acceptodds
Under review as a conference paper at ICLR 2027

LOTUS: Learning to Understand Time-series by Learning to Forecast

10 views

Abstract

Time-series language models typically align numerical signals with language using labels, captions, or question–answer pairs. Although such supervision encourages linguistic representations, it does not explicitly require them to preserve information about how the underlying time-series signal evolves. We introduce forecast grounding, a pretraining principle that requires language-aligned time-series representations to remain predictive of the future evolution before they are adapted to downstream tasks. We instantiate this principle in LOTUS, which divides each time-series channel into non-overlapping patches and maps them into the LLM embedding space. It then predicts the next patch from the LLM hidden states of the time-series tokens, encouraging the learned representation to retain temporal information. To capture both temporal changes and periodic patterns, we also propose a temporal–spectral patch representation that combines each raw time-series patch with its spectral representation before mapping it into the LLM embedding space. Across three language model backbones and six downstream evaluation datasets, LOTUS achieves the best average performance with minimal trainable parameters, reaching accuracy and outperforming the strongest matched baseline by . Our representation analyses further show that LOTUS uses the numerical input and preserves temporal structure in its learned patch representations. These results establish forecast grounding as a promising pretraining principle, enabling compact time-series language models to outperform models 25–50× their size on TSRBench.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series application, Time series understanding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_H4WpOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

LOTUS proposes a time-series language model based on “forecast grounding.” During Phase 1, the model maps time-series patches into LLM soft tokens and uses next-patch prediction from LLM hidden states to encourage the representations to preserve temporal information. It also combines raw temporal patches with FFT-based spectral features. The pretrained model is then adapted to downstream time-series classification and QA tasks using language supervision, and is evaluated across several backbones and benchmarks.

Critical strengths
  1. The paper is clearly written and easy to follow. The motivation, architecture, and two-stage training pipeline are presented clearly.
  2. The authors provide several diagnostic analyses and ablations, although some of them do not fully isolate the claimed contribution.
Critical weaknesses
  1. Limited novelty.
    The main idea is close to aLLM4TS, which already uses next-patch prediction for LLM-based time-series representation learning. LOTUS mainly adds time-series–language alignment and a standard FFT-based temporal–spectral fusion, so the method feels largely compositional.

  2. The forecast-grounding ablation is confounded.
    The full model scores 55.19, while the language-only Phase 1 variant scores 51.79. However, this comparison changes both the forecasting objective and the use of Monash/synthetic forecasting data, so it does not clearly show whether the gain comes from next-patch prediction itself or simply from additional numerical pretraining data.

  3. Potential future-information leakage.
    For multivariate inputs, channels are provided as separate blocks. If complete channel sequences are serialized sequentially, causal attention over token order may still expose later timestamps from earlier channels. The forecasting prompt also includes mean/std computed from the full series, which may contain future information.

  4. Possible answer leakage in ECG-QA.
    In one appendix example, the clinical context already says “burst noise present,” while the question asks “Does this ECG show burst noise?”. This suggests that some samples may be answerable from textual metadata without reading the ECG signal.

  5. Baseline coverage is relatively weak.
    The main matched comparisons are mostly against text-tokenized and OpenTSLM variants. Several more recent time-series language or multimodal methods are discussed but not directly compared, making it difficult to assess competitiveness against stronger recent baselines.

  6. Some claims are overstated and robustness is limited.
    On TSRBench, LOTUS-270M scores 38.5 versus 38.2 for TS-Reasoner-7B, so the margin is very small despite the “25–50× larger” claim. Key ablations are also mainly done on Gemma3-270M, and ECG-QA uses only a fixed 40k subset out of about 159k training samples.

Questions for authors
  1. Beyond combining aLLM4TS-style next-patch prediction with FFT features, what is the main methodological novelty of LOTUS?

  2. Why are stronger and more recent time-series language or multimodal baselines not included in the main comparison?

  3. How do the authors avoid future-information leakage in multivariate next-patch prediction, especially with channel-wise serialization and full-window statistics?

  4. Some ECG-QA prompts appear to contain the answer in the textual metadata. How well would a text-only LLM perform without using the ECG signal?

10 comments
AuthorsOpenReview ↗

Response to Reviewer H4Wp (1/4)

We thank the reviewer for the critical comments. We address each of the concerns below

W1 / Q1: "Limited novelty"

We acknowledge causal next-patch prediction through an LLM is shared with aLLM4TS, as discussed in Section 2.2, and we do not claim that objective as new.** We kindly note that LOTUS differs from existing time-series understanding models in two key aspects: (1) forecast grounding, which explicitly encourages language-aligned representations to preserve future-relevant temporal information, and (2) a temporal–spectral patch representation that jointly incorporates raw time-domain values and rFFT features before entering the LLM. Specifically, next-patch prediction supervises the post-LLM hidden states of soft time-series tokens, alongside language supervision. The forecasting decoder is subsequently removed, allowing us to evaluate whether the learned temporal representations transfer to downstream language tasks. For the reviewer’s clarity, we list the differences below.**

  • Language interface. aLLM4TS takes numerical patches as input and decodes numerical patches. It is evaluated on numerical tasks such as forecasting, anomaly detection and classification, and it produces no language output. LOTUS interleaves time-series tokens with text in one LLM and is evaluated on question answering with written rationales.

  • New finding. aLLM4TS shows that next-patch pretraining yields transferable numerical representations. LOTUS shows that the same supervision transfers to language understanding. In our same-budget study, forecasting-only pretraining with no language supervision raises the three-task mean by 4.9 points over no pretraining and by 2.8 points over language-only pretraining (tables in our response to Reviewer 9Lhr).

  • Joint context. LOTUS contextualizes channels, temporal patches and text together within one LLM, whereas aLLM4TS processes channels independently.

  • Evidence beyond FFT. Every arm in our new controls keeps the temporal-spectral encoder. With data, batches and decoder fixed, the next-patch target is ahead of a current-patch target in all three seeds, by 2.1 points on average, so the gain is not explained by the FFT branch or by the additional corpora.

W2: "The forecast-grounding ablation is confounded."

We acknowledge that the original comparison changes the objective and the data together, because the Monash and synthetic corpora contain no text and enter training only through the forecasting loss.

Data-matched control. We ran a three-arm study to separate the forecasting objective from the forecasting data. The first two arms share the same Phase-1 corpora, batches and decoder, and differ only in the decoder target: the decoder either predicts the next patch (LOTUS) or reconstructs the current patch (control). The third arm uses language-only pretraining. Each arm is pretrained with three seeds, and every pretrained model is then fine-tuned on Sleep-EDF alone with a fixed budget and no model selection (Gemma3-270M, full test set).

Phase 1Seed 42Seed 43Seed 44Mean
Forecasting, next patch (LOTUS)65.2769.4666.6767.13
Same data, current patch62.6968.3963.9865.02
Language only66.7764.6266.6766.02
Comparative EvaluationPerformance Gain95% CI
Next patch vs. current patch, same data+2.1[0.5, 3.7]
Current patch, same data vs. language only−1.0[−2.5, 0.6]
With the same data, the forecasting target is better than the current-patch target in all three seeds, and the interval for the gain excludes zero. The same corpora without a forecasting target gives no gain over language-only pretraining. On this task, the improvement therefore follows the forecasting objective and is not explained by the additional numerical corpora. Against language-only pretraining, the forecasting arm is higher on average, although this single-dataset contrast varies across seeds; the three-task study below gives the broader estimate of that contrast.
AuthorsOpenReview ↗

Response to Reviewer H4Wp (2/4)

Additional same-budget study. We also compared four Phase-1 settings under an identical budget. Every pretrained arm receives 6,000 Phase-1 updates, and all arms are then fine-tuned on each task with the same recipe and evaluated on the Table 1 test sets (Gemma3-270M, univariate forecasting corpora, one seed). The three-task mean is distinct from the six-split average in Table 3.

Phase-1 objectiveHARSleep-EDFECG-QAMean
None67.8464.4113.5648.60
Language only69.7662.0420.3850.73
Forecasting only64.8767.7427.9453.52
Language + forecasting70.7066.5619.7852.35
Comparative EvaluationPerformance Gain (pp)95% Confidence Interval (over test items)
Language + forecasting vs. language only+1.62[0.48, 2.76]
Forecasting only vs. none+4.92[3.85, 5.99]
Forecasting only vs. language only+2.79[1.76, 3.82]

Adding forecasting to language pretraining improves the three-task mean at the same update count, so this gain cannot be attributed to additional optimization steps. Forecasting-only pretraining achieves the highest mean and improves over both no pretraining and language-only pretraining. Numerical predictive supervision therefore transfers to downstream language tasks even without Phase-1 language supervision. All three intervals exclude zero.

Taken together, the two controls show that the benefit of forecast grounding is explained neither by a larger optimization budget nor by the additional numerical corpora.

W3/ Q3: Potential future-information leakage

We would like to provide clarifications regarding the channel ordering and normalization below.

  • Channel ordering. Separate blocks per channel are used in downstream fine-tuning, where the forecasting objective is absent. During Phase-1 pretraining, a multivariate series occupies one block with patches interleaved by time: all channels of patch 1, followed by all channels of patch 2, and so on. With patch-wise encoding and causal attention in the LLM and next-patch decoder, the state predicting patch n+1 cannot attend directly to patch n+1 or later patch tokens from any channel. We verified this ordering in the dataset code and through perturbation checks across all three backbones. We will revise Figure 12 correspondingly to the above description.

  • Mean and standard deviation. Our normalization follows prior autoregressive time-series pretraining: LLM4TS [1] applies instance normalization to the whole input sample before its autoregressive next-patch alignment, so its statistics also cover the target patches. LOTUS likewise computes per-channel statistics over the observed input window, rather than the complete underlying series. We acknowledge that these statistics aggregate information from later target patches, although causal attention prevents direct access to their tokens. Our next-patch objective is auxiliary supervision for representation learning, not an evaluation of causal forecasting. The forecasting decoder is removed before adaptation, and all downstream classification and question-answering tasks operate on fully observed signals (Section 3.3). We will clarify this distinction in the revision.

AuthorsOpenReview ↗

Response to Reviewer H4Wp (3/4)

W4 / Q4: ECG-QA metadata and performance without the signal.

We note that the cited example does not demonstrate answer leakage, as the correct answer could contradict what is given in the context statement. For the record of the appendix example: its correct answer is “no,” although the context says “burst noise present.” Copying that statement would therefore give the wrong answer. The phrase is boilerplate from the public OpenTSLM prompt builder shared by the compared methods. Our audit found it to be identical in all 159,313 training prompts and all 5,000 evaluated test prompts. It therefore cannot distinguish positive from negative recordings. We double check the label counts, and the results we have confirm this:

Yes/no test questionsCountContext ruleAlways "no"
Burst noise85132.8%67.2%
All six context-related conditions3,04232.3%67.7%

Of the 851 burst-noise questions, only 279 have the answer “yes.” Across all 3,042 questions concerning the conditions mentioned in the context, copying its affirmative statements performs substantially worse than always answering “no.” The highlighted boilerplate therefore does not provide a recording-specific answer cue.

To further study the effect of the context, as well as the ECG itself, we additionally run a separate input-ablation and report the results below:

Test-time inputECGClinical contextQuestionAccuracy (%)Change from full input (points)
Full input✓\checkmark✓\checkmark✓\checkmark33.16
Without clinical context✓\checkmark✓\checkmark30.06−3.1
Without ECG✓\checkmark✓\checkmark19.56−13.6
Question only✓\checkmark19.76−13.4

Signal dependence. Removing the ECG reduces accuracy by 13.6%, while removing the clinical context reduces it by 3.%. Without the ECG, clinical context plus question achieves 19.56%, compared with 19.76% for the question alone. These controls show that the trained checkpoint uses the ECG signal and that the highlighted boilerplate does not explain its performance.

W5 & Q2: "Baseline coverage is weak"

We would like to defend our choice of baseline here. First, in terms of framework comparison, we compared our with the direct competitor OpenTSLM variants [2] and to our knowledge, one of the most recent baselines ([2] was accepted in ICML 2026). Second, in terms of backbones, we compared against multimodality models like Gemma-3-4B-pt (image) [3] and large time-series reasoning models like TS-Reasoner [4] (accepted in TMLR 2026), TimeOmni-1 [5] (accepted in ICLR 2026), Chat-TS [6] (accepted in VLDB 2025). We reckon that, the baselines included for comparison is comprehensive and up-to-date. However, we respect the reviewer's opinion, and we run additional experiments to include GEM [7] (accepted in Neurips 2025) for comparison. GEM is a multimodal ECG language model, and we compare our 0.6B model against 2 versions of it: zero-shot and fine-tuned. The results are shown below.

ModelHARSleep-EDFECG-QAPAMAP2MCQT/FAvg.
GEM-7B, zero-shot3.5317.5334.643.8534.3331.9120.96
GEM-7B, fine-tuned43.8661.9438.4049.8348.4270.8552.22
LOTUS, Qwen3-0.6B63.6677.2031.5853.4256.7276.9059.91

As the results show, LOTUS-0.6B is ahead on every split over GEM-7B, except ECG-QA. We found out that one of the main reason GEM-7B leads on ECG-QA because it is an ECG specialist: its instruction-tuning data (ECGInstruct) already contains about 139K ECG-QA samples, so it has been trained on this task specifically before our fine-tuning. Its zero-shot row shows this directly. Without any training on our data, GEM-7B reaches 34.64% on ECG-QA, while staying near or below chance on HAR, Sleep-EDF and PAMAP2. Again, we want to note that the size gap of our model (LOTUS-0.6B) and GEM-7B are nearly 12 times, which further show the effectiveness and efficiency of our method.

Reviewer_9LhrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes LOTUS, a time-series language model in which a small LLM with low-rank adapters reads 32-sample signal patches as continuous token embeddings and answers questions in text. Its first contribution is forecast grounding, a pretraining stage in which a temporary decoder predicts the next patch from the LLM's hidden states at signal positions. Its second contribution is a patch embedding that combines each patch's raw values with its Fourier coefficients. One model per backbone (Gemma3-270M, Gemma3-1B and Qwen3-0.6B) is fine-tuned jointly on five datasets for activity recognition, sleep staging and time-series question answering. Against baselines trained on the same data and backbone, LOTUS has the best average accuracy over six test splits on all three backbones (Table 1).

Critical strengths
  1. Forecast grounding is a simple, well-motivated addition to language-only supervision. Because labels and captions need not preserve temporal dynamics, a next-patch loss is applied to the LLM's hidden states at the signal positions. The decoder is discarded before downstream tuning, leaving the inference path unchanged.
  2. Accuracy gains are consistent across three backbones under a matched protocol. Baselines share the pretraining sources, downstream data mixture, greedy decoding and answer parser. On Gemma3-270M, Gemma3-1B and Qwen3-0.6B, LOTUS exceeds the strongest same-backbone baseline by 2.65 to 3.51 points in average accuracy over six test splits (Table 1).
Critical weaknesses

The evidence that forecasting causes the gains is limited. The only same-budget control, on Gemma3-270M, replaces forecasting with language-only pretraining and lowers average accuracy from 55.19 to 51.79, with no repeated runs reported (Table 3). That variant also drops the electricity and synthetic forecasting corpora, so objective and data change together, and a data-matched control would be needed to separate them.

Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9Lhr

We thank the reviewer for the positive and constructive assessment. We divide the reviewer’s concern into three parts: (1) same-budget control, (2) no repeated runs, and (3) objective and data changing together. We ran new experiments for all three.

"Data-matched control." We ran a three-arm study to separate the forecasting objective from the forecasting data. The first two arms share the same Phase-1 corpora, batches and decoder, and differ only in the decoder target: the decoder either predicts the next patch (LOTUS) or reconstructs the current patch (control). The third arm uses language-only pretraining. Each arm is pretrained with three seeds, and every pretrained model is then fine-tuned on Sleep-EDF alone with a fixed budget and no model selection (Gemma3-270M, full test set):

Phase 1Seed 42Seed 43Seed 44Mean
Same data, current patch62.6968.3963.9865.02
Language only66.7764.6266.6766.02
Forecasting, next patch (LOTUS)65.2769.4666.6767.13

With the same data, LOTUS’s next-patch target is ahead of the current-patch control in all three seeds, by 2.1% on average. The same corpora with the current-patch target gives no gain over language-only pretraining. We therefore reckon the gain comes from the forecasting objective, not from the additional corpora.

Additional same-budget study. We also compared four Phase-1 settings under an identical budget. Every pretrained arm receives 6,000 Phase-1 updates, and all arms are then fine-tuned on each task with the same recipe and evaluated on the Table 1 test sets (Gemma3-270M, univariate forecasting corpora, one seed):

Phase-1 objectiveHARSleep-EDFECG-QAMean
None67.8464.4113.5648.60
Language only69.7662.0420.3850.73
Forecasting only64.8767.7427.9453.52
Language + forecasting70.7066.5619.7852.35

Adding forecasting to language-only pretraining raises the three-task mean by 1.6%. Forecasting alone, with no language supervision, is 2.8% above language-only pretraining and 4.9% above no pretraining. Forecasting therefore improves the mean at a fixed update budget, with or without Phase-1 language supervision.

Why objective and data change together in Table 3. We kindly note that the electricity and synthetic corpora contain no text. A language-only objective gets no training signal from them, so they cannot be kept while the forecasting loss is removed. The data-matched control in the data-matched control above is the experiment that separates the two.

Based on these findings, we have updated the paper with a more detailed ablation analysis, including the data-matched control over three seeds and the same-budget comparison of four Phase-1 settings. We hope that the reviewer is satisfied with our response and will increase their score accordingly.

Thanks for the responses, which addressed all my concerns. I will maintain my positive score.

AuthorsOpenReview ↗

Thank you for the response

Dear Reviewer 9Lhr,

Thank you for the response. We really appreciate your time and consideration.

Best regards,

The authors

Reviewer_P8Q5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LOTUS, a framework for aligning time series with pretrained LMs for time-series Q&A and classification. The method introduces two main components: (1) a temporal–spectral patch representation that concatenates raw time-series patches with their rFFT frequency features before projecting them into LMs, and (2) "forecast grounding" in Phase-1 pretraining, which attaches a next-patch regression head to the LM's hidden states. In Phase 2, the forecasting head is discarded and the model is jointly fine-tuned via LoRA across five downstream datasets (HAR, PAMAP2, Sleep-EDF, ECG-QA, and TimeMQA). Experiments across Gemma3 and Qwen3 backbones show good accuracy and favorable memory scaling compared to baselines.

Critical strengths
  1. The method combining local time-domain patches with rFFT spectral features is well-suited for physiological and sensor signals (EEG, ECG, accelerometers), where frequency bands carry key diagnostic information (verified by the delta-band ablation in Figure 4b). Coupled with a patch size of 32, LOTUS achieves good accuracy on tiny backbones (270M–1B) with much lower memory and token overhead than baselines.

  2. The paper clearly exposes how sequential task fine-tuning (used in prior work like OpenTSLM) suffers from catastrophic forgetting, dropping early-task accuracy to 0%, and sets up a much fairer balanced joint-mixture baseline for all compared methods.

Critical weaknesses

(1) While adding rFFT features is helpful for periodic sensor data, the method itself does not fundamentally change existing paradigms—it mainly combines a standard two-stage pretrain-then-finetune pipeline with a conventional next-patch MSE loss and joint multi-task training.

(2) In the Phase-1 ablation (Appendix Table 3), the full model trains on 844K extra multivariate series that the language-only baseline never sees, making the comparison unfair. Even with this extra data, forecast grounding brings almost no gain on general MCQ reasoning (39.41% vs. 39.23%), as most improvements actually come from the rFFT features on sensor classification tasks.

(3) The overall average score on TSRBench (Figure 3) is noticeably skewed by a single high-scoring subset (EP at 83.5%), while LOTUS still trails the baselines on most of the other 14 subsets, including the time-series forecasting (TSF) task (17.6%).

Questions for authors

no more questions

3 comments
AuthorsOpenReview ↗

Response to Reviewer P8Q5 (1/2)

We thank the reviewer for the constructive comments. We address your concerns and questions below.

W1. "The method itself does not fundamentally change existing paradigms."

We kindly note that LOTUS differs from existing time-series understanding models in two key aspects: (1) forecast grounding, which explicitly encourages language-aligned representations to preserve future-relevant temporal information, and (2) a temporal–spectral patch representation that jointly incorporates raw time-domain values and rFFT features before entering the LLM. Specifically, next-patch prediction supervises the post-LLM hidden states of soft time-series tokens, alongside language supervision. The forecasting decoder is subsequently removed, allowing us to evaluate whether the learned temporal representations transfer to downstream language tasks. While prior work has explored predictive supervision for time-series representation learning, the key distinction is its intended purpose. We summarize the closest approaches below:

MethodPredictive supervisionDownstream objective
PatchTST [1]Masked-patch reconstructionNumerical time-series tasks
aLLM4TS [2]Next-/multi-patch prediction with an LLMNumerical time-series tasks
OpenTSLM [3], ChatTS [4]Primarily language-based alignmentLanguage understanding
TS-Reasoner [5]Pretrained numerical encoder followed by language alignmentLanguage reasoning
LOTUSNext-patch regression from post-LLM soft-token states, alongside language supervisionLanguage understanding

LOTUS connects predictive representation learning with time-series–language alignment. Unlike forecasting-oriented approaches, our goal is not to improve numerical prediction itself, but to use future prediction to strengthen representations for language-based understanding. Our results support this motivation: forecasting-based pretraining improves the six-task average from 51.79% to 55.19% compared with language-only pretraining (Appendix Table 3), and LOTUS achieves the highest average accuracy across all three matched backbones (Table 1). We kindly note that the surrounding pipeline (two-stage training, next-patch MSE, joint multi-task adaptation) is not claimed as a contribution, and we will state this positioning in Section 2.2.

rFFT beyond sensor data. In the ablation of Table 3, we retrain LOTUS on Gemma3-270M with the frequency branch removed, so that each patch is embedded from its time-domain values only, either in Phase 2 alone or in both phases. All other settings follow Table 1. We report the two Time-MQA test splits, which pose general questions over univariate series from several domains:

VariantMCQT/F
LOTUS39.4174.47
w/o frequency branch (Phase 2 only)37.1072.95
w/o frequency branch (both phases)37.3172.64

Both splits are lower in both variants, by 1.5 to 2.3 points, so the benefit of the frequency branch is not confined to the sensor classification datasets, although it is smaller here than on ECG-QA and PAMAP2. We attribute it to questions about properties such as volatility and periodicity (please refer to our FIgure 17), which are compact in the frequency domain and harder to read from the raw values of a single patch.

W2. “...the full model trains on 844K extra...”

First, both LOTUS and the language-only Phase-1 variant use the same temporal–spectral representation, including rFFT features. Thus, rFFT alone cannot explain their performance difference. Second, we reckon that while additional data can be beneficial, its effectiveness depends on the learning objective and downstream transfer. LOTUS uses next-patch prediction to encourage representations to preserve future-relevant temporal information, rather than merely increasing data exposure. Moreover, fixed mixture weights in Phase-1 sampling (Appendix E) prevent the large forecasting corpus from dominating training solely due to its size. Appendix Table 3 shows improvements across all six evaluation splits, with a +3.4% average gain over language-only pretraining.

We also run additional experiments to clarify this (revised Sec. 5.1.2). The first keeps the Phase-1 data, batches, and decoder fixed and changes only the decoder target (Sleep-EDF, 3 seeds):

Phase 1424344Mean
Language only66.7764.6266.6766.02
Current patch62.6968.3963.9865.02
Next patch (LOTUS)65.2769.4666.6767.13

The second gives every setting the same budget of 6,000 updates:

Phase 1HARSleepECGMean
None67.8464.4113.5648.60
Language only69.7662.0420.3850.73
Forecasting only64.8767.7427.9453.52
Language + forecasting70.7066.5619.7852.35

With the same data, the next-patch target wins in all three seeds, and forecasting still helps at an equal budget, so the gain follows the forecasting objective, not extra data or steps.

AuthorsOpenReview ↗

Response to Reviewer P8Q5 (2/2)

W3. "The overall average score on TSRBench (Figure 3) is noticeably skewed by a single high-scoring subset (EP at 83.5%)"

Our evaluation follows TSRBench's official protocol, which reports accuracy over all evaluation instances so that each question contributes equally, to ensure a fair and direct comparison with existing baselines. Under this protocol, LOTUS achieves competitive overall performance with a 270M backbone, approximately 26× smaller than TS-Reasoner-7B and 52× smaller than ChatTS-14B, and it outperforms the OpenTSLM variants with the same backbone size.

Regarding TSF, Phase 1 uses forecasting as a pretraining objective, but Phase 2 removes the forecasting decoder and optimizes only the language objective (Section 3.3), so some forecasting-specific capabilities may be weakened or forgotten during adaptation. Our contribution is forecast grounding as a pretraining principle for language-based time-series understanding, rather than a complete framework for downstream forecasting.

To investigate this further, we add the forecasting task to Phase 2. For this experiment, in Phase 2, LOTUS is trained jointly on the five downstream datasets (HAR, PAMAP2, Sleep-EDF, ECG-QA, and TimeMQA) together with the forecasting data from Phase 1 (Monash electricity demand and our synthetic multivariate corpus), so forecasting supervision is retained alongside the language objective. We note that the decoder is retained during training for the newly added forecasting task, while the rest of the architecture is kept the same. The below results for OpenTSLM-SP and OpenTSLM-Flamingo baselines follow their sequential recipe, with forecasting added to Phase 2 under a similar protocol. TS-Reasoner-7B is the reference result from the paper. Update: Our phase-2 training has now finished, and we also report the trained OpenTSLM-SP and OpenTSLM-Flamingo (sequential) baselines with a similar protocol. Avg. is the weighted mean over the subsets, strictly following the TSRBench protocol.

ModelPRNUADCAERCDARTRNRIRTSFEPQualDMQuantDMAvg.
OpenTSLM-SP (sequential) + forecasting in Phase 2 (fully trained)40.951.332.149.230.326.426.925.425.832.124.315.725.020.030.4
OpenTSLM-Flamingo (sequential) + forecasting in Phase 2 (fully trained)30.337.234.022.532.820.123.911.023.530.715.835.123.512.825.2
LOTUS26.250.029.060.927.625.725.019.234.229.917.683.525.823.738.5
LOTUS + forecasting in Phase 2 (epoch 2, preliminary)24.545.633.351.624.329.322.721.930.027.033.158.123.628.032.4
LOTUS + forecasting in Phase 2 (fully trained)49.157.540.153.227.935.133.425.341.934.634.982.432.125.640.9

With full training, TSF rises from 17.6 to 34.9 and 12 of the 14 subsets improve over the original checkpoint. The average rises from 38.5 to 40.9 while EP performance even marginally drops from 83.5 to 82.4, so we believe the gain does not depend on that single subset. As observed from our additional results, LOTUS is well ahead of both OpenTSLM baselines trained with a similar protocol, and it has the highest accuracy among these three baselines on 12 of the 14 subsets. We reckon the results confirm our contribution, which is forecast grounding as a pretraining principle for language-based time-series understanding.

Based on these findings, we have updated the paper with a more detailed positioning and TSRBench analysis, including the comparison with the closest prior work in Section 2.2 and the results with forecasting retained in Phase 2 that backed our framework design and supports our model's forecasting ability. We hope that the reviewer is satisfied with our response and will increase their score accordingly.

References

[1] A Time Series is Worth 64 Words: Long-term Forecasting with Transformers - ICLR

[2] Multi-Patch Prediction: Adapting Language Models for Time Series Representation Learning - ICML

[3] OpenTSLM: Time-Series Language Models for Reasoning over Multivariate Medical Text- and Time-Series Data - ICML

[4] ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning - VLDB

[5] TS-Reasoner: Aligning Time Series Foundation Models with LLM Reasoning - TMLR 2026

Dear Reviewer P8Q5,

With the discussion period ending in approximately 2 days, we would greatly appreciate your feedback on our rebuttal. We believe we have thoroughly addressed your concerns and would welcome your assessment. If you find that we have successfully resolved your major points, we respectfully request that you consider adjusting your score to reflect these improvements.

Thank you for your time and consideration.

Best regards,

The Authors

Comments on the paper

5 comments
AuthorsOpenReview ↗

General response from the authors

We thank the reviewers for their constructive feedback. In addition to the individual response to each reviewer, we have uploaded a revised paper based on the changes requested, with all changes marked in blue

We first thank the reviewers for the strengths listed:

  • A simple and well-motivated principle. Forecast grounding applies a next-patch loss to the LLM's hidden states, and the decoder is discarded before downstream tuning, leaving the inference path unchanged (Reviewer 9Lhr).
  • A representation suited to real signals at low cost. Combining time-domain patches with rFFT features suits physiological and sensor signals, with good accuracy on small backbones at much lower memory and token overhead (Reviewer P8Q5).
  • Consistent gains under a matched protocol. LOTUS has the best average accuracy on all three backbones against baselines that share the same data, decoding, and parser (Reviewer 9Lhr), and the paper sets up a fairer joint-mixture baseline (Reviewer P8Q5).
  • Clear presentation. The motivation, architecture, and training pipeline are clearly presented, with several diagnostic analyses (Reviewer H4Wp, Reviewer P8Q5).

For each concern raised, we ran additional experiments and added clarifications:

Reviewer P8Q5

  • Novelty of the method (W1): Section 2.2 now clearly states the contribution, which is using future prediction to strengthen language-based understanding. Appendix A.3 and Table 6 show that the frequency branch also helps on the two Time-MQA splits, beyond sensor data.
  • Extra forecasting data in the Phase-1 ablation (W2): new Section 5.1.2 and Table 3 separate the objective from the data. With the same corpora, batches, and decoder, the next-patch target is ahead of a current-patch target in all three seeds, and a same-budget comparison shows that the gain is not due to additional optimization steps.
  • TSRBench overall score and low TSF accuracy (W3): our evaluation follows the official TSRBench protocol. A new paragraph in Section 4.3 and Table 2 add the forecasting task to Phase 2: TSF accuracy rises from 17.6% to 34.9%, 12 of the 14 subsets improve, and LOTUS stays well ahead of the OpenTSLM baselines trained with a similar protocol.

Reviewer 9Lhr

  • Evidence that forecasting causes the gains (W1): Section 5.1.2 and Table 3 add a data-matched control repeated over three seeds, which isolates the objective from the data, and a same-budget comparison of four Phase-1 settings, where adding forecasting improves the mean. The electricity and synthetic corpora contain no text, which is why the objective and the data changed together in the original ablation.

Reviewer H4Wp

  • Novelty relative to aLLM4TS (W1, Q1): Section 2.2 acknowledges the shared causal next-patch objective and states the distinction: aLLM4TS decodes numerical patches with no language output, whereas LOTUS uses future prediction to strengthen language-based understanding.
  • Confounded forecast-grounding ablation (W2): addressed by the data-matched and same-budget controls in Section 5.1.2 and Table 3.
  • Future-information leakage (W3, Q3): a new paragraph in Section 3.2 describes the token order and normalization. In Phase 1, patches are interleaved by time, so the state predicting patch n+1n+1 cannot attend to that patch or any later one. Figure 12 is revised accordingly.
  • Answer leakage in ECG-QA (W4, Q4): Appendix A.1 and Table 4 show that the phrase is boilerplate identical in every prompt, and that removing the ECG lowers accuracy by 13.6% against 3.1 for removing the clinical context, so the model relies on the signal.
  • Baseline coverage (W5, Q2): we additionally evaluate GEM-7B, zero-shot and fine-tuned, in our response. LOTUS with a 0.6B backbone is ahead on every split except ECG-QA, on which GEM is a specialist.
  • Overstated claims and robustness (W6): we describe the TSRBench result as equal or better performance with a backbone 25 to 50 times smaller. Appendix A.2 and Table 5 extend the two main ablations to Qwen3-0.6B, and Appendix A.4 and Table 7 show that training on all ECG-QA examples does not help ECG-QA and lowers the six-split mean.

We hope these changes address the reviewers' concerns, and we are happy to answer any further questions during the rebuttal period.

AuthorsOpenReview ↗

Thank you for the reassessment

Dear Reviewer P8Q5,

We thank you for reassessing our work. We really appreciate your time and consideration.

Best regards,

The authors

Area_Chair_EgpxOpenReview ↗

Reminder

Dear Reviewer P8Q5,

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.