acceptodds
Under review as a conference paper at ICLR 2027

Does More History Help? Understanding Context Length in Time-Series Foundation Models

Abstract

*Time series foundation models* (TSFMs) for time series forecasting are designed to operate on variable input context lengths. This leads to the question: *how much context is worth using for a given task?* While the current trend favors longer contexts to enable more flexible predictors, their beneficial effect on inference-time accuracy is not guaranteed. In this work, we systematically investigate the influence of context length on forecasting performance across state-of-the-art TSFMs, and show consistent diminishing returns as the context length grows: an effect called *context saturation*. Through experiments in controlled environments, we provide insight into the factors driving context saturation, and show signs of recency bias in TSFMs, where their ability to leverage long-range dependencies varies substantially by model. Furthermore, through oracle context-length selection, we demonstrate the existence of operating points with both lower aggregate forecasting error and lower computational cost than those resulting from maximum context length. Finally, we propose a methodology for *zero-shot* inference-time context length selection, leveraging only synthetic data. Evaluated across all TSFMs on the GiftEval benchmark, our approach achieves substantial efficiency gains, validating the practical impact of our findings.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time series foundation models, Context Length, Time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qik5OpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the phenomenon of context saturation in timeseries foundation models, examining how the predictive performance of 11 pre-trained models varies with input history length. Through controlled synthetic experiments and GiftEval-based evaluation, the study analyzes task- and model-specific context requirements and proposes a lightweight selector trained solely on synthetic data, aiming to minimize theoretical computational costs while maximizing predictive accuracy.

Critical strengths
  1. The problem setting has clear practical value, as context length selection helps balance forecasting accuracy and computational cost when deploying time series foundation models.
  2. The paper provides a comprehensive experimental evaluation across 11 models, covering both controlled synthetic settings and real-world forecasting tasks.
Critical weaknesses
  1. The paper provides a theoretical efficiency analysis but lacks actual runtime measurements.
  2. The method is calibrated on synthetic data, but the real-world evaluation mainly reports aggregate MASE, potentially masking threshold violations and severe error degradation in individual cases.
Questions for authors

None

3 comments

We thank the reviewer for their review and comments. We address the reviewer's concerns in the following.

W1: The paper provides a theoretical efficiency analysis but lacks actual runtime measurements.

We agree that actual runtime measurements can provide a more direct perspective about the computational speedup achievable under a particular policy. We favored the theoretical efficiency as runtime improvements depend on many hardware and implementation dependent factors. Nonetheless, in the revised manuscript we added Table 16, reporting a case study for the Max Efficiency policy, showing how MAC savings translate to actual computation speedup on both GPU and CPU. Indeed, the results show that throughput can be significantly improved, even when explicitly accounting for the cost of the full context selection pipeline (i.e., feature extraction and candidate lengths risk scoring). This highlights how dynamic context length selection can provide concrete practical advantages, beyond the theoretical MAC quantification.

W2: The method is calibrated on synthetic data, but the real-world evaluation mainly reports aggregate MASE, potentially masking threshold violations and severe error degradation in individual cases.

Let us clarify that it is common practice to evaluate TSFMs performance in terms of aggregate metrics [1,2,3,4]. Indeed these can mask severe degradations in individual cases, as much as they can mask outstanding improvements. To assess whether variations around the reported aggregate metrics are substantial, in the revised manuscript we added Appendix D.7, where we show the standard deviation, as well as 55-th and 9595-th percentile, of the relative MASE change for dynamic and fixed length context selection policies. In particular, Table 13 reports values across the Gift-Eval benchmark, while Table 14 shows per-model measurements. Overall, these results suggest that dynamic context length selection does not present excessive variability, in particular, compared to fixed length policies, relative error variability is lower when targeting higher computational savings. Intuitively, this might be due to the additional flexibility in selecting an appropriate context length on an input-dependent basis.

Regarding threshold violations, indeed those could occur and we do not claim otherwise. The calibration targets compliance in aggregation over the target data distribution, hence individual samples may surpass the thresholds. To provide information about this, we added Section D.8 in the revised manuscript's appendix. It reports the fraction of GIFT-Eval samples on which the threshold is violated by each individual dynamic policy for the two thresholds considered by the methodology: (i) the mean relative error degradation and (ii) the percentage of samples that exceed a 5%5\% error degradation. The results (Table 15) show how the first threshold is occasionally violated, mainly for more aggressive policies, while the second threshold is almost never violated. Let us remark that we do not expect the thresholds to hold strictly, as the context length selector is calibrated on synthetic data, whose input-dependent error distribution differs from that of the real-world data we evaluate the methodology on.

We thank the reviewer again for their comments. Please let us know if there are any outstanding concerns preventing a higher score. We remain open to further discussion during the rebuttal period.

[1] Petralia, Adrien, et al. "Timestral." arXiv, 27 Oct. 2026, doi:10.48550/arXiv.2610.37585.

[2] Ansari, Abdul Fatir, et al. "Chronos-2: From Univariate to Universal Forecasting." arXiv, 17 Oct. 2025, doi:10.48550/arXiv.2510.15821.

[3] Shchur, Oleksandr, et al. "fev-bench: A Realistic Benchmark for Time Series Forecasting." arXiv, 30 Sept. 2025, doi:10.48550/arXiv.2509.26468.

[4] Aksu, Taha, et al. "GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation." arXiv, 14 Oct. 2024, doi:10.48550/arXiv.2410.10393.

Reviewer_qik5OpenReview ↗

Response to Authors

Thank you for the detailed response and additional analyses, which address most of my concerns. I have therefore decided to raise my rating from 2 to 3. Best wishes.

We thank the reviewer for acknowledging our response and raising the score. We are happy that it addressed their concerns.

Reviewer_QfH3OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends previous research on context saturation in time-series foundation models (TSFMs) through a systematic evaluation of 11 pretrained models. Using controlled synthetic experiments, the authors investigate how task characteristics, including seasonality, forecasting horizon, regime changes, and noise, relate to the context length required for near-optimal forecasting performance. Additional perturbation and attention analyses examine model-specific recency biases and long-range dependency utilization. The authors further conduct oracle context-length sweeps on GIFT-Eval to identify potential accuracy–efficiency improvements, and propose a zero-shot context selector trained entirely on synthetic data. The proposed method achieves substantial theoretical computational savings with relatively small forecasting accuracy degradation.

Critical strengths

The paper systematically investigates context saturation across 11 pretrained TSFMs, extending previous research through diverse controlled synthetic environments and real-world forecasting tasks. The analysis connects context-length requirements with temporal characteristics such as seasonality, forecasting horizon, regime changes, and noise. Additional perturbation and attention experiments provide useful insights into model-specific recency biases and long-range dependency utilization.

The proposed approach requires no modification of pretrained TSFM parameters and trains its context selector entirely on synthetic data. The oracle analysis demonstrates opportunities to improve both forecasting accuracy and computational efficiency through appropriate context selection. The practical selector further achieves substantial theoretical computational savings with relatively small forecasting accuracy degradation.

Critical weaknesses

1.We believe that this article is too engineering and empirical, and the choice of historical length and performance benefits is not significant. Table 1 shows considerable variation in the accuracy–efficiency trade-offs across different TSFMs. The Balanced policy provides limited average computational savings, while more aggressive policies generally increase forecasting errors. Although the results demonstrate potential efficiency gains, the improvements are not consistent across models, making it difficult to determine how reliably the proposed selector identifies appropriate context lengths for different forecasting tasks.

2.Although there are many TSFMs, the comparison between baselines and benchmarks in this article is limited. The main evaluation relies on GIFT-Eval, and the proposed selector is primarily compared against a fixed 4K context policy. Additional evaluation on benchmarks such as fev-bench and TSFM-Bench, together with comparisons against shorter fixed lookback windows and simple property-based selection strategies, would strengthen the evidence for the method's generalizability and practical advantages.

3.The impact of different pretraining configurations on context sensitivity remains insufficiently analyzed. Differences in training data, training budgets, and context-length distributions may influence the optimal history length, making it difficult to attribute the observed saturation patterns primarily to time-series characteristic

Questions for authors
  1. In conventional long-term time-series forecasting, the lookback window is often treated as a model hyperparameter, and relatively short windows are commonly used. For example, LTSF-Linear (AAAI 2023) and iTransformer (ICLR 2024) investigate different lookback lengths, including 96, 192, 336, and 720, which are substantially shorter than 4K. Moreover, task-specific models can outperform TSFMs in some settings. Could the authors discuss how their findings relate to these established practices and provide comparisons with shorter fixed windows?

  2. Could the authors extend the evaluation to additional benchmarks, such as fev-bench and TSFM-Bench, to verify whether the observed context-saturation patterns and efficiency gains generalize beyond GIFT-Eval?

  3. Could the authors further quantify the relationship between oracle-selected context lengths and time-series characteristics on real-world datasets, beyond the controlled synthetic experiments?

7 comments

We thank the reviewer for their review and feedback. In the following we address the reviewer's concerns point by point.

W1: We believe that this article is too engineering and empirical, and the choice of historical length and performance benefits is not significant.

We respectfully disagree that the choice of historical lengths and performance benefits are not significant. The lengths we take into account are relevant when considering TSFMs, which are trained with context lengths far larger than those of task-specific models [2,3,4]. Moreover, recently proposed TSFMs suggest the current trend is going for even longer contexts [1]. While such context lengths might appear unnecessarily large, that is exactly the point of our discussion: TSFMs are trained with context lengths that are not specific to any particular task, yet discussion about the factors that determine efficient length selection at inference time is absent in the literature. While our work is empirical (as most works on TSFMs), our analysis is grounded in the perspective that TSFMs use the input context to infer the target task [5], and the amount of observations that is effective largely depends on the task itself, as we demonstrate in Section 3.

Table 1 shows considerable variation in the accuracy–efficiency trade-offs across different TSFMs. The Balanced policy provides limited average computational savings, while more aggressive policies generally increase forecasting errors. Although the results demonstrate potential efficiency gains, the improvements are not consistent across models, making it difficult to determine how reliably the proposed selector identifies appropriate context lengths for different forecasting tasks.

The fact that the achievable gains depend on the considered TSFM is expected, that would be true of any approach that is based on a pre-trained component, much like for a fine-tuning strategy the final performance would depend on the pre-trained backbone. The fact that more conservative policies allow modest efficiency gains, while more aggressive ones incur in higher errors is also expected, as the policies themselves reflect different target trade-offs where the relative importance of the error compared to the computational savings is different. The goal of Section 5 is to demonstrate the feasibility of dynamic selection and show that it can allow better trade-off points than fixed-length policies. Please refer to the answer to the following question where we discuss how we improved the manuscript to better convey this message.

W2: Although there are many TSFMs, the comparison between baselines and benchmarks in this article is limited. The main evaluation relies on GIFT-Eval, and the proposed selector is primarily compared against a fixed 4K context policy. Additional evaluation on benchmarks such as fev-bench and TSFM-Bench, together with comparisons against shorter fixed lookback windows and simple property-based selection strategies, would strengthen the evidence for the method's generalizability and practical advantages.

We agree with the reviewer that the single 4k4k baseline might not sufficiently establish the effectiveness of dynamic context length selection. To provide further evidence of the practical utility of the approach, we extended the comparison in the revised version of the paper. In particular, we compare the trade-offs achievable with the dynamic policies to those achievable by fixed-length policies of two kinds: (i) those selecting the fixed length that solves the same optimization problem of the dynamic policies, by calibration on the same synthetic data, and (ii) those that select a fixed length a-priori. Results are reported in Figures 4 and 5 of the revised manuscript and show that, for most models, the dynamic policies enable better trade-off points than fixed-length policies, in particular when targeting higher computational savings.

The revised manuscript now also includes analogous results for the Fev-Bench benchmark in Figure 23 of the appendix. Here, the dynamic context selection approach often offers similar trade-offs to fixed-length policies, though it still enables better higher computational savings trade-off points that would not be achievable by fixed-length policies. This shows the design is practically useful: intuitively, selecting the context length in an input-dependent way allows the flexibility required to mitigate accuracy degradation when targeting larger computational savings. Allow us to stress that the objective of the methodology in Section 5 is to provide evidence of the feasibility of zero-shot dynamic context length selection (ll. 69-70). This has to be considered within the broader scope of the paper, that discusses the problem of context saturation and inference-time context length selection, which was not considered in previous literature on TSFMs.

Regarding simple property-based selection strategies, in Table 17 of the revised manuscript's appendix, we compare the efficiency trade-offs of the dynamic policies of Section 5, to those achievable with rule-based approaches that select the context length as a function of the target forecasting horizon and of the estimated seasonality of the series. As the results show, these approaches incur in significantly higher error.

W3: The impact of different pretraining configurations on context sensitivity remains insufficiently analyzed. Differences in training data, training budgets, and context-length distributions may influence the optimal history length, making it difficult to attribute the observed saturation patterns primarily to time-series characteristic

Let us clarify that the TSFMs we consider in our work already span a wide variety of pre-training configurations, in terms of parameter count, training data, budget, and context-length. While this is not equivalent to a systematic evaluation where these parameters are consistently varied, we believe our results in Section 3 clearly establish a strong connection between context saturation and time-series characteristics. Indeed, regardless of the configuration variability across evaluated TSFMs, saturation patterns are often dominated by the varied task characteristic (e.g., seasonality, horizon, regime shift). This effect intuitively connects to how the most effective input length selected with hyperparameter tuning depends on the target task for conventional task-specific models. Indeed, model-specific effects might be present, as we acknowledge in the paper (ll. 63-64, 308-309), and studying the impact of the pre-training configuration would be an interesting direction for future work.

Q1:In conventional long-term time-series forecasting, the lookback window is often treated as a model hyperparameter, and relatively short windows are commonly used. For example, LTSF-Linear (AAAI 2023) and iTransformer (ICLR 2024) investigate different lookback lengths, including 96, 192, 336, and 720, which are substantially shorter than 4K. Moreover, task-specific models can outperform TSFMs in some settings. Could the authors discuss how their findings relate to these established practices and provide comparisons with shorter fixed windows?

As the reviewer noted, in conventional time-series forecasting the lookback length is usually an hyperparameter selected at training time. In TSFMs however, the model is not tailored to a task in particular, hence the pre-training lookback length is usually chosen to be as large as possible within the computational budget for training. Our findings show that this length is often wasteful at inference time, where performance on individual target tasks may saturate at much shorter lookback lengths. As such, this choice which conventionally occurred at training time, for TSFMs becomes and inference-time choice. Our work establishes this distinction. The connection with conventional practices lies in the observed effect of task characteristics on the saturation length: as the appropriate lookback length of task-specific models depends on the task, also context saturation length in TSFMs depends consistently on the specific inference-time task (as our results in Section 3 suggest).

We are aware that conventional deep learning models for time-series forecasting usually rely on shorter input lengths, however those input lengths are often inadequate for TSFMs and lead to severe performance degradation (see Appendix D.6 and Table 11 in the revised manuscript). The difference in lookback length to performance trade-offs between task-specific models and TSFMs has been already discussed in the literature [5]: as they are "generalist" predictors, TSFMs usually require longer lookbacks compared to task-specific models, as the additional observations are needed to resolve the uncertainty about which is the inference-time task. Our work deals with a different problem: as the lookback length of TSFMs is not tailored to any target task, how should one select an efficient lookback length at inference time? Ultimately, our goal is to provide evidence that TSFMs can be used in a more efficient way than their pre-training length suggests, and that zero-shot context length selection is a promising direction for future developments.

Q2:Could the authors extend the evaluation to additional benchmarks, such as fev-bench and TSFM-Bench, to verify whether the observed context-saturation patterns and efficiency gains generalize beyond GIFT-Eval?

Please refer to our previous answer to W2, where we discuss results on fev-bench. Note that TSFMBench does only consider tasks with context length up to 512512 time samples [6]. As such, it is not appropriate to evaluate dynamic context length selection, as the provided context length is already severely short compared to the pre-training length of TSFMs. To further expand the considered evaluation settings, we have also computed results for the recent TIME benchmark [7], in Figure 24, as suggested by Reviewer wZHW. We include them in the revision as well. The results are in line with our observation that dynamic context length selection enables efficiency trade-offs that may not be available to fixed-length policies, and this can be advantageous when targeting higher computational savings.

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

The paper investigates the context saturation problem of time series foundation models and proposes a new method for zero-shot inference-time context length selection. The context length selection module are trained only on synthetic data and used on the real datasets without further adjustment. To demonstrate the effectiveness of the proposed method for efficiency gains through context length selection, the paper conducts experiments on GIFT-Eval across 11 backbone checkpoints.

Critical strengths
  1. The paper is well-motivated and easy to follow.

  2. The experiments are conducted across diverse model architectures.

  3. The proposed context selection module are learned only on synthetic data while works well on GIFT-Eval real dataset, which is interesting.

Critical weaknesses
  1. The current 4k baseline in Table 1 is not enough to demonstrate the effectiveness of the context length selection strategy. Since the backbones are different with the maximum available context length, only using 4k as the baseline is not enough and not fair for all checkpoints.

  2. Also, the current results in table 1 seems that 4k is part of the trade-off setting of the proposed method similar with the other three rows, especially under the univariate setting. The efficiency gains are not consistent across backbones. For some backbones, 4k has intermediate delta MASE and delta MAC between Balanced and Very aggressive (and even better than both of them for Moirai-2), which makes the selection strategy look not important.

  3. The MAC reported by the paper only includes the forward pass of the TSFMs. It is better to report the efficiency results of both the context selection process and the forward pass of TSFM.

Questions for authors
  1. Do the authors try the proposed method on some other benchmarks? such as TIME and fev-bench?

  2. Can the authors also provide baselines of each checkpoint with some max-related context length? such as 78Lmax\frac{7}{8} L_{max}, 34Lmax\frac{3}{4} L_{max}, 12Lmax\frac{1}{2} L_{max}?

  3. How is the proposed method sensitive to the context length candidate sets? Do the authors try more candidates number or even a continuous set for the context length set?

4 comments

We thank the reviewer for their review and the thoughtful comments. We address the reviewer's concerns point by point in the following.

W1: The current 4k baseline in Table 1 is not enough to demonstrate the effectiveness of the context length selection strategy. Since the backbones are different with the maximum available context length, only using 4k as the baseline is not enough and not fair for all checkpoints.

We agree with the reviewer that the single 4k4k baseline might not sufficiently establish the effectiveness of dynamic context length selection. To provide further evidence of the practical utility of the approach, we extended the comparison in the revised version of the paper. Instead of comparing to 4k4k only, we compare the trade-offs achievable with the dynamic policies, to those achievable by fixed-length policies that optimize the same objective (Figure 4 and Figure 5 in the revised document). In particular, for each dynamic policy, we find the fixed length the solves the same optimization problem, i.e., the minimum length for which the average error on the synthetic calibration data does not exceed the prescribed ratio to the error achieved using the maximum context length of the pre-trained model. The results show that, for most models, the dynamic policies allow better efficiency trade-offs than the fixed-length policies ones, in particular when targeting higher computational savings.

W2: Also, the current results in table 1 seems that 4k is part of the trade-off setting of the proposed method similar with the other three rows, especially under the univariate setting. The efficiency gains are not consistent across backbones. For some backbones, 4k has intermediate delta MASE and delta MAC between Balanced and Very aggressive (and even better than both of them for Moirai-2), which makes the selection strategy look not important.

Having efficiency gains that vary across different pre-trained models is to be expected, and we do not claim otherwise. Indeed, model-specific effects influence the achievable trade-offs, yet that is inherent to any approach that involves a pre-trained component (e.g., performance achievable by transfer learning depends on the backbone). The fact that fixed-4k4k context length can provide trade-offs that are in-between proposed policies does not invalidate the importance of the selection strategy. It is simply an effect of the fact that different dynamic policies aim for different cost-to-error trade-offs, while fixed-4k4k inherently reflects a particular point in the cost-to-error space. As acknowledged in the previous answer, we agree this might make Table 1 hard to interpret, and we replaced it with Figure 4 and Figure 5 in the revised version. The efficiency trade-off curves clarify how in most cases dynamic policies allow to exploit trade-offs that are not available to policies based on a fixed length.

W3: The MAC reported by the paper only includes the forward pass of the TSFMs. It is better to report the efficiency results of both the context selection process and the forward pass of TSFM.

We agree that including the MAC cost of the whole context selection pipeline would be more precise. We omitted its cost as it was negligible. Across the considered foundation models, the selection pipeline has a MAC between 0.01%0.01\% and 0.00005%0.00005\% of the model's forward pass, with a median of 0.0012%0.0012\%. The revised manuscript now reports this information in Table 10 of the appendix. To provide a practical perspective on the impact of the context selection pipeline, in the revised manuscript we added Table 16, reporting a case study for the Max Efficiency policy, showing how MAC savings translate to actual computation speedup on both GPU and CPU. Indeed, the results show that throughput can be significantly improved, even when explicitly accounting for the cost of the full context selection pipeline (i.e., feature extraction and candidate lengths risk scoring).

Q1: Do the authors try the proposed method on some other benchmarks? such as TIME and fev-bench?

In the revised manuscript's appendix, we added results for fev-bench (Figure 23 and Table 8B) and TIME (Figure 24 and Table 8C). The figures report efficiency trade-offs achievable under both dynamic and fixed length context selection policies, following the approach described in our earlier answer to W1. The tables instead show the trade-offs achievable by trying to match the computational savings of dynamic length selection policies by means of a fixed context length. On these benchmarks the dynamic context selection policies often offer similar trade-offs to fixed-length policies, yet the former can still enable better trade-offs when targeting higher computational savings. These results show that dynamic context length selection is practically useful; intuitively, when higher computational savings are desired, choosing the context length in an input-dependent manner provides a useful flexibility that individual fixed lengths lack. Allow us to stress that the objective of the methodology in Section 5 is not to establish dominance over fixed length baselines, but rather to provide evidence of the feasibility of zero-shot dynamic context length selection (ll. 69-70). This must be considered within the broader scope of the manuscript, that discusses how the problem of context length selection, which was classically a training-time hyperparameter selection problem, becomes an inference-time decision for TSFMs, which depends on the task at hand. In fact, a TSFM pre-training length is not chosen to satisfy the needs of any particular downstream task and, as we demonstrate, using the maximum supported length can be wasteful. As there are no established baselines for this problem, our goal was to discuss it, demonstrate solutions' feasibility, and provide a first step towards further research on the matter.

Q2 Can the authors also provide baselines of each checkpoint with some max-related context length? such as 78Lmax⁡\frac{7}{8}L_{\max}, 34Lmax⁡\frac{3}{4}L_{\max}, 12Lmax⁡\frac{1}{2}L_{\max}?

We included this baseline in the revised manuscript's appendix, covering Gift-Eval (Figures 21 and 22), Fev-Bench (Figure 23), and TIME (Figure 24). In most cases dynamic policies are better or comparable, often allowing to exploit trade-off points that are not available to fixed-length approaches (in this case taking a fixed fraction of the maximum length). Note that selecting a fixed fraction of the maximum length a-priori, can have significantly different impacts depending on the maximum length itself. Indeed, our results in Section 3 suggest that the context saturation length depends on properties of the task, hence halving the context length of a model with 32k32k maximum context cannot be considered as semantically equivalent to halving that of a model with 2k2k maximum length.

Q3 How is the proposed method sensitive to the context length candidate sets? Do the authors try more candidates number or even a continuous set for the context length set?

This is an interesting question. To provide evidence that the method is not particularly sensitive to the set of candidate context lengths, we evaluated its performance when, instead of scoring only the context lengths used for training and calibration, we score all possible context lengths between 11 and the TSFM maximum context length. Results are reported in Table 9 of the revised manuscript's appendix. Overall, the results suggest that the approach is not particularly sensitive to the choice of candidate grid and, in this case, scoring all possible options results in trade-offs towards slightly higher computational savings, at the cost of a small error increase. Future work may consider studying more advanced approaches for candidate grid selection, e.g., by assigning model-specific a-priori risks to different context lengths to provide an inductive bias for dynamic context length selection.

We thank again the reviewer for their questions and remain open to further discussion during the rebuttal period.

I thank the authors for the detailed response, which solved most of my concerns. The comparisons with fixed caps and fixed lengths calibrated on the same synthetic data address my main concern about relying on a single 4K baseline. The additional evaluations on TIME and fev-bench, the candidate-grid sensitivity test, and the end-to-end timing results including selector overhead also address my questions. I have raised my score to weak accept accordingly.

We thank the reviewer for acknowledging our response and raising the score.

We remain available should the reviewer wish to discuss any other aspect of our work.

Comments on the paper

2 comments

We thank all the reviewers for their comments and the time dedicated to reviewing our paper reviews. In the following, we summarize the main changes made to the manuscript in order to address the reviewers' concerns.

  • We broaden the evaluation of the dynamic context length selection methodology of Section 5 to include more informative baselines. In particular, Figure 4 and Figure 5 show the efficiency trade-offs for dynamic context length selection, and compare them to fixed length selection policies. We consider two kind of fixed-length policies: (i) choosing the length a-priori, (ii) choosing the length by finding the one which solves the same optimization problem as dynamic selection, by calibration on the same synthetic data. Additionally, in Appendix D.3, we include comparisons to: (i) a-priori fixed lengths chosen as a fraction of the TSFM's maximum context length (Figure 21 and Figure 22), and to (ii) fixed lengths chosen in hindsight, to match dynamic policies computational savings (Table 7 and Table 8A). Overall, these results clarify the advantage of dynamic context length selection, in particular when targeting higher computational savings.

  • In Appendix D.3, we extend the evaluation described at the previous point to Fev-Bench (Figure 23 and Table 8B) and TIME (Figure 24 and Table 8C) benchmarks. This makes the evaluation of the dynamic context length selection approach particularly broad, supporting its overall practical applicability, and showing it allows efficiency trade-off points that are precluded to fixed-length policies.

  • We extended the evaluation to include 22 additional TSFMs: Timestral-209M and Timestral-123M [1]. These very recent pre-trained models further confirm the trend towards increasingly longer pre-training context lengths, reaching 3232K time-steps in this case. This suggests that inference-time context length selection might become increasingly important for efficient TSFM utilization.

  • We expanded the appendix to include further analysis to address individual reviewers' concerns. In particular, Appendix A.4, and Appendix D.4 through D.10. Individual results are discussed specifically in each response. Changes in the revised text are highlighted in blue for the reviewers' convenience.

Let us remark that the methodology evaluated in Section 5 should be considered within the broader scope of the paper, which asks whether TSFMs should be used at their maximum context length at inference time. We support this discussion with empirical studies in Sections 2, 3 and 4. The methodology in Section 5 is then introduced to demonstrate the feasibility of exploiting the insight that the effective context length depends on the target task at inference time.

We thank again the reviewers for their time.

[1] Petralia, Adrien, et al. "Timestral." arXiv, 27 Oct. 2026, doi:10.48550/arXiv.2610.37585.

Dear reviewers,

Thank you again for reviewing our work.

As the discussion period is close to its end, we would like to know if there is any remaining concern. We are available to provide any additional clarification.

Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.