PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Tony Chen, Timo Stoffregen, Maxwell Xu, Thomas Kaar, Martin Maritsch, Geremia Pompei, Nicolas Zumarraga, Robert Jakob, Paul Schmiedmayer, Patrick Langer, Juncheng Liu

arXiv:2609.40265v1cs.LG

TL;DR

Time-series models often separate numerical forecasting from context-conditioned prediction and temporal reasoning, leaving open how to combine these capabilities without diluting specialist performance. OpenTSLM TeeMoE trains distinct experts over a shared backbone and uses a learned controller to combine them per request. It ranks among the top three on GIFT-Eval, Context is Key, and TimeSeriesExam.

  • Problem

    Numerical forecasting and language-based contextual analysis are often handled by different model families, raising the question of whether a generalist can broaden capabilities without diluting specialist performance.

    Source: §1 · p. 1

  • Method

    OpenTSLM TeeMoE independently trains aggregation, native forecasting, and temporal analysis experts over a shared backbone, then learns a controller to weight frozen experts and select an output path per request.

    Source: §1 · p. 2

  • Results

    OpenTSLM TeeMoE ranks among the top three on GIFT-Eval, Context is Key, and TimeSeriesExam.

    Source: §1 · p. 3, §6 · p. 9

  • Takeaways & Limitations

    The results support training each capability with its needed objectives and resources, then composing the experts while retaining their strengths and independent development flexibility.

    Source: §6 · p. 10

    The benchmarks assess capabilities separately, so workflows alternating forecasting and reasoning remain untested; evaluation beyond one backbone family is also a next step.

    Source: §6 · p. 10

Abstract

from arXiv · show

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.

1 INTRODUCTION

Time-series requests span numerical forecasting, context-conditioned prediction, and temporal reasoning, but existing model families address these capabilities separately. OpenTSLM TeeMoE combines independently trained capability experts with a controller, aiming for broad coverage without sacrificing specialist performance.

1 INTRODUCTION

  • Time-series applications need numerical forecasts, context-conditioned predictions, and temporal reasoning, motivating a generalist model that retains specialist accuracy.A demand-planning system may need to forecast demand, assess an intervention, or interpret a historical pattern.Source: §1 · p. 1
  • OpenTSLM TeeMoE trains three low-rank experts over a shared Qwen3.6-27B backbone for forecast aggregation, native forecasting, and temporal analysis.The aggregation expert refines external forecasts; the native expert predicts from history and text; the analysis expert answers questions about signals.Source: §1 · p. 2
  • A controller trained over frozen experts assigns their weights and selects an output path for each request.The proposed approach builds on modular expert composition, which lets specialists use distinct objectives and be revised independently.Source: §1 · p. 2
  • The paper presents TeeMoE as a generalist model with competitive performance across benchmarks for distinct capabilities, where specialized models usually dominate.The authors describe this as the first use of MoE to construct a generalist time-series language model.Source: §1 · p. 2
  • OpenTSLM TeeMoE ranks among the top three on GIFT-Eval, Context is Key, and TimeSeriesExam.The paper also studies capability preservation, expert interaction, and training cost under alternative compositions.Source: §1 · p. 3

2 RELATED WORK

Related work spans pretrained numerical forecasters, language-based time-series prediction and analysis, and general-purpose temporal models. TeeMoE draws on modular expert composition while extending forecast ensembles with language-backbone refinement.

Time-series foundation models

  • Numerical foundation models reuse patterns learned across many series to forecast new datasets with little or no task-specific training (Das et al., 2024; Ansari et al., 2024).Forecast ensembles such as FFORMA, Chroma, Synapse, and CastStar address variation in model performance; TeeMoE exposes ensemble predictions and constituent distributions to a language backbone for refinement.Source: §2 · p. 3

Language models for prediction and understanding

  • Time-series language models adapt pretrained language representations or connect observations to text for forecasting and temporal understanding (Zhou et al., 2023; Jin et al., 2024; Xie et al., 2025).Prior approaches include sentence-to-sentence forecasting, numerical-text continuation, and instruction tuning aligned with a frozen time-series model.Source: §2 · p. 3

Contextual forecasting

General-purpose temporal models

Expert specialization and composition

  • Modular adaptation composes task-specific LoRAs over a shared backbone; TeeMoE applies request-level weights shared across adapted layers (Ostapenko et al., 2024; Wu et al., 2024).The paper studies capability retention alongside independent expert training and revision.Source: §2 · pp. 3–4

3 OPENTSLM TEEMOE

OpenTSLM TeeMoE connects independently trained aggregation, native forecasting, and analysis adapters through one backbone and a learned controller. Its request interface routes each input to numerical forecast refinement or language generation using a mixture selected for that request.

3.1 DESIGN AND REQUEST INTERFACE

  • A request includes an observed series and an instruction, with optional context or a question; an adapter-disabled pass forms the controller representation before a second pass produces the response.The controller selects adapter weights and either the numerical decoder or language-model head.Source: §3.1 · p. 4

3.2 THE FOUNDATION MODEL AGGREGATION EXPERT

  • The aggregation expert builds a reference by combining pretrained forecasters, then uses candidate agreement and disagreement to guide refinement through the shared backbone.The XGBoost ranker scores eight forecasters; blending with Toto-FnF yields a union of thirteen candidates.Source: §3.2 · p. 4
  • The numerical encoder represents forecast levels, uncertainty, and candidate disagreement, while the decoder adjusts forecast location and preserves quantile order and interval widths.A disagreement gate reduces the edit when candidate disagreement is large relative to reference uncertainty.Source: §3.2 · p. 4

3.3 THE NATIVE FORECASTING EXPERT

  • The native forecasting expert generates timestamp–value pairs from historical observations and textual context without using external forecasters.Multiple sampled responses form an empirical predictive distribution.Source: §3.3 · p. 5

3.4 THE TEMPORAL ANALYSIS EXPERT

  • The temporal analysis expert answers questions by identifying patterns and anomalies, comparing series, or interpreting statistical relationships.It produces a selected label and text for multiple-choice questions, or a free-form answer.Source: §3.4 · p. 5

3.5 LEARNED EXPERT COMPOSITION

  • A learned controller combines frozen expert parameter updates using request-level softmax weights shared across adapted layers and fixed for the response.The numerical decoder is selected when the aggregation weight exceeds 0.5; otherwise, the model uses token generation.Source: §3.5 · p. 5

4 TRAINING

Training fits an external-forecast ensemble, then trains separate aggregation, native forecasting, and analysis experts with task-specific data and objectives. A final controller stage learns to combine the frozen experts and select an output format.

4 TRAINING

  • The shared pretrained backbone stays frozen while the numerical ensemble, capability experts, and controller are trained in sequence.Appendix A is cited for adapter configurations and optimization settings.Source: §4 · p. 5

4.1 AGGREGATION TRAINING

  • The numerical ensemble is fitted on 576,920 forecasting windows spanning seven application domains, with XGBoost predicting candidate ranks under empirical CRPS.The training also fits a scalar blend of XGBoost and Toto-FnF forecasts.Source: §4.1 · p. 5
  • The aggregation adapter, numerical encoder, and forecast decoder train jointly on 4,096 GIFT training examples using a weighted SmoothL1 correction loss and squared-edit penalty.The disagreement gate is applied at inference rather than supervised by this loss.Source: §4.1 · p. 5

4.2 NATIVE FORECASTING TRAINING

  • The native expert trains for one epoch on 20,000 contextual forecasting examples using teacher-forced response-token cross-entropy and forward KL regularization to the frozen base model.The examples combine real and synthetic time series.Source: §4.2 · p. 5

4.3 TEMPORAL ANALYSIS TRAINING

  • The analysis expert trains for one epoch on 12,000 question-answering examples with response-token cross-entropy and forward and reverse KL regularization to the frozen base model.Targets include categorical and free-form answers.Source: §4.3 · p. 6

4.4 CONTROLLER FITTING

  • The controller trains for one epoch on 1,000 sampled examples, combining capability prediction losses with supervision for numerical-versus-textual output selection.Text capabilities use response-token cross-entropy; aggregation uses normalized-correction loss without the squared-edit penalty.Source: §4.4 · p. 6

5 EXPERIMENTS

Experiments compare OpenTSLM TeeMoE across forecasting, context-conditioned prediction, and temporal reasoning, then examine benchmark standing, specialist retention, expert composition, training costs, and forecast refinement.

5.1 BENCHMARKS AND COMPARISONS

  • Evaluation uses 97 GIFT-Eval cells (371,330 forecast windows), 355 CiK instances across 71 task types, and 746 TSE questions.Reports with identified train–test leakage, incompatible evaluation populations, or unavailable ranking metrics are excluded.Source: §5.1 · p. 6

5.2 PERFORMANCE ACROSS CAPABILITIES

  • TeeMoE ranks among the top three across benchmarks: GIFT-Eval mean MASE rank 19.990, CiK RCRPS 0.115, and TSE accuracy 78.552%.It outperforms selected numerical foundation models on GIFT, locally evaluated time-series language models on CiK, and leads the TSE comparison. At approximately 40B total parameters, it also outperforms 405B Llama-3.1-405B-Instruct on CiK and the approximately 117B GPT-oss-120B hybrid agent on TSE.Source: §5.2 · pp. 6–7, Table 1 · p. 7

5.3 UNDERSTANDING TEEMOE’S PERFORMANCE

  • Comparisons distinguish specialist retention from expert-interaction gains and assess training costs for separate specialization versus joint optimization.The analyses compare experts, adapter combinations, ensembling, and refinement.Source: §5 · p. 6, §5.3 · p. 7

5.3.1 INDIVIDUAL EXPERTS

  • Native forecasting leads specialist CiK at 0.123 and analysis leads TSE at 78.418%; TeeMoE composition further improves both.Composition reduces RCRPS by 6.6% relative to the native expert and improves TSE accuracy by 0.134 percentage points over the analysis expert; independently trained experts retain their strengths in the shared backbone.Source: §5.3.1 · pp. 7–8, Table 2 · p. 8

5.3.2 EXPERT COMBINATION

  • Learned weighting balances tasks better than tested fixed mixtures; top-1 routing remains comparable, and composition outperforms the joint baseline on all three tasks.Equal weights slightly improve GIFT rank but worsen CiK and reduce TSE accuracy by 0.402 percentage points; full-strength adapters improve GIFT rank but worsen CiK and reduce TSE accuracy by 2.949 percentage points.Source: Table 3 · p. 8, §5.3.2 · p. 8
  • Separate specialization costs 30.756 H100 GPU-hours versus 58.657 for the joint recipe and allows one expert and the controller to be revised while others stay fixed.The reported costs exclude shared numerical preparation.Source: §5.3.2 · pp. 8–9, Table 4 · p. 9

5.3.3 NUMERICAL ENSEMBLING AND FORECAST REFINEMENT

  • On GIFT, learned weighting improves over uniform pooling and refinement further improves the fitted blend; on CiK, history-only ensembles score 0.283–0.292 RCRPS versus TeeMoE’s 0.115.Because GIFT rank measures relative ordering rather than error magnitude, the rank gains do not directly establish which component contributes more to forecast quality.Source: Table 5 · p. 9, §5.3.3 · p. 9

6 DISCUSSION AND CONCLUSION

OpenTSLM TeeMoE ranks among the top three systems on all three benchmarks, and its comparisons show that modular composition preserves specialist strengths while offering training and revision benefits. The evaluation covers established separate-capability benchmarks; workflows alternating forecasting and reasoning, and other backbone families, remain next steps.

6 DISCUSSION AND CONCLUSION

  • OpenTSLM TeeMoE ranks among the top three systems on all three benchmarks.The comparisons examine breadth across numerical forecasting, context-conditioned prediction, and temporal reasoning.Source: §6 · p. 9
  • Independently trained experts retain their strengths when composed; learned composition improves contextual forecasting, with smaller gains on GIFT and TSE.The discussion identifies preservation as the clearest benefit and describes routing’s predictive gains as limited.Source: §6 · p. 9
  • Separate specialization lowers training cost and allows an expert and controller to be retrained without retraining the other experts.Each capability can retain its own objective and optimization settings, unlike joint training in one shared adapter.Source: §6 · p. 9
  • The evaluation uses established benchmarks that assess capabilities separately; alternating forecast-and-reason workflows and other backbone families remain future directions.The authors say broader backbone coverage would clarify transfer to other models and training recipes.Source: §6 · p. 10
  • The authors present modular training as a practical route to combining capabilities while retaining specialist strengths and independent development.Each capability can use the objectives and resources it needs.Source: §6 · p. 10

REPRODUCIBILITY STATEMENT

The code and model checkpoints are publicly available, alongside workflows and appendices covering training and evaluation details.

REPRODUCIBILITY STATEMENT

ETHICS STATEMENT

The ethics statement cautions against treating benchmark accuracy as evidence of fitness for consequential autonomous decisions. It calls for domain validation and human oversight.

ETHICS STATEMENT

  • Forecasts and temporal interpretations may inform decisions in healthcare, finance, and public infrastructure.Source: ETHICS STATEMENT · p. 10
  • Benchmark accuracy does not establish that the model is fit for autonomous decisions in those settings.Source: ETHICS STATEMENT · p. 10
  • Users should validate the model on the intended domain and retain appropriate human oversight.Source: ETHICS STATEMENT · p. 10

AI USE STATEMENT

The appendices document AI use and the model’s request, controller, numerical aggregation, and text-training interfaces. Numerical forecasting refines external forecasts, while native forecasting and analysis use language-model generation.

AI USE STATEMENT

  • Generative AI assisted ideation, methods, implementation, dataset cleanup, literature review, drafting, figures, and experiment monitoring; authors directed decisions and retain responsibility for verification and final content.Source: AI USE STATEMENT · p. 10

A ARCHITECTURE AND TRAINING DETAILS

  • The appendices describe the request interface and numerical computation before training recipes, with separate sections for ablations and evaluation protocols, prompts, and comparison models.Source: A ARCHITECTURE AND TRAINING DETAILS · p. 17

A.1 SHARED LANGUAGE INTERFACE

  • A shared system instruction serves four user-message formats: forecasting with or without context, multiple-choice analysis, and free-form analysis.Formats include the example’s history and, where applicable, timestamps, context, question, and options.Source: §A.1 · p. 17
  • Native forecasting and analysis use the language-model output head, while aggregation uses a numerical encoder and forecast decoder.Source: §A.1 · p. 17

A.2 CONTROLLER AND OUTPUT SELECTION

  • A request-level controller assigns a three-way softmax mixture from text-only and full-request views; aggregation weight above 0.5 selects the forecast decoder.Otherwise, the response is generated as language-model tokens.Source: §A.2 · p. 17

A.3 NUMERICAL AGGREGATION INTERFACE

  • The aggregation expert blends a reference forecast from eight XGBoost-ranked forecasters with Toto-FnF; the combined pool supplies 13 candidates.Source: §A.3 · p. 17
  • The numerical encoder turns forecast levels, widths, asymmetries, positions, and candidate disagreement into embeddings; a 24-token ordering presents the horizon in both directions before output.The 53-value features and grouped agreement summaries are projected into embeddings for the backbone.Source: §A.3 · pp. 17–18
  • A 14-channel head interpolates corrections across the horizon; inference reduces correction when candidate disagreement is large relative to reference uncertainty, while shifts preserve quantile order and width.The head is initialized so the initial forecast equals the reference.Source: §A.3 · p. 18

A.5 AGGREGATION DATA AND HYPERPARAMETERS

The appendix specifies how TeeMoE’s aggregation, native-forecasting, and analysis experts are trained, then describes composition and control protocols. Its aggregation pipeline cross-fits reference forecasts before an expert refines them.

A.5 AGGREGATION DATA AND HYPERPARAMETERS

  • Cross-fitting supplies held-out reference forecasts for the aggregation expert, which refines a reference built from an XGBoost ensemble and a learned Toto-FnF blend.The scalar blend assigns Toto-FnF weight α ≈0.386; the pooled candidate forecast uses rank-derived CDF weights, and Toto-FnF is omitted when unavailable.Source: §A.5 · pp. 19–20

A.6 NATIVE FORECASTING DATA AND HYPERPARAMETERS

  • The native expert generates future values from observed history and supplied context, using synthetic contextual pairs, forecast discussions, contextual series, and augmented LOTSA windows.The LOTSA augmentation clips future values to stated bounds while leaving history unchanged; the CAF pool excludes 55 reviewed examples with contradictory bounds and targets.Source: §A.6 · pp. 21–22

A.7 ANALYSIS DATA AND HYPERPARAMETERS

  • The analysis expert maps observed signals and questions to categorical or free-form answers across activity, anomaly, comparison, readout, and temporal-description tasks.Its mixture includes Time-MQA, TSQA, HiTSR, and synthetic examples from the ChatTS authors’ published generator.Source: §A.7 · p. 22

A.8 COMPOSITION DATA AND HYPERPARAMETERS

  • Composition trains a request-conditioned controller over the three completed experts while keeping the backbone, adapters, encoder, and numerical output head frozen.The controller uses a 5,120-dimensional request representation to produce three logits; its example-level mixture remains fixed through each request’s editing or generation.Source: §A.8 · pp. 23–24
  • The composition objective combines each capability’s prediction loss with an output-format loss that supervises numerical-versus-text selection, not native-versus-analysis weights.The format loss is −log πagg for numerical examples and −log(1 − πagg) for text examples.Source: §A.8 · p. 23

B ABLATION PROTOCOLS

  • The ablation appendix defines expert, composition, and numerical-ensemble controls for comparison with the composed model.Its stated control scope covers individual experts, composition variants, and numerical-ensemble variants.Source: B ABLATION PROTOCOLS · p. 24

B.2 INDIVIDUAL EXPERTS AND COMPOSITION CONTROLS

  • Individual-expert controls activate a named frozen adapter, while composition controls compare learned, top-1, equal-weight, and full-strength adapter settings.For CiK and TimeSeriesExam, individual experts use the full textual request and language-model head; the aggregation expert uses its complete numerical interface on GIFT.Source: §B.2 · p. 24

B.3 NUMERICAL ENSEMBLE CONTROLS

  • Numerical-ensemble controls compare uniform candidate pools, fitted XGBoost weights, Toto-FnF, and their fitted blend, with the aggregation expert’s refinement omitted from the blend condition.CiK ensemble comparisons use history without textual context and no fitting on CiK; Toto-FnF alone cannot cover the full benchmark.Source: §B.3 · pp. 24–25

B.4 JOINT-TRAINING BASELINE

The joint-training baseline shares an adapter under matched training exposure and prediction interfaces; the accompanying sections specify benchmark comparisons, model inputs, and evaluation settings. The reported results place TeeMoE third on GIFT-Eval and above a GPT-oss-120B hybrid configuration on TimeSeriesExam.

B.4 JOINT-TRAINING BASELINE

  • The joint baseline matches specialist data exposure and prediction interfaces while using one shared rank-32 adapter; broader joint-training configurations are left to future work.Rank 32 matches the largest specialist’s rank without expanding capacity beyond every specialist’s training-time adapter width.Source: §B.4 · pp. 25–26

C EVALUATION PROTOCOLS AND COMPARISON SOURCES

  • Comparison protocols exclude reports with evaluation-data fitting, different test populations or supervision, or missing benchmark ranking metrics.For GIFT-Eval, the paper recomputes ranks across 134 entries; selected public-model comparisons preserve unsupported interfaces and report OOT when projected compute exceeds budget.Source: C EVALUATION PROTOCOLS AND COMPARISON… · p. 26, §C.1 · p. 26

C.1 GIFT-EVAL

  • 19.990 mean rank places TeeMoE third among 134 GIFT-Eval entries after adding its submission and recomputing per-cell midranks.The roster covers 97 evaluation cells and 371,330 forecast rows, including individual forecasting models, ensembles, and agent-based systems.Source: §C.1 · p. 26

C.2 CONTEXT IS KEY

  • Context Is Key evaluation uses weighted RCRPS on 71 tasks, five deterministic instances per task, and 25 forecast trajectories, with the trained checkpoint applied without updates or labelled demonstrations.The authors report no overlap between native-expert training sources and CiK evaluation instances and list IC-DP separately because it uses a ground-truth worked example.Source: §C.2 · pp. 28–29

C.3 TIMESERIESEXAM

  • TimeSeriesExam evaluation uses the 746-question v1.1 release, greedy 64-token decoding, and scoring of returned responses without constrained decoding.Empty or unmatched answers count as incorrect, and length-capped responses remain in the denominator.Source: §C.3 · p. 30
  • 78.552% accuracy exceeds the GPT-oss-120B hybrid agent’s 78.000% on the version-matched TimeSeriesExam comparison.The comparison preserves each published configuration’s model size, input representation, and tool access.Source: §C.3 · p. 30

D.2 ANALYSIS INPUTS

  • Analysis inputs pair raw observations with fixed numerical summaries of level, variability, trends, dependence, periodicity, and inter-series relationships.Summaries are computed solely from observed values, independently of the question, answer options, and reference answer.Source: §D.2 · p. 32

D.1 SHARED SYSTEM INSTRUCTION; D.4 EXAMPLE OF ROUTING AND EXPERT INPUTS

  • For contextual forecasting, the controller combines representations of task-text and the full request, while the expert receives the full request under the shared system instruction.The system instruction directs the model to choose structured aggregation or native contextual forecasting from supplied evidence, not task identity or candidate presence.Source: §D.1 · p. 32, §D.4 · p. 34, Figure 2 · p. 34

E.2 CONTEXT IS KEY

This block reports local Context is Key and TimeSeriesExam evaluations, release assets, and computational costs. The evaluation passages also specify scoring protocols and model-specific inputs.

E.2 CONTEXT IS KEY

  • CiK local evaluations used 355 instances and 25 trajectories per instance with an official per-instance RCRPS cap of 5; numerical foundation predictors received history and horizon, not text context.Table 11 lists local capped and uncapped results separately from published scores, for which paired values were unavailable.Source: §E.2 · p. 35, Table 11 · p. 36
  • Among local CiK transfers, ChatTS-14B scored 0.322 capped RCRPS, TS-Reasoner 0.374, TimeOmni-VL 0.386, and OpenTSLM 0.779.TeeMoE’s main score was identical with and without the per-instance cap.Source: §E.2 · pp. 35, 37

E.3 TIMESERIESEXAM

  • 57.507% strict accuracy was ChatTS-14B’s reported TimeSeriesExam score; TS-Reasoner scored 52.681% strict, OpenTSLM 26.810% strict, and TimeOmni-VL 14.745% flexible.Competitor results report the higher of the official flexible and strict scores, while TeeMoE and its ablations retain the flexible scorer.Source: §E.3 · p. 37

F.1 RELEASE AND SOURCE ASSETS

  • The release includes implementation, selected configurations, and guides for source preparation, expert training, composition, ablations, and comparison-model evaluation.Fixed source coordinates and membership metadata specify training selections; datasets, caches, and learned weights are handled separately from the source checkout.Source: §F.1 · p. 38

F.2 COMPUTATIONAL RESOURCES

  • 130 H100 GPU-hours were estimated for composed-model cache preparation, training, and evaluation; full GIFT-Eval, CiK, and TimeSeriesExam evaluations used 8.000, 5.780, and 0.550 GPU-hours.The GIFT-Eval estimate assumes foundation-model forecasts and request-routing states are already cached.Source: §F.2 · p. 38

F.3 MODULAR AND JOINT TRAINING COSTS

Joint training had a higher measured loop ratio despite equal example presentations, while the paper reports modular training costs separately from shared preparation. The discussion also identifies deployment safeguards.

F.3 MODULAR AND JOINT TRAINING COSTS

  • The output selector fits on cached request representations on CPU in about 29 seconds; loop timings exclude model loading and request-feature extraction.The selector adds no GPU training time.Source: §F.3 · p. 38
  • Shared numerical preparation and evaluation caches are excluded from Table 4 and reported separately in Table 12; both recipes use the same training forecast caches and fitted ensemble.The accounting separates shared preparation, model training, and capability-specific revision.Source: §F.3 · pp. 38–39
  • 1.907 was the measured joint-training loop ratio; equal example presentations still cost differently because synchronous updates wait for the longest rank-local workload and combine distinct losses.The joint run processed the same 4,096 aggregation, 20,000 native, and 12,000 analysis example presentations as the modular recipe.Source: §F.3 · p. 39

F.4 POTENTIAL IMPACTS

  • High-stakes deployment requires validation on the intended population, data protections, and human oversight; benchmark performance alone does not establish fitness for autonomous use.Incorrect forecasts or interpretations may lead to harmful decisions, and sensitive time-series inputs raise privacy concerns.Source: §F.4 · p. 39

Related papers

Loading 2609.40265v1…