acceptodds
Under review as a conference paper at ICLR 2027

Recursive Inference for Test-Time Compute Allocation in Time-Series Foundation Models

Abstract

In time-series forecasting tasks, some forecasts require carrying seasonal demand \em months into the future whereas in other settings, one must anticipate changes in electricity load over the \em next few hours. Still, a pretrained forecasting model typically runs the same sequence of blocks \em once for each task. In this paper, we ask whether a small amount of additional inference budget can be utilized for tangible performance gains, in a way that leaves the model’s parameters frozen. A fixed budget can be variously used to distribute compute. We explore these choices through compute programs that revisit selected consecutive blocks, control the size of each update, and determine which token representations can change. Their cost in encoder-block evaluations is known in advance. But their impact on forecast error for the particular task must be measured. Recursive inference uses a reasoning model to propose programs, read their results on historical forecasting windows, and revise subsequent proposals. Held-out later windows decide whether to deploy a program, and it then stays \em fixed throughout testing. We find that this simple idea, across 56 tasks from 14 datasets, reduces test MSE by 4.02% on average, at only 1.28 times the baseline encoder-block evaluations per forecast.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, time-series foundation models, test-time compute, adaptive computation, looped transformers, frozen models, LLM reasoning

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

This paper proposes a test-time method that improves forecasting accuracy by repeating a selected window of encoder blocks in a frozen time-series foundation model (Chronos-2). The extra cost of each repetition scheme is known exactly in advance. An LLM searches for a scheme per task using errors measured on historical windows, and the chosen scheme is deployed only after passing a held-out verification check. Across 56 tasks, it reduces test MSE relative to the base Chronos-2 model, at controllable computation cost relative to base.

Critical strengths
  1. The paper treats the internal execution path of a frozen time-series foundation model as test-time compute allocation. It has a certain degree of novelty and parallels layer-looping work in language models.
  2. The extra cost of repeating encoder blocks is known exactly before running, which makes budget-aware deployment capable.
Critical weaknesses
  1. The paper provides some initial analysis, but it is mostly correlational, lacks intervention experiments, and does not explain why different tasks benefit from repeating blocks at different positions, or how different encoder block position changes the forecast.
  2. The method relies on Chronos-2's specific context and future token structure and is evaluated only on Chronos-2, so it is unclear whether it transfers to other time-series foundation models.
  3. Each task uses at most 18 candidate programs, while Chronos-2 has only 12 encoder blocks and the program space is low-dimensional. Maybe rule-based methods such as random or grid search might perform comparably? But the paper does not compare against them, so the necessity of the LLM reasoner is not fully demonstrated.
Questions for authors
  1. Did you observe typical forecast changes when repeating encoder blocks at different positions (e.g., corrections to trend, seasonality, or volatility)? Is there a systematic relationship between the repetition position and series characteristics such as autocorrelation, volatility shift, or cross-channel correlation?
  2. How would this design transfer to other models that not follow Chronos-2's joint context and future token structure?
  3. The method is conceptually close to loop transformer architecture. Beyond being training-free, what is the essential difference between your work and loop transformer?
6 comments

We sincerely thank Reviewer pAjG for the thoughtful feedback and constructive suggestions, which have helped us further improve our work. We address each point below.

W1: Transfer to another TSFM.

The concern about generalization beyond Chronos-2 is very reasonable and was also raised by Reviewers VmaH and w4Zx. To address this shared concern, we conducted additional experiments on a second time-series foundation model, Timestral-209M. Here, we also address the reviewer's more specific question: whether RI depends on Chronos-2's particular context and future-token structure.

RI can transfer through repetition of native Transformer blocks without requiring Chronos-2’s joint context and future-token structure. Selectively updating future-token representations is an execution option available for Chronos-2, but not a requirement of our RI idea. For a model without that partition, the model adapter can instead expose recipes that repeat complete blocks over their full hidden state, preserving the model’s native attention and inference semantics. The reasoner then proposes recipes using the execution choices supported by that architecture.

We implemented and evaluated this approach in Timestral-209M with Claude Haiku 5.5. Timestral encodes context causally and uses separate prediction tokens that attend to the context, while context tokens do not attend back to them [1]. Our implementation repeats complete native blocks across both streams, retaining their original attention masks, positions, and feed-forward computations. The reasoner gets Timestral’s block structure and task information, such as the forecasting horizon, and revises its proposals using forecasting feedback from training data. All forecaster weights remain frozen.

The transferred workflow finds useful recipes under these different token interactions. Across the reported 12 datasets covering 48 tasks, RI with Haiku achieves mean dataset-level reductions of 5.97% in MSE and 1.08% in MAE, after averaging four horizons within each dataset. MSE improves on 11 of 12 datasets, and the mean reduction remains 2.18% even when excluding Covid-19, which has a particularly large gain. The results support transferring the proposal-and-feedback workflow through the target model's native computation (with its own token structure preserved).

Our Timestral implementation also produces average gains with Sonnet in the eight-task comparison. Under the same workflow and candidate budget, average MSE/MAE reductions are 13.77%/2.53% with Haiku and 14.52%/3.00% with Sonnet, providing supplementary evidence across reasoner choices. So, our main point is that RI does not need a particular token-update mechanism. It assigns/allocates compute using the execution choices supported by the target forecaster. Our Timestral experiments show that this approach remains effective with a different attention structure and native full-block repetition.

Table 1. Generalization to Timestral-209M with Claude Haiku 5.5. Results cover 12 datasets and 48 forecasting tasks. MSE and MAE are averaged over four horizons within each dataset. The Average row reports the mean dataset-level percentage change. Δ denotes percentage change relative to Base. Negative values indicate lower forecasting error.

DatasetMSE↓ (Base→RI)Δ MSEMAE↓ (Base→RI)Δ MAE
Covid-192.793→1.463-47.63%0.041→0.039-3.24%
ETTh10.400→0.4000.00%0.402→0.4020.00%
ETTh20.344→0.340-1.15%0.369→0.370+0.10%
ETTm10.378→0.356-5.83%0.362→0.358-1.16%
ETTm20.269→0.267-0.46%0.302→0.303+0.56%
ILI1.595→1.592-0.20%0.765→0.7650.00%
NASDAQ0.933→0.913-2.13%0.650→0.648-0.35%
NN50.673→0.672-0.10%0.548→0.548-0.01%
NYSE0.442→0.436-1.26%0.410→0.405-1.40%
PEMS080.495→0.437-11.74%0.438→0.404-7.84%
Wind1.146→1.143-0.26%0.735→0.735-0.03%
ZafNoo0.521→0.517-0.84%0.409→0.411+0.35%
Average—-5.97%—-1.08%

W2: Comparisons with conventional search methods.

This concern (is LLM-guided search needed?) was also raised by Reviewers VmaH and w4Zx. To address this concern, we conducted the requested comparisons against conventional search methods. Our results show that LLM-guided recipe generation achieves much larger forecasting improvements under the same candidate-evaluation budget.

Our Sol-based reasoner finds more effective recipes than uniform random search and the three adaptive search methods we evaluated under the same 18-candidate budget. Although Chronos-2 has only 12 encoder blocks, the search involves much more than selecting a block position. These comparisons directly check whether conventional search can recover the gains of LLM-guided recipe generation within the paper’s search budget.

All methods search the same legal recipe space. Chronos-2’s 12 blocks define 78 contiguous block ranges, with additional choices of repetition count, solver, and update strength. We evaluate 60 tasks spanning 15 datasets and four horizons per dataset, using the same candidate-evaluation budget, final validation selector, and test protocol for every method.

The baselines include the suggested uniform random search, which samples recipe parameters without feedback. The Tree-structured Parzen Estimator (TPE) [1] uses previous evaluations to propose promising configurations. Differential evolution (DE) [2] mutates and recombines parameter vectors. Monte Carlo tree search (MCTS) [3] explores a tree of recipe choices using Upper Confidence Bounds applied to Trees (UCT). The adaptive searches receive their own training MSE feedback. The LLM-guided workflow additionally uses model knowledge and task information to construct and revise proposals.

The differences are quite meaningful. Random search does not reproduce the observed gains. Its average MSE is essentially unchanged (+0.001%), while MAE increases by 1.41%. Sol reduces average MSE by 4.15% and MAE by 0.94%, improving MSE on 53/60 tasks. MCTS, the strongest of the four conventional searches by average MSE, achieves a 1.97% reduction and improves 33/60 tasks. All four conventional searches worsen average MAE. With final selection held fixed, these results support the role of LLM-guided candidate generation under the limited evaluation budget.

The advantage also appears with a smaller local reasoner. Qwen3-4B-Thinking-2507 reduces average MSE by 3.00% and MAE by 0.71%, exceeding the tested searches on both aggregate metrics.

[1] Bergstra et al. Algorithms for Hyper-Parameter Optimization. NeurIPS, 2011.

[2] Storn and Price. Differential Evolution: A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces. Journal of Global Optimization, 1997.

[3] Kocsis and Szepesvári. Bandit Based Monte-Carlo Planning. ECML, 2006.

Table 2: Local and API reasoners versus conventional search for Chronos-2 looping recipes. All methods are evaluated on 60 tasks. B/T/W = Better/Tie/Worse. API models use low reasoning effort.

TypeModelMSE Avg Δ% ↓MSE B/T/WMAE Avg Δ% ↓MAE B/T/W
LocalQwen4B-3.00%48/7/5-0.71%36/7/17
LocalDeepSeek8B-1.42%50/7/3+0.15%37/7/16
APILuna-low-3.36%46/3/11-0.72%40/3/17
APISol-low-4.15%53/4/3-0.94%48/4/8
APITerra-low-3.78%49/7/4-0.80%40/7/13
SearchUniform+0.001%17/29/14+1.41%5/29/26
SearchTPE-1.14%28/23/9+0.69%16/23/21
SearchDE-0.43%25/23/12+1.14%15/23/22
SearchMCTS-1.97%33/18/9+0.54%19/18/23

Q1: Analysis of looping positions.

This is a very interesting question. We have done additional analyses to address what looping changes in the forecasts, why position of repetition matters, and how these effects relate to time-series characteristics. We go beyond the correlational analysis by decomposing the forecast corrections and directly measuring how their individual components affect prediction error.

First, we examine which changes in the forecasts are responsible for the improvements. For the 56 selected recipes on 16 test origins per task, we orthogonally decompose the forecast correction (Loop minus Base) into interpretable components and evaluate each component by adding it separately to the Base forecast. Level (constant shift) and trend (slope change) account for 61.5% and 14.1% of the net macro MSE gain, respectively, totaling 75.6%. A harmonic at a frequency estimated from training contributes 3.0%. Remaining changes contribute 21.4% and may include curvature, local adjustments and other frequencies. So**, the largest improvements come from correcting the forecast's level and trend.**

Second, we check how these forecast corrections differ depending on where blocks are repeated. Looping adjusts both periodic amplitude and peak timing, with amplitude changes occurring in both directions. Across 41 matched pairs of beneficial recipes from 1,000 reasoner proposals, early repetition increases the fitted periodic amplitude in 31.1% of channel forecasts and decreases it in 64.5%. At late positions, the corresponding rates are 12.9% and 82.9%, with approximately 4% unchanged at either position. These rates are averaged equally across datasets. Later repetition also produces larger shifts in the timing of fitted periodic peaks and troughs relative to Base, with mean absolute phase changes of 0.14 versus 0.04 radians. So, these results show that repetition at different block positions produces different forecast corrections. Basically late repetition more frequently attenuates periodic amplitude and produces larger phase adjustments.

Third, we check whether the relative benefit of early and late repetition is related to properties of the input series. Later repetition tends to offer a larger advantage when channels are more strongly related. Across eight datasets with 512-step histories, historical absolute cross-channel correlation is associated with a larger late-over-early level calibration benefit (Spearman rank correlation is 0.67). Omitting each dataset in turn preserves the positive association, so its direction does not depend on one dataset.

We believe these experiments give a clean and accurate account of how our recursive inference improves forecasting.

Reviewer_w4ZxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Recursive Inference (RI), a method for allocating additional computation within a frozen time-series forecasting model. It defines compute programs that specify which consecutive encoder blocks to repeat, the number and strength of updates, and the token representations to write back. An external reasoning model proposes and revises candidate programs using feedback from historical forecasting windows. Later held-out windows determine which program to deploy, and the selected program remains fixed throughout testing. On a single frozen Chronos-2 checkpoint, the paper reports a mean test MSE reduction of 4.02% across 56 tasks from 14 datasets, with a mean deployment cost of 1.28 times the baseline number of encoder-block evaluations per forecast.

Critical strengths

The method reports forecasting gains without updating the pretrained forecaster’s parameters. Its evaluation across 56 tasks from 14 datasets provides evidence that modifying the execution of a frozen Chronos-2 model can be useful across datasets. This gives the approach practical potential, although the concerns below limit my assessment of its contribution.

Critical weaknesses

1.Missing Several Necessary Experiments

a. The necessity of using an LLM needs to be explained and demonstrated. Could conventional search achieve the same gains? The test results primarily compare the full RI method with Base. Although the method delivers gains, I believe comparisons with conventional search methods may be missing to demonstrate the necessity of using LLM-based search.

b. All experiments in the paper use the same Chronos-2 checkpoint. Therefore, the current experiments support only effectiveness across datasets within Chronos-2, which cannot be directly equated with effectiveness across TSFM architectures. Adding a second foundation model would substantially strengthen the contribution.

2.Substantial Room for Improvement in the Paper’s Writing

The paper’s writing has considerable room for improvement, and its logical flow is not very coherent. For example, the abstract does not clearly introduce the problem, and the proposed method is not adequately connected to the problem presented. The organization of the abstract also makes it difficult to understand. I recommend making substantial improvements to the paper’s writing quality.

Questions for authors
  1. The idea of “freezing consecutive layers + repeated execution + damping + configuration selection” has a clear precedent in Training-Free Looped Transformers. The authors may need to focus on clarifying the novelty of their work.

  2. The necessity of using an LLM needs to be explained and demonstrated. Could conventional search achieve the same gains? The test results primarily compare the full RI method with Base. Although the method delivers gains, I believe comparisons with conventional search methods may be missing to demonstrate the necessity of using LLM-based search.

  3. All experiments in the paper use the same Chronos-2 checkpoint. Therefore, the current experiments support only effectiveness across datasets within Chronos-2, which cannot be directly equated with effectiveness across TSFM architectures. Adding a second foundation model would substantially strengthen the contribution.

  4. The paper’s writing has considerable room for improvement, and its logical flow is not very coherent. For example, the abstract does not clearly introduce the problem, and the proposed method is not adequately connected to the problem presented. The organization of the abstract also makes it difficult to understand. I recommend making substantial improvements to the paper’s writing quality.

4 comments

We sincerely thank Reviewer w4Zx for the thoughtful comments and constructive suggestions, which have helped us. We address the four points below.

W1: Comparison with conventional search methods.

This is a very reasonable question! We have conducted the conventional-search comparisons as suggested. The results show that LLM-guided recipe generation achieves much larger forecasting gains than random search and three quite well established adaptive search methods under the same candidate-evaluation budget. These experiments directly help us check the contribution of LLM-guided candidate generation but keeping the forecaster, recipe space, and final selection procedure fixed.

We evaluated Chronos-2 on 60 tasks spanning 15 datasets and 4 horizons per dataset. All methods search the same allowable recipe parameters (including block range, repetition count, solver, and update strength). Each method gets a budget of 18 candidate evaluations and uses the same validation selector and test protocol. Uniform random search samples recipe parameters with no feedback. Tree-structured Parzen Estimator (TPE) [1] models previous evaluations to propose promising configurations. Differential evolution (DE) [2] mutates and recombines parameter vectors. Monte Carlo tree search (MCTS) [3] explores a tree of recipe choices using Upper Confidence Bounds applied to Trees (UCT) to balance exploration/exploitation. The adaptive searches get their respective training MSE feedback. The LLM-guided workflow additionally uses model knowledge and task information to construct and revise proposals.

The performance differences are quite meaningful. Sol reduces average MSE by 4.15% and improves 53/60 tasks, compared with 1.97% and 33/60 for MCTS (strongest conventional search baseline by average MSE). Sol also reduces average MAE by 0.94%, whereas all four search baselines increase it. We note that the final validation selector is identical across methods, so these improvements are not due to a different deployment-selection procedure. We believe the results support the advantage of LLM-guided proposal generation over the conventional searches tested under this limited evaluation budget.

It was interesting/reassuring that useful gains also extend to a smaller, locally deployed reasoner. With the same model knowledge, task information, and iterative feedback workflow, frozen Qwen3-4B-Thinking-2507 reduces average MSE by 3.00% and MAE by 0.71%, improving MSE on 48/60 tasks. Its advantage over the tested searches provides additional evidence that effective recipe generation is attainable beyond the default API reasoner. Please see table below.

[1] Bergstra et al. Algorithms for Hyper-Parameter Optimization. NeurIPS, 2011.

[2] Storn and Price. Differential Evolution: A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces. Journal of Global Optimization, 1997.

[3] Kocsis and Szepesvári. Bandit Based Monte-Carlo Planning. ECML, 2006.

Table 1: Local and API reasoners versus conventional search for Chronos-2 looping recipes. All methods are evaluated on 60 tasks. B/T/W = Better/Tie/Worse. API models use low reasoning effort.

TypeModelMSE Avg Δ% ↓MSE B/T/WMAE Avg Δ% ↓MAE B/T/W
LocalQwen4B-3.00%48/7/5-0.71%36/7/17
LocalDeepSeek8B-1.42%50/7/3+0.15%37/7/16
APILuna-low-3.36%46/3/11-0.72%40/3/17
APISol-low-4.15%53/4/3-0.94%48/4/8
APITerra-low-3.78%49/7/4-0.80%40/7/13
SearchUniform+0.001%17/29/14+1.41%5/29/26
SearchTPE-1.14%28/23/9+0.69%16/23/21
SearchDE-0.43%25/23/12+1.14%15/23/22
SearchMCTS-1.97%33/18/9+0.54%19/18/23

W2: Evaluate the pipeline on an additional TSFM architecture.

We thank the reviewer for this suggestion (consistent also with Reviewer VmAH). We additionally evaluated our pipeline on Timestral-209M to assess its generalizability. This model also benefits from our pipeline with Claude Haiku 5.5.

Timestral gives an architectural contrast to Chronos-2. Chronos-2 uses bidirectional context encoding, Timestral encodes context causally and uses separate prediction tokens that attend to the context, while context tokens do not attend back to them [1]. The transfer preserves the core recipe proposal and feedback workflow while using Timestral’s native computation. The reasoner gets model-specific knowledge (like the block structure) and task information (like the forecasting horizon). It proposes looping recipes and refines subsequent proposals using forecasting feedback from training data. Timestral’s complete Transformer blocks serve as the looping units, keeping their native attention and feed-forward computations. All forecaster weights remain frozen. So, transferring the framework does not need reproducing Chronos-2's particular token-update mechanism or information flow.

Across the reported 12 datasets covering 48 tasks, RI with Haiku achieves mean dataset-level reductions of 5.97% in MSE and 1.08% in MAE, after averaging four horizons within each dataset. These results provide evidence that the core workflow can improve a forecaster beyond the Chronos-2 checkpoint and architecture.

The Timestral gains are also present with a second API reasoner on the eight-task comparison. Under the same workflow and candidate budget, average MSE/MAE reductions are 13.77%/2.53% with Haiku and 14.52%/3.00% with Sonnet. This supplements the second-forecaster results by showing that useful recipe generation on Timestral is attainable with both tested reasoners.

[1] Petralia et al. Timestral. arXiv:2610.37585, 2026.

Table 2. Generalization to Timestral-209M with Claude Haiku 5.5. Results cover 12 datasets and 48 forecasting tasks. MSE and MAE are averaged over four horizons within each dataset. The Average row reports the mean dataset-level percentage change. Δ denotes percentage change relative to Base. Negative values indicate lower forecasting error.

DatasetMSE↓ (Base→RI)Δ MSEMAE↓ (Base→RI)Δ MAE
Covid-192.793→1.463-47.63%0.041→0.039-3.24%
ETTh10.400→0.4000.00%0.402→0.4020.00%
ETTh20.344→0.340-1.15%0.369→0.370+0.10%
ETTm10.378→0.356-5.83%0.362→0.358-1.16%
ETTm20.269→0.267-0.46%0.302→0.303+0.56%
ILI1.595→1.592-0.20%0.765→0.7650.00%
NASDAQ0.933→0.913-2.13%0.650→0.648-0.35%
NN50.673→0.672-0.10%0.548→0.548-0.01%
NYSE0.442→0.436-1.26%0.410→0.405-1.40%
PEMS080.495→0.437-11.74%0.438→0.404-7.84%
Wind1.146→1.143-0.26%0.735→0.735-0.03%
ZafNoo0.521→0.517-0.84%0.409→0.411+0.35%
Average—-5.97%—-1.08%

W3: Writing.

We agree that in the abstract we should clearly state the main problem. We have adjusted the abstract that describes to the reader, the compute-allocation problem first, explains our main proposal and how programs are selected. And only then we move to the experimental results. Abstract is now as follows.

Pretrained time-series foundation models typically execute the same sequence of blocks once, independent of what the forecasting task is. But different tasks may benefit from spending additional computation, possibly at different locations within a frozen model. We investigate here how a limited inference budget can be allocated to improve forecasting without updating model parameters. We define compute programs that revisit selected consecutive blocks, control the number and size of repeated updates, and specify which token representations are modified. Their compute cost is known in advance, but their forecasting benefit must be measured. Recursive inference uses a reasoning model to propose programs, evaluate their effects on historical forecasting windows, and revise subsequent proposals using the measured results. Later held-out windows can determine whether a program should be deployed, after which it will remain fixed throughout testing. [Experimental findings we will state after this]

We will also make the same progression explicit in the main text: Sec. 2 will define the compute programs and their costs. Sec. 3 explains how the reasoner proposes and revises programs using measured responses. This is followed by independent verification and deployment. Sec. 4 then evaluates the resulting allocations and examines what repeated computation changes within the forecaster. We will strengthen the transitions between these sections.

All these changes are localized and will improve the presentation!

Q1: Clarify the novelty.

Yes, your point is well taken. It is true that repeating frozen Transformer blocks, damping their updates, and selecting looping configurations are not new to our paper. We also do not claim these operations as our novel insights. The central contribution of our work is different: we study how an external reasoning model can allocate computation within a separate, frozen forecasting model, using task-specific information and measured forecasting outcomes to decide which execution strategy to deploy.

The paper Training-Free Looped Transformers paper describes how repeated execution can improve a pretrained language model. This is a really nice result. But in our framework, the LLM instead acts as a reasoner that designs and revises the internal execution of a separate numerical forecaster. The difference is not simply the use of looping, but the introduction of a task-level, feedback-driven process for deciding how the forecaster should actually do compute.

We hope you will agree that the reasoner combines model knowledge with temporal characteristics, such as autocorrelation, recent scale changes, and forecasting horizon, to propose where to repeat blocks, how often, and with what update strength. The execution choices also reflect different roles of observed history and future predictions. In Chronos-2, the reasoner can propose updating their representations jointly or refining forecast representations but holding historical representations at a fixed reference between repetitions. TSFM checks these recipes on historical forecasting windows. Resulting errors inform the reasoner's later proposals. Then, chronologically held-out verification windows independently check whether a program should be deployed. After this, the selected execution remains fixed throughout testing.

Our new experiments also show that the choice of how to generate these programs matters. Across 60 Chronos-2 tasks, our Sol-based pipeline reduces MSE by 4.15% on average, compared with 1.97% for MCTS (strongest of four conventional search methods that we evaluated under the same candidate budget and final selection rule). And now, experiments with Timestral and Haiku further show that this framework extends beyond the original model combination.

So, yes we build on the established mechanism of training-free looping, but address a different problem of how to use task information, measured responses, and a limited evaluation budget to obtain an effective execution strategy for a frozen forecasting model.

Dear Reviewer w4Zx,

Thank you again for your time and feedback. We wanted to kindly follow up on our responses above. We have added comparisons with conventional search methods and experiments on a second TSFM. The results show that our approach achieves larger forecasting gains than the tested conventional searches under the same evaluation budget, and that the improvements extend to a different TSFM architecture. We have also clarified how our work differs from prior looping methods and revised the abstract to better explain the problem and our method. We hope these additions help address your main concerns. We would be grateful for your consideration of our responses.

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

This work proposes recursive inference, a test-time compute allocation framework for frozen time-series foundation models. Instead of running each encoder block once, it defines compute programs that revisit a selected block window with damped updates and token-scoped write-backs at a known cost, and uses a reasoning LLM to propose, revise, and select one program per task from measured responses on historical windows, with held-out later windows deciding deployment.

The evaluation is reasonably broad and the gains are solid: 56 tasks across 14 datasets show an average 4.02% reduction in test MSE at 1.28x baseline compute, with improvements on 48/56 tasks, plus ablations confirming that measured responses improve proposals and verification tracks test performance.

Critical strengths

The overall idea is sound and well-motivated: using an LLM reasoner to identify hard regions and re-execute selected frozen blocks with damped, token-scoped revisits is a reasonable and elegant way to spend extra test-time compute without touching model parameters.

The empirical results are generally broad and promising. The evaluation spans 56 tasks across 14 diverse datasets with proper train/verification/test separation and dataset-level bootstrapped CIs, showing a consistent 4.02% MSE reduction, while the additional deployment cost is modest at only 1.28x baseline block evaluations.

Critical weaknesses
  1. All main results use one frozen forecaster (Chronos-2) and one reasoner (GPT-5.6 Sol). For a pipeline paper, this leaves open whether the gains come from the framework or from this specific combination. At least a subset-level validation on another TSFM (e.g., TimesFM 3.0) is needed to show robustness of this pipeline.

  2. Table 12 shows low vs. high reasoning effort perform nearly identically (-3.30% vs. -3.35%), which raises two competing explanations: (1) weak reasoning ability is already enough, or (2) no real reasoning is needed and the LLM acts only as a weak selector. The with-vs-without-responses study (Fig. 5) is a good start but does not separate these. I would suggest:

(1) An ablation with a smaller/older/local model (e.g., a small Qwen that can be deployed locally, which also mitigates the privacy concern of sending data statistics to a closed API) and/or a non-LLM search baseline would clarify this.

(2) A case study showing whether the pipeline allocates more effort where it should, e.g., non-stationary vs. stationary segments. I would expect the pipeline revisits more on non-stationary segments, the performance gain are largely relies on improving predictions on these segments, which would strengthen the claim that the reasoner truly "reasons about how it computes."

Questions for authors

Please refer to the weakness

4 comments

We sincerely thank Reviewer VmaH for the detailed and constructive feedback!

W1: Validate the framework on an additional TSFM.

We have now conducted the cross-model experiment as suggested. We transferred our pipeline to a second model that just came out, Timestral-209M, using Claude Haiku 5.5 as the reasoner (Petralia et al., 2026). While Chronos-2 uses bidirectional context encoding, Timestral encodes the historical context causally and also separates prediction tokens from context representations. This can help us see if RI transfers to a model with different information-flow and attention mechanisms, and also tests whether the gains extend beyond the original Chronos-2/Sol combination.

The implementation uses Timestral's native Transformer blocks. The reasoner receives model-specific knowledge (such as block structure) and task information (such as forecast horizon). It proposes looping recipes and uses feedback from their evaluation on training data to guide subsequent proposals. Timestral's complete Transformer blocks serve directly as looping units. They keep their native attention and feed-forward setup as is. All forecaster weights remain frozen. Thus, the core proposal-and-feedback workflow transfers without requiring Chronos-2's particular block structure or token-update mechanism.

The transferred pipeline finds strong/beneficial recipes across multiple datasets, with gains that vary by task. On the reported 12-dataset subset covering 48 tasks, RI with Haiku achieves average relative reductions of 5.97% in MSE and 1.08% in MAE, after averaging four horizons within each dataset. Importantly, MSE improves on 11 of the 12 datasets. Even excluding Covid-19, which has a particularly large improvement, the mean MSE reduction remains 2.18%. These results show that the transferred workflow identifies useful execution changes for a forecaster with a different block structure.

[1] Petralia et al. Timestral. arXiv:2610.37585, 2026.

Table 1. Generalization to Timestral-209M with Claude Haiku 5.5. Results cover 12 datasets and 48 forecasting tasks. MSE and MAE are averaged over 4 horizons within each dataset. The Average row reports the mean dataset-level percentage change. Δ denotes percentage change relative to Base. Negative values indicate lower forecasting error.

DatasetMSE↓ (Base→RI)Δ MSEMAE↓ (Base→RI)Δ MAE
Covid-192.793→1.463-47.63%0.041→0.039-3.24%
ETTh10.400→0.4000.00%0.402→0.4020.00%
ETTh20.344→0.340-1.15%0.369→0.370+0.10%
ETTm10.378→0.356-5.83%0.362→0.358-1.16%
ETTm20.269→0.267-0.46%0.302→0.303+0.56%
ILI1.595→1.592-0.20%0.765→0.7650.00%
NASDAQ0.933→0.913-2.13%0.650→0.648-0.35%
NN50.673→0.672-0.10%0.548→0.548-0.01%
NYSE0.442→0.436-1.26%0.410→0.405-1.40%
PEMS080.495→0.437-11.74%0.438→0.404-7.84%
Wind1.146→1.143-0.26%0.735→0.735-0.03%
ZafNoo0.521→0.517-0.84%0.409→0.411+0.35%
Average—-5.97%—-1.08%

The forecasting benefit is also present with a second API reasoner on the tested subset. Under the same workflow and candidate budget, the eight-task comparison gives average MSE/MAE reductions of 13.77%/2.53% with Haiku and 14.52%/3.00% with Sonnet. So, both APIs support effective recipe search for Timestral, providing us experimental evidence of portability across reasoner choices and also forecasters.

Table 2. Comparison of Haiku and Sonnet API reasoners on Timestral. Results use the same pipeline and budget. H denotes the forecasting horizon. The Average row reports the mean task-level percentage change relative to Base. Δ denotes percentage change relative to Base. Negative values indicate lower forecasting error.

DatasetHMSE↓ BaseHaiku (Δ%)Sonnet (Δ%)MAE↓ BaseHaiku (Δ%)Sonnet (Δ%)
Covid-19241.71411.1007 (-35.79%)1.1007 (-35.79%)0.02940.0286 (-2.50%)0.0286 (-2.50%)
Covid-19483.14041.5533 (-50.54%)1.5224 (-51.52%)0.04420.0423 (-4.29%)0.0420 (-5.09%)
ETTh1960.35440.3541 (-0.08%)0.3507 (-1.03%)0.36730.3691 (+0.50%)0.3683 (+0.29%)
ETTh13360.41920.4223 (+0.73%)0.4159 (-0.81%)0.41190.4116 (-0.08%)0.4159 (+0.96%)
PEMS08960.32920.2959 (-10.12%)0.2875 (-12.67%)0.35170.3403 (-3.23%)0.3268 (-7.07%)
PEMS083360.55090.4791 (-13.04%)0.4791 (-13.04%)0.45990.4133 (-10.12%)0.4133 (-10.12%)
Wind960.87850.8785 (0.00%)0.8785 (0.00%)0.59790.5979 (0.00%)0.5979 (0.00%)
Wind7201.37741.3596 (-1.29%)1.3596 (-1.29%)0.84660.8424 (-0.50%)0.8424 (-0.50%)

Overall, we hope that we have addressed the main concern if the gains are limited to the original Chronos-2/Sol combination. Thanks.

W2: Evaluate a local LLM and a non-LLM search baseline.

We have now conducted both experiments suggested in the review. We replaced the API reasoner with smaller, locally deployed models and compared LLM-guided proposals against conventional search methods. We note that similar performance at low and high reasoning effort does not tell us that the LLM-guided proposal process is unnecessary. Instead we can check whether it finds better programs than conventional search under the same evaluation budget.

Our pipeline supports both API and locally deployed reasoners. We replace the API reasoner with locally hosted Qwen3-4B-Thinking-2507 and DeepSeek-R1-0528-Qwen3-8B, retaining the model knowledge, task information, and iterative training feedback used by the pipeline. These models directly propose recipe parameters, with the local reasoner and forecaster weights frozen. We evaluate on Chronos-2 across 15 datasets and 4 horizons per dataset, totaling 60 tasks. Qwen reduces average MSE by 3.00% and MAE by 0.71%, improving MSE on 48 tasks.

We find that Sol and Qwen find more effective recipes than the tested conventional searches under the same candidate-evaluation budget. All methods search the same allowed recipe parameters (includes block range, repetition count, solver, update strength). Uniform random search in this case samples parameters but without using feedback. The Tree-structured Parzen Estimator (TPE) [1] models previously evaluated candidates to suggest further promising parameters. Differential evolution (DE) [2] mutates and recombines parameter vectors. Monte Carlo tree search (MCTS) [3] explores a tree of parameter choices, using Upper Confidence Bounds applied to Trees (UCT) to balance exploration of new branches with exploitation of promising ones. The adaptive searches receive their respective training MSE feedback. The main point is that every method receives the same budget of 18 candidate evaluations and uses the same final validation selector and test protocol.

The differences are quite clear. Sol reduces average MSE by 4.15% and improves 53/60 tasks, compared with 1.97% and 33/60 for MCTS (strongest search baseline by average MSE). Sol and Qwen also reduce average MAE, whereas all four search baselines end up increasing it. Since the recipe space, evaluation budget, and final selector are held fixed, these results give good evidence that LLM-guided proposal generation offers an advantage over the conventional search methods tested here.

Reasoners differ in how effectively they use the same proposal budget. Sol achieves the strongest average results, while Terra and Luna also outperform the local models on average MSE. Qwen nevertheless improves both metrics and outperforms DeepSeek, showing that parameter count alone will not determine recipe-generation effectiveness. The results also address the concern about dependence on a powerful closed API: a locally deployed 4B reasoner already gives meaningful gains and outperforms the conventional searches we tested. API models give higher observed forecasting accuracy, while local models can allow self-hosted deployment.

[1] Bergstra et al. Algorithms for Hyper-Parameter Optimization. NeurIPS, 2011.

[2] Storn and Price. Differential Evolution: A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces. Journal of Global Optimization, 1997.

[3] Kocsis and Szepesvári. Bandit Based Monte-Carlo Planning. ECML, 2006.

Table 3: Local and API reasoners versus conventional search for Chronos-2 looping recipes. All methods are evaluated on 60 tasks. B/T/W = Better/Tie/Worse. API models use low reasoning effort.

TypeModelMSE Avg Δ% ↓MSE B/T/WMAE Avg Δ% ↓MAE B/T/W
LocalQwen4B-3.00%48/7/5-0.71%36/7/17
LocalDeepSeek8B-1.42%50/7/3+0.15%37/7/16
APILuna-low-3.36%46/3/11-0.72%40/3/17
APISol-low-4.15%53/4/3-0.94%48/4/8
APITerra-low-3.78%49/7/4-0.80%40/7/13
SearchUniform+0.001%17/29/14+1.41%5/29/26
SearchTPE-1.14%28/23/9+0.69%16/23/21
SearchDE-0.43%25/23/12+1.14%15/23/22
SearchMCTS-1.97%33/18/9+0.54%19/18/23

W3: Case study.

Good point! Based on this comment, we checked the main hypothesis about tasks showing greater nonstationarity should get more computation and maybe benefit more from recursive inference. Since RI selects a fixed program for each forecasting task, we can study this relationship at the task level instead of varying compute across individual test segments. The results we get support this hypothesis, so thanks for asking us this question.

We found that at the task level the pipeline tends to allocate more looping compute to tasks with greater historical scale changes, and these tasks also tend to achieve larger forecasting gains. We analyze the paper’s 56 Sol-selected Chronos-2 tasks, averaging four horizons within each of 14 datasets. Variability is measured by median absolute deviation (MAD) from the segment median. Scale change is the absolute log ratio of MAD between two recent segments of a training history. Its Spearman correlations are 0.486 with additional encoder-block executions, 0.310 with full-test MSE reduction and 0.411 with MAE reduction relative to Base. Looping tends to help us get larger gains on tasks with greater historical scale changes, and the pipeline also tends to allocate more compute to them.

NYSE illustrates beneficial looping on nonstationary data. Standard ADF and KPSS diagnostics support unit-root nonstationarity in all five channels. Selected recipes average 17.5 block executions versus 12 for Base and reduce full-test MSE by 5.10%, averaged across four horizons.

The reasoner can use training feedback to expand its proposed computation. On the Weather dataset, for forecasts 720 steps ahead, it proposed one additional pass through the last three encoder blocks, increasing total block executions from 15 to 18. The validation stage selected this recipe, which reduced full-test MSE by 14.33% relative to Base.

Based on your suggestion, these results have strengthened the paper by showing that the computation selected by RI is meaningfully associated with task characteristics and forecasting benefits.

Dear Reviewer VmaH,

Thank you again for your positive assessment of our work and your suggestions. We wanted to kindly follow up on our responses above. We have added experiments on a second TSFM, evaluated locally deployed reasoners, and compared our approach with conventional search methods. We also examined how the selected computation relates to task characteristics and forecasting gains, as you suggested. We believe these new results provide further support for our approach and hope they help address your main concerns. We would be grateful for your consideration of these additions.

Comments on the paper

1 comment

Thank you for reading our response. We would be happy to provide any further clarification if helpful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.