversification. These metrics answer complementary questions and should be interpreted together.

Q1. Matched-budget gains and the role of EM versus MV

Response. Thank you for this suggestion. We have added a standard-sampling baseline at N=64N=64 (Std@64) alongside the diversified inference results in the main RobustMSE table (Table 3). Std@64 applies MV aggregation to 64 forecasts generated from the unperturbed input, whereas MV@64 applies the same aggregation rule to 64 forecasts generated with input diversification. This comparison fixes both the candidate budget and aggregation rule, directly measuring the marginal benefit of diversification beyond standard multi-sample inference.

For a concise comparison, the table below summarizes the dataset-averaged MSE across ETTh1, Traffic, and Electricity. Full dataset-level results and standard deviations are provided in Table 3.

ModelStrategyMV@64Std@64
Chronos-T5-BaseTask-Agnostic0.13260.1382
Chronos-T5-BaseTask-Specific0.13620.1382
Moirai-1.1-R-BaseTask-Agnostic0.12640.1272
Moirai-1.1-R-BaseTask-Specific0.12530.1272
TimesFM-2.5-200MTask-Agnostic0.12150.1246
TimesFM-2.5-200MTask-Specific0.12470.1246
TimeMoE-200MTask-Agnostic0.11300.1156
TimeMoE-200MTask-Specific0.11310.1156

Diversified MV achieves lower dataset-averaged MSE in seven of the eight model–strategy comparisons. For example, task-agnostic diversification reduces Chronos-Base’s average MSE from 0.1382 to 0.1326, while task-specific diversification reduces Moirai’s from 0.1272 to 0.1253. At the individual dataset level, MV@64 improves over Std@64 in 19 of the 24 configurations, with two ties and three small degradations. These results demonstrate incremental benefits beyond standard sampling, while also showing that the gains depend on the model, dataset, and perturbation strategy.

We focus on MV for these matched-budget comparisons and treat EM solely as an oracle best-of-NN reference. MV aggregates candidates without consulting the ground-truth future, whereas EM selects the candidate with the lowest ground-truth error. The improvements reported above therefore do not rely on oracle candidate selection or comparisons against single-pass inference.

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

Inference-time scaling remains underexplored for time series foundation models. This paper first examines whether standard sampling at inference time improves forecasting, and finds that its gains quickly saturate because the sampling space is too narrow. The authors then develop diversified inference scaling, which adds varied perturbations to the input context to broaden sampling, explain why it works with a theoretical analysis, and propose RobustMSE, a new metric for comparing TSFMs’ headroom under a fixed inference budget. Core experiments cover four representative TSFMs on widely used long-term forecasting benchmarks and analyze how model size, context length, temperature, perturbation type, and aggregation strategy affect performance. Overall, inference-time scaling is a promising direction for TSFMs, and this paper offers a valuable early exploration of it.

Critical strengths
  1. Inference-time compute has been well studied for LLMs but rarely for TSFMs. The paper presents an early exploration in this direction, reporting not only the performance improvements observed under the proposed method but also the limitations of TSFMs and negative results on method design.
  2. The paper offers some helpful insights for practical usage, like the distinction between probabilistic and deterministic models is a sensible framing that helps explain the different scaling behaviors across model families, and the observations on sample budget and perturbation strength can guide practical configuration.
Critical weaknesses
  1. Proposition 4.2 states something almost self-evident. It discusses that if perturbations widen the range of possible forecasts, the best forecast in that wider range is at least as good as before, which is true for any added noise and only matters when the ground truth is used to pick the best candidate. The following Proposition 4.3 is also built on an overly simple model, and the resulting critical sample size is only a rough illustration and cannot guide how many samples to use in practice. Importantly, this section only analyzes EM, which requires the ground truth and is therefore impossible to use at forecasting time, while it provides no theoretical support for MV, the aggregator that could actually be deployed.

  2. Some standard deviations in Table 1 are as large as the means, and many of the differences in Tables 2 and 10 are small, while Table 10 reports no standard deviations. A significance tests for the comparisons in these tables would be more convincing.

  3. Form Figure 3, the best results occur when the perturbed input is nearly identical to the original, suggesting that the gains may not come from exploring new sample space as the paper claims. Moreover, no perturbation works reliably across models and datasets; if perturbations must be selected case by case, the method can hardly provide a reliable benefit.

  4. The perturbation choice lacks a clear motivation and guidence, and their extra computational cost is not clear in the inference time.

Questions for authors
  1. How are the perturbations configured in experiments, like Table 1 and Figure 4 ?
  2. How the time series with different cosine-similarity levels in Figure 3 are generated ?
2 comments
AuthorsOpenReview ↗

Response to Reviewer CNZd

Dear Reviewer CNZd,

We sincerely appreciate your careful reading and constructive comments. Please find our point-by-point responses below.

W1. Theoretical guarantees and practical aggregation

Response. Our theory is designed to analyze oracle headroom, and practical inference is evaluated separately through MV and EM. Section 4.1 explicitly scopes the analysis to best-of-NN and states that MV can behave differently or even degrade. We therefore do not infer an MV guarantee from the EM propositions. Proposition 4.3 addresses what support expansion alone cannot establish: whether useful candidates occur frequently enough to outweigh harmful perturbations under a finite budget (lines 283–297). Thus, the analysis does not imply that arbitrary noise improves forecasting. Practical budget recommendations additionally rely on observed MV convergence, but a good upper bound can provide us with a reference for practice.

We also clarified that EM is an oracle evaluator rather than a deployable forecasting-time aggregator (line 147), since it selects candidates using ground-truth errors. Our practical claims do not rely on EM alone. Crucially, Section 5 evaluates both EM and MV under matched sampling budgets across multiple TSFMs and datasets. MV aggregates the generated candidates without accessing the ground truth; for example, Gaussian improves Chronos-T5-Tiny from 0.1564 to 0.1406 (10.1%). Its empirical improvements over standard inference provide our evidence for deployable diversified inference.

W2. Variability of tables

Response. Table 1 explicitly includes unstable and failed perturbations (lines 270); their large variance is part of the phenomenon being investigated, rather than a contradiction of our findings. And we clarify that the results in Tables 1, 2, and 10 are all averaged over five runs. Our central finding also includes substantial gains beyond small numerical differences: under the same N=128N=128 and MV aggregation, Gaussian diversification improves Chronos-T5-Tiny by 10.1% and Chronos-T5-Large by 7.0% (Table 11, lines 1422–1426).

W3. About Figure 3

Response. We respectfully disagree with this reading of Figure 3. Across the three datasets, the minimum MSE generally lies within the cosine-similarity range of 0.95–1.00, rather than necessarily at 1.00. The relationship between input similarity and forecasting performance is non-monotonic: moderate perturbations can improve performance over the unperturbed input, whereas overly aggressive perturbations may introduce harmful noise. This behavior is consistent with the fidelity–diversity trade-off summarized in the takeaway (lines 360–363).

Moreover, the absence of a universally optimal perturbation does not imply that every forecast requires case-by-case selection. We explicitly recommend noise-based perturbations as a default, with the selector being optional (lines 428–430). The paper also explicitly excludes a claim of universal improvement across every model–dataset pair (lines 1615–1632). Model-dependent effectiveness is a finding of this systematic study, not a contradiction of its contribution. Our goal is to systematically investigate how controlled input diversity affects TSFM inference, rather than to identify one universally optimal perturbation for every model–dataset pair.

W4. Guidance and cost of perturbations

Response. The motivation, configuration guidance, and latency measurements are already provided in the manuscript. Appendix C explains the design rationale: task-agnostic perturbations probe sensitivity to input diversity, while task-specific perturbations exploit temporal structure to preserve fidelity. Table 3 specifies their definitions and default settings. The Actionable Insights (lines 416-431) further recommend noise-based defaults, high-similarity perturbations, and moderate sampling budgets.

Computational cost is explicitly discussed in Appendix I: total computation increases with NN, while parallel execution reduces its impact on latency. Table 13 provides end-to-end measurements: for Chronos-Tiny on ETTh1, N=16N=16 reduces MSE from 0.1949 to 0.1544 (20.8%), while latency increases from 326.66 to 353.82 seconds (8.3%). Thus, the reported benefits are accompanied by measured inference-time trade-offs. Moreover, we additionally report perturbation-generation overhead in Table 12. The conclusion is that the cost caused by these perturbed inputs only adds a modest overhead compared to the prediction time.

Q1. Experimental perturbation configurations

Response. The perturbation definitions and parameter settings were provided in Table 3, Appendix C (lines~883--898). For example, Gaussian noise uses standard deviation ησX\eta\sigma_X, with η∈{0.05,0.1,0.2}\eta\in\{0.05,0.1,0.2\} and prefix padding use lengths sampled from {16,32}\{16,32\} and padding values sampled from U(−1,1)U(-1,1). Task-sensitivity and task-dependency perturbations use an intensity sampled from U(0.01,0.05)U(0.01,0.05), with their procedures detailed in Algorithms~1--2.

Q2. Cosine-similarity

Response. The cosine similarity in Figure 3 measures the similarity between the perturbed input X′X' and the original input XX. Specifically, we generate perturbed inputs by varying the perturbation strength and calculate their cosine similarity to the corresponding original inputs: sim⁡(X′,X)=X′⊤X∥X′∥2∥X∥2\operatorname{sim}(X',X)=\frac{X'^\top X}{\|X'\|_2\|X\|_2}. Each perturbed input is then fed into TSFMs to obtain a forecast, and the forecast MSE is evaluated against the ground truth. Thus, the horizontal axis measures input similarity (or fidelity), while the vertical axis measures forecasting error.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Clarifications to Reviewer Concerns

Dear Reviewers and Area Chairs,

Thank you for your constructive feedback. We summarize the main clarifications below:

  • Quantified gains beyond standard sampling. We added Std@64 to the main table. With the same sampling budget and MV aggregation, diversification improves 19/24 configurations and 7/8 dataset-averaged comparisons. (Reviewers edko and arKv) 
  • Separated oracle analysis from practical inference. We clarified that the theory characterizes oracle headroom, while practical gains are evaluated through MV without oracle candidate selection. (Reviewers CNZd and arKv) 
  • Added comparisons with SPRINT. Under matched historical information, DSI achieves lower MV MSE in 22/24 comparisons. We explicitly distinguish equal candidate counts from equal computational costs. (Reviewer arKv) 
  • Made perturbation guidance and overhead explicit. We clarified perturbation settings and practical defaults, and separately reported generation costs alongside end-to-end latency measurements. (Reviewer CNZd and arKv)

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.