acceptodds
Under review as a conference paper at ICLR 2027

OpenTSLM TeeMoE: A Unified Time-Series Language Model for Forecasting, Contextual Prediction, and Reasoning

13 views

Abstract

Real-world time-series applications increasingly require models that can handle time series forecasting, context-conditioned prediction, and language-based temporal reasoning. Yet current time-series foundation models remain fragmented across these capabilities: numerical specialists often provide the strongest forecasts, while language-based models offer broader contextual understanding and analysis. A central challenge is to unify these heterogeneous capabilities without reducing their individual performance. We introduce OpenTSLM TeeMoE, a generalist time-series language model that can forecast directly from observed time series, reason over textual context and temporal patterns, and synthesize and refine predictions from external numerical forecasting specialists. We independently train three low-rank experts for forecast aggregation, native forecasting, and temporal analysis over a shared backbone. A learned LoRA mixture-of-experts controller then weights their frozen parameter updates for each request. Our proposed model achieves strong performance on widely used benchmarks for time series forecasting, context-conditioned prediction, and language-based temporal reasoning, ranking among the top three on GIFT-Eval by mean MASE rank, Context is Key by RCRPS, and TimeSeriesExam by accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Language Models, TSLM, Multimodal LLM, Time Series Foundation Models, TSFM

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

The paper introduces OpenTSLM TeeMoE, a shared language backbone with independently trained LoRA experts for forecast aggregation, native contextual forecasting, and temporal analysis. A controller learns request-level mixtures of the frozen adapters and selects a numerical or language output path. The aggregation path refines a learned ensemble of 13 numerical forecasters, while the text path generates forecast trajectories or answers analytical questions. The model is evaluated on GIFT-Eval, Context is Key, and TimeSeriesExam, with controls for individual experts, adapter composition, joint training, and numerical ensembling, plus training-cost accounting.

Critical strengths

The paper evaluates a single composed checkpoint across three substantially different capabilities and addresses the main question of whether broad coverage dilutes specialist performance.

The authors performed a good amount of ablation studies such as separating the numerical ensemble from the editor to understand the gain of the proposed model.

Critical weaknesses

The joint training comparison does not seem fair. The three separate adapters together have more parameter count and hence probably more capacity than the single adapter. the size of adapters in the two settings are not matched so its hard to understand the benefit of the three separate adapters.

The evidence for a benefit from soft composition beyond selecting a specialist is weak outside Context is Key benchmarks. For the other two benchmarks, selecting the best expert seem well sufficient and the training cost would also possibly be cheaper if we only train that one adapter?

Questions for authors

in figure 1, it's a little confusing that the contextual forecasting uses the language model head since the output is numerical. why is that? the clarity can be improved.

quite a few table entries are OOT, this makes it hard to compare. there is no TSLM selected for comparison for the GIFT-EVAL task. and for table 2 GIFT-EVAL comparison, two experts have OOT. why is using all three experts ok but a single expert becomes OOT. and the only available expert (aggregate expert) performs very similar to your method in that case.

5 comments
AuthorsOpenReview ↗

Response to Review (Part 1 of 3)

Weakness 1. Parameter-matched joint training

We thank the reviewer for identifying the parameter-count difference. To address it directly, we provide a shared rank-52 joint model, matching the three specialists' combined rank (4+32+16=524+32+16=52) over the same adapted matrices. Both settings use the same frozen backbone, training examples, number of training presentations, and output paths.

ModelGIFT rank ↓CiK RCRPS ↓TSE (%) ↑
Shared rank-52 joint19.9790.1185777.748
TeeMoE (4+32+164+32+16)19.9900.1147978.552

With adapter parameters matched, TeeMoE achieves 3.19% lower CiK RCRPS and 0.80 percentage points higher TSE accuracy, while joint training obtains a slightly better GIFT mean rank. TeeMoE thus remains competitive on numerical forecasting and performs better on the two language-based tasks without an advantage in total LoRA parameter count.

The comparison also retains a practical benefit: TeeMoE requires 30.756 H100 GPU-hours of training versus 58.657 for the joint model, a 47.6% reduction under these training recipes. Once trained, an individual expert can also be updated without retraining the other two. Together, these results support separate specialization as an effective way to build the generalist at matched adapter size.

We have replaced the rank-32 joint baseline in Section 5.3.2 with this parameter-matched rank-52 model and updated the joint-training appendix and training-cost table accordingly.

Weakness 2. Value of composition and specialist retention

We thank the reviewer for this observation. For a user who needs only numerical forecasting or only temporal analysis, training the corresponding expert is indeed a lower-cost option. Our GIFT and TSE results do not suggest that all three experts are necessary for either task alone. TeeMoE addresses the broader need for one model that performs strongly across numerical forecasting, contextual forecasting, and temporal analysis. Covering complementary time-series capabilities also motivates UniTS [2], TimeOmni-1 [3], TimeOmni-VL [4], and ChatTS [5]. For example, a demand-planning workflow may require a baseline forecast, a forecast conditioned on an announced promotion, and analysis of recent trends. TeeMoE's top-three results across the three benchmarks, together with the specialist comparisons, show that this broader coverage can be achieved while retaining the experts' performance.

Within that generalist, the reviewer is right that top-1 selection already preserves most of the GIFT and TSE performance. We do not claim a substantial soft-mixing gain on those two benchmarks. MoE lets us develop capabilities with different training requirements separately and combine them in a shared backbone; however, the contribution is the resulting generalist model and training recipe, rather than a new MoE framework whose principal claim is additional gains from expert interaction. We therefore believe that preserving strong specialist performance across the combined capabilities is valuable even where mixing does not improve each specialist further.

CiK indeed shows a clear additional benefit: soft composition reduces RCRPS from 0.12296 to 0.11479, a 6.65% reduction. To assess the consistency of this gain, we performed a paired bootstrap analysis of the existing predictions for all 355 instances across 71 task types. Each resample draws task types with replacement, keeping their five paired instances together and retaining the benchmark weights. Across 100,000 resamples, the 95% interval for the absolute reduction is [0.00211, 0.01482], entirely above zero. The improvement also remains positive when any single task type is omitted.

We have revised Section 5.3.2 to distinguish these findings: strong performance and specialist retention across all three capabilities, with additional predictive gains from composition most clearly supported on CiK. The paired analysis is included in the appendix. We appreciate the suggestion to make the value of the generalist and the more specific benefit of soft mixing clearer.

AuthorsOpenReview ↗

Response to Review (Part 2 of 3)

Q1. Numerical forecasts from the language-model head

We thank the reviewer for pointing out this ambiguity. We retain language-model generation for contextual forecasting to use the LLM's ability to interpret textual information when producing forecast values. This choice is supported by CiK, where several LLMs generating forecasts as text outperform the evaluated LLM-based models with numerical forecasting heads, such as UniTime and Time-LLM [1, Section 5.3]. Our expert thus generates timestamp–value pairs as tokens, which we parse into numerical trajectories.

By contrast, for numerical forecasting, strong forecasts are already available from TSFMs so the numerical decoder lets TeeMoE refine their quantiles directly, without generating new trajectories token by token and incurring the associated cost of autoregressive generation. We ultimately preserve the prediction mechanism suited to each capability rather than force both through the same output head.

Figure 1 now labels the language-model head's outputs as “Forecasts and analytical answers,” and Section 3.3 explains this rationale and how generated timestamp–value pairs form numerical trajectories.

Q2. Missing GIFT comparisons and OOT entries

We thank the reviewer for raising this comparison gap. The selected public TSLMs were evaluated through their documented forecasting interfaces, but their measured runtimes project (far) beyond the stated 1,000 H100 GPU-hour budget per model on full GIFT-Eval (Appendix E.1). At 371,330 windows and 25 samples per window, each model must generate over nine million trajectories, which was not feasible given our compute. The same bottleneck applies to our standalone native and analysis experts, which generate forecasts autoregressively and also exceed this budget.

The full TeeMoE model avoids this cost by selecting the numerical decoder for GIFT requests and refining TSFM forecasts directly. Combining three adapters within one backbone does not require running three separate forecasting procedures, so the composed model can complete the evaluation even when the two standalone language-based experts cannot.

The aggregation expert is trained specifically for numerical forecasting so its strong GIFT performance is expected and provides the specialist reference we aim to retain. Matching that performance after composition shows that adding contextual forecasting and analysis does not sacrifice this capability. The additional experts extend what the model can do, rather than being intended necessarily to improve the aggregation expert on its own task.

To fill the missing comparisons, we have now evaluated all four selected public TSLMs and both previously OOT experts on a common GIFT subset, alongside TeeMoE and its numerical baselines. We selected one window from each of the 97 dataset/frequency/horizon categories without using prediction quality, preserving the original forecast horizons. Each language-based model generates five forecast samples per window, and numerical models use their quantile forecasts on the same windows.

ModelMean MASE rank ↓MASE ↓CRPS ↓
OpenTSLM21.1677.93931.9711
TimeOmni-VL19.8965.65141.0529
TS-Reasoner18.7341.55450.4239
ChatTS16.2760.83900.4010
Native expert15.2140.61630.1904
Analysis expert14.6200.58500.1857
Aggregation expert6.8230.57360.1062
TeeMoE6.8130.57350.1062

Mean MASE rank is computed within a common 22-system roster that includes the 13 constituent TSFMs, so these subset ranks are separate from the full-benchmark leaderboard ranks.

Both standalone language-based experts outperform the four external TSLMs on all three subset metrics. The analysis expert is particularly close to TeeMoE in geometric MASE (0.5850 versus 0.5735), although its CRPS and mean rank are substantially worse. TeeMoE retains the aggregation expert's performance and obtains the lowest geometric MASE and CRPS in the full subset roster. These results fill the OOT comparison gap while showing why retaining the numerical forecasting path is useful. We have added the subset comparison to Appendix E.1 and linked it from the captions of Tables 1–2.

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

The paper introduces a time-series language model built from three independently trained LoRA experts for forecast aggregation, native forecasting, and temporal analysis. A request-level controller mixes their frozen parameter updates and selects numerical or textual output; numerical forecasting also uses an ensemble of external forecasting models with learned refinement. In the reported GIFT-Eval, Context is Key, and TimeSeriesExam comparisons, the unified system ranks among the top three and largely retains the individual experts’ performance.

Critical strengths
  1. The method is well motivated by the practical need to unify numerical forecasting, context-conditioned prediction, and temporal analysis while retaining the strengths of specialized models.

  2. The proposed method achieves competitive results across the reported GIFT-Eval, Context is Key, and TimeSeriesExam comparisons, demonstrating broad task coverage while largely preserving the performance of the individual specialists.

Critical weaknesses
  1. The contribution of the unified architecture could be articulated more precisely at the methodological level. The paper credits LoRA-library composition [1], MoLE [2], and X-LoRA [3], and describes its request-level shared weights and dual output paths. Building on these acknowledged connections, a clearer account of which design choices address difficulties specific to combining numerical forecasting and language-based analysis would help establish the significance of the adaptation.

  2. The additional predictive benefit of soft composition is supported less strongly than capability retention. The paper acknowledges that top-1 routing preserves most of the gains, and the TSE difference amounts to one additional correct answer. The larger CiK improvement is more encouraging, although its robustness across evaluated instances is less clear. A focused analysis of the existing CiK predictions would help establish the strength of this benefit.

[1] Oleksiy Ostapenko, et al. Towards Modular LLMs by Building and Reusing a Library of LoRAs. 2024.
[2] Xun Wu, et al. Mixture of LoRA Experts, 2024.
[3] Eric L. Buehler et al. X-LoRA: Mixture of Low-Rank Adapter Experts, a Flexible Framework for Large Language Models with Applications in Protein Mechanics and Molecular Design. 2024.

Questions for authors
  1. What is the key methodological insight in adapting existing frozen-LoRA composition approaches [1–3] to these three time-series capabilities? Beyond the architectural differences already described, it would be helpful to explain the time-series-specific considerations behind request-level shared weighting and numerical-versus-textual output selection. A conceptual explanation of these choices and their trade-offs would help distinguish the methodological contribution from the breadth of the integrated system.

  2. How consistent is the CiK improvement from soft composition over top-1 routing across the evaluated instances? If the individual predictions are available, could you provide a paired uncertainty analysis that accounts for instances sharing a task type? This would help assess whether the aggregate difference provides reliable evidence of an additional benefit from mixing experts. The request concerns analysis of existing outputs and does not require additional model training.

2 comments
AuthorsOpenReview ↗

Response to Review (Part 1 of 2)

Weakness 1 and Question 1. Design rationale and routing granularity

We thank the reviewer for this helpful suggestion. We have expanded Sections 3.1 and 3.5 to clarify the following design rationale, also addressing question 1.

The key adaptation is to compose experts within a shared backbone while preserving the prediction mechanism suited to each capability.

For aggregation, since a strong TSFM forecast is already available, the numerical path lets the language model refine it directly rather than incur the cost of generating multiple new trajectories as text. The numerical ablation demonstrates the benefit: adding the aggregation expert improves mean MASE rank from 20.660 to 20.000 while holding the reference ensemble fixed. For contextual forecasting, however, we retain language-native generation; CiK finds that several LLMs generating forecasts as text outperform the tested LLM-based models with numerical forecasting heads, such as UniTime and Time-LLM [4, Section 5.3].

Accordingly, while ensemble refinement operates on numerical forecasts, contextual forecasting and analysis use autoregressive text generation. To accommodate these different inputs and outputs within the composed model, we introduce a controller-selected numerical encoder–decoder path alongside the language-model head. The numerical path is selected when the aggregation weight exceeds 0.5, and token generation is used otherwise. This places numerical forecasting on the same footing as the other capabilities: all three are learned through adapters in the same backbone and participate in the same composition mechanism.

For routing, we choose the request as the unit of selection, rather than an entire conversation or an individual token. A generalist should be able to produce a forecast for one request and analyze a series for the next; selecting a new mixture for each request lets it switch capabilities accordingly. Within each response, however, we retain the same weights to keep composition simple and avoid repeating the controller's backbone pass at every generation step. This request-level design already retains specialist-level performance in our evaluations. X-LoRA's optional cached-scaling mode uses the same computational principle to reduce inference latency [3, Section 4.3].

We also share the weights across layers, assigning one weight to each complete adapter rather than learning separate mixtures at different depths. Following the input-level Task Predictor scheme in Ostapenko et al. [1, Section 4.1.1], this reduces the number of routing parameters and keeps composition simple to implement and train.

To justify these choices empirically, we additionally fitted layerwise and tokenwise composers over the same frozen experts and evaluated them on full GIFT-Eval, CiK, and TSE:

ComposerGIFT MASE rank ↓CiK RCRPS ↓TSE accuracy ↑
TeeMoE (request-level)19.9900.1147978.552%
Layerwise20.0000.1409578.42%
Tokenwise20.0000.1144578.28%

Layerwise weighting does not improve GIFT rank and increases CiK error, and the tokenwise fit gives nearly unchanged GIFT and CiK performance. Thus, the results support retaining the simpler request-level composer rather than adding finer routing to the current model. We have added this comparison to the revised appendix subsection, “Routing granularity.”

We thank the reviewer again for helping us clarify these design choices. Sections 3.1 and 3.5 now explain the output interfaces and request-level composition design. Section 3.3 cites the CiK comparison supporting language-native forecasting, and Section 3.5 cites the input-level Task Predictor precedent.

AuthorsOpenReview ↗

Response to Review (Part 2 of 2)

Weakness 2 and Question 2. Predictive gains and CiK consistency

We thank the reviewer for this distinction and the suggested analysis. Indeed, as the reviewer observes, composition does not produce large gains across all three tasks. However, further gains from mixing were neither the primary motivation nor the sole measure of success for this work. TeeMoE's main contribution is a model that performs strongly across all three capabilities, ranking among the top three on each benchmark in Table 1. It achieves this while retaining expert performance, reducing training cost relative to our joint baseline, and allowing experts to be replaced independently. Our objective was therefore to build a strong generalist from independently trained capabilities, with composition serving that objective. Nonetheless, we agree that the limited additional gains on GIFT and TSE are a limitation, and we have now made this more explicit in Section 5.3.2.

For the larger CiK gain, we performed the requested analysis of existing predictions, also addressing question 2. TeeMoE's soft composition and top-1 routing are evaluated on the same 355 instances with 25 trajectories per instance. Soft composition reduces weighted RCRPS from 0.12296 to 0.11479, a 6.65% reduction. We assess the uncertainty in this reduction with a paired bootstrap comparison. To account for instances sharing a task type, each bootstrap sample draws 71 task types with replacement and keeps all five paired instances within each selected type. We then recompute the weighted RCRPS difference using the original instance weights and scoring cap. Across 100,000 resamples, the 95% interval for the absolute reduction is [0.00211, 0.01482], entirely above zero.

The improvement occurs in 41 of 71 task types and 216 of 355 instances, with five instance-level ties. Removing any single task type leaves the aggregate reduction positive, between 0.00634 and 0.00904. We have added these results to the new appendix subsection, “Paired CiK comparison of soft and top-1 composition,” which provides the equation, resampling procedure, and random seed, with a pointer in Section 5.3.2.

We hope the additions to Section 5.3.2 and the paired-CiK appendix address the reviewer's concern, and thank them for suggesting an analysis that strengthens the evidence for this result.

References

[1] Ostapenko et al. Towards Modular LLMs by Building and Reusing a Library of LoRAs. ICML, 2024. Section 4.1.1 (Task Predictor routing).

[2] Wu et al. Mixture of LoRA Experts. ICLR, 2024.

[3] Buehler and Buehler. X-LoRA: Mixture of Low-Rank Adapter Experts, a Flexible Framework for Large Language Models with Applications in Protein Mechanics and Molecular Design. 2024. Section 4.3 (cached scaling weights).

[4] Williams et al. Context is Key: A Benchmark for Forecasting with Essential Textual Information. ICML, 2025. Section 5.3.

Reviewer_Z3q3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes OpenTSLM TeeMoE, a generalist time-series language model designed to handle three distinct tasks within a unified framework: native numerical time-series forecasting, language-based temporal analysis/reasoning (including context-conditioned prediction), and forecast aggregation from external numerical specialists. The architecture builds upon a shared backbone by independently training three task-specific low-rank (LoRA) experts and dynamically combining their frozen parameter updates via a learned MoE controller.

Critical strengths
  1. Generalist Capability: The proposed framework expands the functional scope of time-series language models by supporting multiple heterogeneous tasks (native forecasting, temporal reasoning, and forecast aggregation) under a single routing architecture.
  2. Competitive Empirical Performance: The model achieves solid quantitative results across several benchmarks, including GIFT-Eval, Context is Key, and TimeSeriesExam.
Critical weaknesses
  1. Lack of Synergy and Unclear Motivation for Unification: Beyond the engineering goal of packaging three tasks into a single framework, the paper fails to demonstrate any intrinsic motivation or mutual benefit for combining them. Since the three LoRA experts are trained independently, there is no evidence that temporal analysis helps improve forecast aggregation, or that forecast aggregation enhances temporal analysis and contextual forecasting. Without proven cross-task synergy, the architecture feels like a mechanical stitching of disjoint components rather than a truly unified model.
  2. Unjustified Selection of a 27B Backbone and Poor Efficiency-Performance Trade-off: Existing time-series language models typically rely on ~7B parameter backbones, which generally provide sufficient capacity for temporal tasks. This work directly adopts a heavyweight 27B backbone without thorough justification or scaling analysis. The paper fails to evaluate the trade-off between model parameter size/computational cost and task difficulty, leaving it unclear whether a 27B model is necessary or if a 7B backbone could achieve comparable results.
  3. Confusing Design and Unclear Role of the LLM in the Forecast Aggregation Branch: The forecast aggregation pipeline is confusing and appears largely disconnected from the core capabilities of a frozen language model. The external module (e.g., Toto/FNF) is already designed to aggregate predictions from time-series foundation models. It is neither clearly explained nor empirically justified why these aggregated numerical outputs need to be routed through a massive language model for further processing, or how the LLM actually provides further gains over an already capable forecast ensemble module.
  4. Potentially Unfair Benchmark Comparisons and Lack of True OOD Evaluation: The evaluation setup raises fairness concerns. While OpenTSLM TeeMoE is explicitly trained on the corresponding task distributions, many compared baselines—such as proprietary general LLMs and other time-series language models—were never trained on datasets aligned with Context is Key or TimeSeriesExam. A truly fair comparison requires evaluating all models with frozen weights on strictly held-out, third-party out-of-distribution (OOD) benchmarks.
  5. Insufficient Exploration of Stronger Ensemble Baselines and Backbone Variations: In the forecast aggregation task, the comparison is limited to relatively simple ensemble baselines such as XGBoost. It remains unclear whether replacing XGBoost with stronger, state-of-the-art forecast ensembling methods would narrow or eliminate the performance gap. Furthermore, the paper lacks ablation studies across different LLM backbone sizes and families.
  6. Subpar Presentation and AI-Generated Illustrations: The overall formatting and writing structure are not easy to follow. In particular, Figure 1 appears to be generated by AI tools rather than drawn as a rigorous academic diagram, which obscures the technical pipeline and degrades the presentation quality of the paper.
Questions for authors

Please refer to the points raised in the Weaknesses above.

6 comments
AuthorsOpenReview ↗

Response to Review (Part 1 of 3)

W1. Motivation for unification

We thank the reviewer for raising the motivation for unification. Numerical forecasting, contextual forecasting, and temporal analysis answer complementary questions: what will happen, how external information changes that forecast, and what the observed series reveals, and supporting these needs in one model is the intrinsic motivation for TeeMoE. This precise objective of covering multiple time-series capabilities also motivates UniTS [1], TimeOmni-1 [2], TimeOmni-VL [7], and ChatTS [3].

Building on this line of work, our goal is to develop a generalist that performs strongly across its capabilities. TeeMoE ranks among the top three systems on all three evaluated benchmarks, while Tables 2–3 show that bringing these capabilities together preserves specialist-level performance. By comparison, the existing TSLMs evaluated in Table 1 fall behind the leading benchmark entries on CiK and TSE. This is what makes TeeMoE's unification useful: a user can turn to the same model for numerical forecasting, context-conditioned prediction, and temporal analysis without giving up the accuracy of the corresponding specialist-level experts.

MoE is an important part of how we achieve this result: it allows capabilities with different training requirements to be developed separately and then combined within a shared backbone. However, our contribution is the resulting generalist model and training recipe, rather than a novel MoE framework whose principal claim is additional gains from expert interaction.

We agree with the reviewer that cross-task improvements would strengthen our results further and are worth studying. We focused on established capability benchmarks because we did not identify a comparably established protocol for measuring mutual gains among these three capabilities with strong specialist baselines. We therefore leave that evaluation to future work as suitable benchmarks develop. In the meantime, we believe that achieving broad capability coverage while retaining specialist-level performance itself contributes a useful result for the community.

We have clarified this motivation and the scope of the contribution in the Introduction and Discussion and Conclusion.

W2. Backbone size and efficiency

We thank the reviewer for requesting a direct capacity comparison. Although 7B is a common choice, we believe that the literature does not establish it as a sufficient capacity for temporal tasks generally. Existing studies span substantially different scales: OpenTSLM evaluates 270M, 1B, and 3B backbones [4]; ChatTS uses a 14B backbone and studies text-based time-series fine-tuning up to 32B [3]; and TimeOmni-1 provides a 9B release and reports 14B experiments [2]. There is therefore no established 7B standard against which a larger backbone is inherently unjustified.

To address this question empirically, we trained all three experts and the composer on Qwen3.5-4B, Qwen3.5-9B, and Qwen3.5-27B using the final TeeMoE recipe, preserving the training data, objectives, adapter ranks, global batch sizes, and optimization schedules. These sizes let us test models below and above the requested 7B scale and measure scaling within the same model family. The submitted Qwen3.6-27B TeeMoE is shown separately below.

BackboneGIFT rank ↓CiK RCRPS ↓TSE (%) ↑Training GPU-hours
Qwen3.5-4B19.9690.2415964.7458.118
Qwen3.5-9B19.8250.1958366.75610.650
Qwen3.5-27B20.1550.1646275.73731.227
Qwen3.6-27B (TeeMoE)19.9900.1147978.55230.756

The smaller models retain strong numerical forecasting, with 9B obtaining the best GIFT rank. In contrast, contextual forecasting and temporal analysis improve consistently from 4B to 9B to 27B within Qwen3.5; compared with 9B, the 27B model reduces CiK RCRPS by 15.9% and improves TSE accuracy by 8.98 percentage points. The submitted Qwen3.6-27B model performs better still on both tasks. These results support the 27B choice for performance across the generalist's capabilities, while the smaller variants offer a lower-training-cost option when numerical forecasting is the priority.

We have added this comparison to Section 5.3.4, with the training settings and cost accounting in the appendix.

AuthorsOpenReview ↗

Response to Review (Part 2 of 3)

W3. Role of the language model in forecast aggregation

We thank the reviewer for asking why the ensemble forecast passes through the language backbone. The primary purpose is unification: numerical forecasting should use the same backbone and adapter composition as contextual forecasting and temporal analysis. The numerical encoder feeds the ensemble forecasts into the backbone, and the decoder produces refined quantiles. This lets us train the aggregation adapter as an expert and combine it with the other two adapters in the same model.

The numerical ablation also demonstrates a forecasting benefit. Holding the reference ensemble fixed, mean MASE rank improves from 20.660 to 20.000 with the aggregation expert, and to 19.990 with TeeMoE. The design therefore integrates numerical forecasting into the generalist while improving an already strong ensemble.

We have expanded Section 3.1 to explain how the numerical path integrates with the shared backbone and language-model head.

W4. Evaluation protocol and frozen weights

We thank the reviewer for raising the evaluation protocol. TeeMoE is evaluated with frozen weights: all parameters are fixed before benchmark evaluation, and no weights are updated on test requests. If the reviewer means that the language backbone should also remain unadapted before evaluation, that would test a different contribution, as the paper contributes a trained time-series model and its training recipe, rather than an inference-only harness for an unchanged general LLM. Evaluating the benefit of that training is a central purpose of the comparison.

Comparing trained time-series models with general LLMs is established practice in OpenTSLM [4], ChatTS [3], TimeOmni-1 [2], TimeOmni-VL [7], and TS-Reasoner [8].

Even so, we additionally isolate our own adaptation benefit using the same Qwen backbone under “Pre-adaptation baseline,” an unadapted Qwen3.6-27B. With training, CiK RCRPS improves from 0.151 to 0.115, and TSE accuracy from 74.933% to 78.552%, establishing the training benefit independently of external systems' training histories.

Our CiK and TSE evaluations are zero-shot: we do not train on either benchmark's evaluation examples or supply labelled demonstrations at inference [5,6]. For GIFT-Eval, we use its designated training split for its intended purpose—training forecasting models. Leading entries, including Toto 2.0-FT and Toto-FnF, also use this split [10].

Section 5.1 and the comparison-protocol appendix now distinguish prior training from frozen-weight evaluation.

AuthorsOpenReview ↗

Response to Review (Part 3 of 3)

W5. Ensemble baselines and backbone variations

We thank the reviewer for asking whether the ensemble is sufficiently strong. XGBoost-based weighting has powered leading forecasting ensembles: FFORMA learns forecast weights from time-series features and placed second in M4 [9]. More recently, Toto's Family-and-Friends uses an FFORMA-style XGBoost meta-learner over ten foundation models, achieving leading GIFT-Eval results across all four metrics [10]. XGBoost is therefore an established choice for high-performing forecast ensembling, including at the foundation-model scale relevant here.

To address the request directly, we compared our XGBoost weighting with two neural weighting models (64 and 128 hidden units), global mean-rank weighting, and uniform weighting. The neural models use the same 576,920 training windows and targets as XGBoost, and are fitted for 20 epochs. All five methods are evaluated on full GIFT-Eval against the paper's common leaderboard roster.

Ensemble / modelMean MASE rank ↓
Uniform weighting (new)31.351
Global mean-rank weighting (new)29.959
Neural weighting, 64 hidden units (new)25.887
Neural weighting, 128 hidden units (new)25.021
XGBoost weighting22.629
Reference blend20.660
Reference blend + aggregation expert20.000
TeeMoE19.990

XGBoost outperforms these alternative weighting methods on the same forecaster pool, supporting our choice, and combining it with Toto-FnF gives the reference blend used for refinement.

The new 4B and 9B experiments address backbone size (W2).

We agree that testing another language-model family would also be valuable, but such a comparison requires training all three experts and the composer anew and evaluating the resulting model across all three benchmarks. Given that computational cost, we have prioritized the requested size and ensemble comparisons for this revision and leave cross-family evaluation to future work.

We have added the ensemble results to Section 5.3.3 and the appendix subsection “Alternative forecast-weighting models,” and identified cross-family evaluation as future work in Discussion and Conclusion.

W6. Presentation and Figure 1

We thank the reviewer for the presentation feedback. We have ensured that the components, connections, and output paths in Figure 1 accurately represent the implemented model and are consistent with the method equations, and we are happy to improve the figure and surrounding explanation. Could the reviewer identify the particular labels, connections, or sections that make the pipeline difficult to follow? This would help us address the source of the confusion directly and ensure that the revised presentation communicates the method clearly.

We have already expanded Sections 3.1 and 3.5 to clarify the two output paths and the controller's role.

References

[1] Gao et al. UniTS: A Unified Multi-Task Time Series Model. NeurIPS, 2024.

[2] Guan et al. TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models. ICLR, 2026. See also the official 9B release.

[3] Xie et al. ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning. PVLDB, 18(8), 2025.

[4] Langer et al. OpenTSLM: Time-Series Language Models for Reasoning over Multivariate Medical Text- and Time-Series Data. arXiv:2510.02410, 2025.

[5] Williams et al. Context is Key: A Benchmark for Forecasting with Essential Textual Information. ICML, 2025.

[6] Auton Lab. TimeSeriesExam, v1.1 dataset release.

[7] Guan et al. TimeOmni-VL: Unified Models for Time Series Understanding and Generation. ICML, 2026.

[8] Yu, Zhao, and Zhou. TS-Reasoner: Aligning Time Series Foundation Models with LLM Reasoning. TMLR, 2026.

[9] Montero-Manso, Athanasopoulos, Hyndman, and Talagala. FFORMA: Feature-based forecast model averaging. International Journal of Forecasting, 36(1):86–92, 2020.

[10] Khwaja et al. Toto 2.0: Time Series Forecasting Enters the Scaling Era. arXiv:2605.20119, 2026. Section 5.3 and Figure 7.

[11] Datadog. Toto-2.0-Family-and-Friends. Official model card.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their time and constructive feedback.

We have uploaded a revised manuscript and provided point-by-point responses under each review.

The revised manuscript adds a parameter-matched rank-52 joint model comparison, backbone-scaling experiments, alternative ensemble baselines, routing-granularity experiments, a paired uncertainty analysis of the CiK improvement, and GIFT subset evaluations for previously OOT models. We have also clarified the architectural rationale and output paths.

We are grateful for the suggestions, which have made the paper stronger. We welcome further questions and would be glad to hear whether these additions address the main concerns or whether anything else can be improved.

Area_Chair_o8kVOpenReview ↗

Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.