acceptodds
Under review as a conference paper at ICLR 2027

Decoupling Perception from Description: Computation-Grounded Representation Alignment between Multivariate Time Series and Language

Abstract

Training multimodal models to align time series with language runs into a self-supervision trap. The usual recipe asks an LLM to read a series and write a description, so label quality is capped by the perceptual skill the model is supposed to learn. The data can never teach more than the labeler already knows. A second gap makes this worse: most datasets use a single variable, but the patterns that matter (cross-channel correlation, lead-lag structure, co-occurring anomalies) appear only with several variables, right where the labeling LLM's limits are most exposed. These two problems create a trilemma: existing methods are reliable, realistic, or scalable, but none achieves all three. We resolve this by decoupling perception from description. Deterministic code computes a set of statistics from real, open-source multivariate series; the LLM verbalizes those precomputed facts. Perception, which LLMs do poorly, is handled by computation, while the LLM handles expression. This produces CGTime, our 4B-parameter computation-grounded time-series–language model. CGTime outperforms far larger general-purpose models on multivariate understanding tasks: it attains the best multivariate fact score on our held-out benchmark (0.283 vs. 0.173 for GPT-4o-mini and 0.203 for GPT-5.4-nano), a gap that survives Holm-corrected paired significance tests against every baseline. It also states verifiable numerical facts in generated captions more accurately and covers a broader range of statistical properties.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time Series Foundation Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FEj7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The candidate paper presents a novel model called CGTime, a 4 Billion parameter time-series language model that is trained on deterministically computed statistics. The authors argue that, instead of asking a large language model to interpret time series when generating labels, to compute 169 statistical properties from real multivariate time series data and use an LLM to verbalize them into training descriptions. The model is encompassed of a frozen Moment encoder, supervised fine-tuning, and GRPO with statistical rewards. The candidate model outperforms general-purpose and time-series models on their internal Metric-QA benchmark. The authors also report competitive results on TSQA.

Critical strengths
  1. The paper is written well and provides extensive implementation details. The statistical reporting is also careful and includes paired tests, multiple-comparison correction, and evaluation audits. The computed versus LLM perceived supervision ablation demonstrates improvements in CGTime to detect these statistical properties in their own internal Metric-QA benchmark.
Critical weaknesses
  1. The central learning objective lacks grounding and justification, and the obvious computational baseline is missing The paper puts forward the argument that deterministic computation of statistics is more reliable than LLM perception, yet the paper trains an LLM model to approximate these same computable statistics at inference time. Why should a model learn to recognize these properties rather than receive their exact computed values (using the authors' existing deterministic computation pipeline) and verbalize them? The authors already implement a pipeline that computes all the 169 statistics, but do not provide a compute-then-verbalize baseline or evidence of advantages in efficiency, generalization or downstream utility. The paper reports that CGTime achieves only 0.355 on Rel25 Metric-QA, leaving around two thirds of the answers outside the specified tolerance. The significance of the learning objective is unclear.

  2. The evaluation primarily demonstrates task specialization rather than time-series generalization Metric-QA and Captioning use the same statistical properties and scoring machinery as the training, while the external baselines do not receive any comparable task-specific supervision. TSQA is also not evaluated zero shot, as CGTime trains on 90 percent of the selected questions. Moreover, Table 15 in the appendix shows that TSQA-only final alignment achieves higher accuracy and ROUGE-L (0.712/0.299) than the full model (0.650/0.278). Thus, in the reported final-alignment ablation, the statistical based supervision that the paper proposes reduces performance on the only broader QA benchmark relative to TSQA-only training. There is no independent evidence that demonstrates transferable numerical reasoning.

Questions for authors
  1. What advantage does learning to approximate deterministic statistics offer over computing the same statistics at runtime and verbalizing them directly? Can the authors provide this baseline with accuracy and inference-cost comparisons?

  2. What is the fundamental difference between the proposed supervision approach and Taco and SensorLM?

  3. Can the authors demonstrate generalization on independent zero-shot benchmarks or downstream tasks?

  4. Whhy does adding statistics based supervision reduce the TSQA performance relative to the TSQA-only training?

3 comments

Thank you for your thoughtful feedback on the learning objective and generalization.

The questions concern the purpose of learning from computed statistics and how the resulting capabilities transfer beyond the constructed tasks. We discuss the learning objective and present additional evaluations below.

Q1. Statistical Supervision and Reasoning

Deterministic computation provides exact statistics, while CGTime focuses on developing time-series analysis capabilities beyond numerical computation. We use computed statistics as the proxy for supervision and CoT to guide the model in organizing metric definitions, signal features, intermediate judgments, and final conclusions into a coherent analysis. Statistical accuracy is both a training objective and a verifiable measure of whether the model has learned to base its analysis on the input signal. CGTime thus provides comprehensive analysis through CoT-guided reasoning, beyond numerical accuracy.

The Rel25 score is considered a complementary measure of numerical approximation. CGTime’s advantage over the baselines in Table 2 holds under both Gaussian and Rel25 scoring.

Using the non-thinking Qwen3-4B-Instruct-2507 as the backbone, we compare its outputs before and after training on the same time-series input and question. In the PCA case shown in the paper, Qwen before training mainly infers the number of principal components through qualitative observations and an assumed allocation of variance, arriving at a final answer of 5. CGTime instead gives a threshold-based judgment using cumulative explained variance during generation and selects 4, matching the computed reference answer. Their complete outputs are provided in Appendix R.4 of the revised manuscript.

Q2. Differences from TACO and SensorLM

TACO, SensorLM, and CGTime all construct text supervision from computed features. TACO converts features into qualitative descriptions; SensorLM aligns time series and text through contrastive and text-generation losses. Both emphasize channel-wise statistics and temporal patterns, with limited analysis of cross-channel and system-level relationships. CGTime uses PCA to capture multivariate system-level and cross-channel relationships.

What is more, these statistics provide QA and Caption targets for SFT and also serve as RL references for verifying generated numerical facts and computing rewards. Thus, multivariate statistics provide both the supervision content and the basis for optimizing numerical accuracy.

Q3. Independent Zero-Shot Evaluation

We have added zero-shot evaluations on TemporalBench and TSRBench, both using the original CGTime-4B checkpoint without additional fine-tuning on these benchmarks. All TemporalBench models were re-evaluated; for TSRBench, we compare our CGTime results with baseline results published in the paper. Detailed results are presented in Table 4 of the revised manuscript.

These results support the transfer of CGTime's learned capabilities to independent time-series understanding and reasoning tasks.

Q4. Mixed Training and Task Proportions

We hypothesize that this difference is related to the trade-off between task proportions and optimization objectives in mixed training. TSQA-only concentrates final-alignment training on TSQA, whereas joint training combines statistical QA, factual captioning, and TSQA, changing the model's relative exposure to each task.

To examine the effect of task proportions, we added two mixture settings, each using 60,000 records per stage and configured with the same training budget. The ratios below follow Metric-QA:Caption:TSQA. We report their TSQA results alongside the original results in Appendix L.1 of the revised manuscript.

Training mixtureTSQA shareTSQA ChoiceTSQA ROUGE-LSource
2:2:120%0.6300.266New experiment
Original mixture36.4%/37.2%0.6500.278Original paper
1:1:250%0.6540.266New experiment
TSQA-only100%0.7120.299Original paper

TSQA shares refer to training-pool record counts; the original mixture's shares are given for SFT/GRPO. Original results retain their original datasets and training settings. Choice is the macro-average accuracy of multiple-choice and true/false questions.

In the two added settings, increasing the TSQA share from 20% to 50% raises Choice from 0.630 to 0.654, while open-ended ROUGE-L remains approximately 0.266. These results support task proportions as a factor affecting TSQA choice performance and show different responses to the mixture change across TSQA subtasks. The optimization requirements of different tasks may also compete, affecting task-specific TSQA performance. Quantifying how task proportions affect performance and how to balance statistical factual accuracy with TSQA performance remains an open question worth further investigation.

I thank the authors for attempting to address my concerns. The zero-shot addition is a nice improvement, as it further illustrates the capabilities of the trained model. However, one of my main points was that the clear baseline I asked for is still missing. I wanted to see a compute-then-verbalize baseline, where these deterministically computable statistics are computed and then verbalized by an LLM, and how that LLM would perform compared to CGTime. This is a very simple baseline, since all of its components already exist in the authors’ pipeline. The one possible benefit I can see of training is that internalizing the statistics could improve the model’s reasoning, but I have not seen evidence of that. Since TemporalBench T1 asks specifically about trend, seasonality, anomalies and other properties, it is not surprising that a model trained specifically on these properties performs well. I keep my score.

Thank you for your questions.

Compute-then-verbalize. This baseline provides a natural upper bound on Metric-QA and Caption because both tasks score the same statistics that our pipeline computes. When the statistics of interest are known, neither CGTime nor other time-series reasoning models can replace direct computation in terms of numerical accuracy, since direct computation yields the exact values. Numerical accuracies on our tasks measure how well the model has learned to recognize statistical properties from supervision based on computed values.

The value of learning. The open question is whether this ability helps the model answer questions whose answers are not predefined statistics. We agree that TemporalBench T1 tests properties the model saw during training. TSRBench Event Prediction has less overlap with those properties. On that task, CGTime achieves 60.56% zero-shot accuracy, compared with the published 55.0% for the strongest time-series-specific baseline.

Scope of the contribution. Our central claim concerns supervision: for models that learn to perceive time series end to end, computed statistics provide more reliable labels than an LLM's own reading of the series (Table 16). Our aim is to develop the model's ability and intuition for working directly with high-dimensional time series rather than to require high-precision calculation. The quantitative and qualitative results illustrate this ability: our 4B model outperforms GPT-6.1 Terra, a strong baseline, in the multivariate setting, and its chains of thought show how it reasons about time series.

Reviewer_kRuSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CGTime for statistical question answering and captioning of multivariate time series. The authors compute statistics from real data and ask an LLM to turn them into descriptions for training. These statistics also provide a way to check numerical claims during reinforcement learning. CGTime connects a time-series encoder to a language model and is trained with SFT and RL. The experiments compare it with general-purpose LLMs and existing time-series–language models. The authors report higher statistical QA scores and better coverage of numerical facts in captions, with ablations on the supervision source and training stages.

Critical strengths

The paper presents a practical approach to constructing statistically grounded supervision from real multivariate time series. Deterministic computations supply reference statistics, including cross-channel relationships, which the teacher LLM uses to generate textual descriptions. This reduces reliance on the teacher’s ability to interpret numerical data and provides a factual basis for both supervised training targets and reinforcement-learning rewards.

Critical weaknesses

The evidence does not sufficiently distinguish learned multivariate analysis from adaptation to the constructed response patterns and statistical priors. Figure 2 shows that external computations supply the statistical facts used by the teacher LLM, whereas the student must generate answers without those computed results (§2.3). Although Tables 1–2 demonstrate improved numerical agreement, the ablations do not clearly establish what analytical capability the student acquires: the zero Metric-QA score without Joint-SFT in Table 16 is attributed to format failure, while Table 3 tests dependence on the time-series input only for GPT-4o, not CGTime. Thus, the claim that CGTime learns to reconstruct complex statistics remains insufficiently supported.

Questions for authors
  1. Appendix L.1 notes that the GRPO prompts were not matched between the two supervision variants in Table 15. Could you clarify whether the training sets were comparable in size and task/statistical-property distributions, and whether both variants used the same RL training budget?

  2. Could you evaluate CGTime with mismatched time-series inputs and with controlled perturbations that alter cross-channel relationships while preserving per-channel marginal statistics? For the perturbed series, comparing predictions against recomputed reference statistics would help establish whether the model’s answers track the actual multivariate structure.

3 comments

We appreciate your detailed questions about the supervision design and the model’s use of time-series inputs.

We clarify the supervision design and provide additional experimental evidence below. CGTime obtains the numerical facts used for supervision through deterministic computation. The LLM verbalizes these facts into natural-language training targets and does not affect the numerical facts themselves.

CGTime also achieves 58.12% zero-shot accuracy on TemporalBench T1 and 60.56% zero-shot accuracy on TSRBench Event Prediction without additional fine-tuning, supporting transfer beyond our constructed statistical tasks rather than being adapted to them, as discussed in Table 4 of the revised manuscript.

Q1. Supervision Consistency and the Prompt-Matched Comparison

The two variants in the lower panel of Table 16 start from the same Replay checkpoint and use the same set of SFT inputs. They differ only in the supervision targets: the computed-statistics variant uses targets derived from deterministically computed statistics, whereas the GPT-perceived variant uses GPT-5-nano generations. For GRPO, the GPT-perceived training set was built from its own teacher-generated examples, so that its reward references are the numerical values extracted from the same teacher outputs used in SFT. This keeps SFT and GRPO consistent within each variant. ReFT [1] follows the same design and reuses the same training questions and reference answers for RL after SFT, and Sun et al. [2] use same-source data across SFT, reward-model training, and RL. As a result, the GRPO prompts of the two variants were not matched prompt by prompt. This mismatch comes from how the data were constructed, not from the ablation itself. The two GRPO training sets are comparable in size and task distribution, and both variants use comparable RL training budgets.

To rule out the possibility that the gap comes from the GRPO prompts rather than from the supervision source, we retrain the GPT-perceived variant with a GRPO set matched prompt by prompt to that of the computed-statistics variant. The raw time series, questions, and task and statistical-property distributions are therefore identical across the two variants in GRPO. Together with the shared SFT inputs, this means the two variants now differ only in the source of supervision, where the advantage of computed-statistics supervision still holds:

Supervision sourceQA overallQA multivariateCaption PCaption R
Computed statistics (original result)0.2690.2820.3830.173
GPT-perceived (original result)0.1620.1570.1210.041
GPT-perceived (prompt-matched, new)0.13970.13390.12380.0482

References:

[1] Trung Quoc Luong et al. ReFT: Reasoning with Reinforced Fine-Tuning. ACL 2024.

[2] Haoran Sun et al. Curiosity-Driven Reinforcement Learning from Human Feedback. ACL 2025.

Q2. Channel-Order Permutation

We test channel-order robustness on the same 1,000 multivariate series using five different random permutations, preserving each channel's values and temporal order. The model generates QA answers and captions for each permutation, scored against the original reference statistics. The table compares original-input scores with the mean and standard deviation across five rounds.

Input settingQACaption PCaption R
Original input0.28320.31540.2020
Channel permutation (mean±standard deviation over five rounds)0.2831±0.00960.3134±0.00180.2008±0.0014

The aggregate scores across five permutation rounds are close to those for the original input, indicating that the model's overall statistical prediction performance is robust to channel presentation order.

We select a caption example of a 5 channel, 535 time step series as an example of consistency in the main structure described. For the original input and all five permutations, the model describes a single dominant factor, highly synchronized variables, and strong volatility comovement. (The generated captions are presented in Appendix R.3 of the revised manuscript.) The corresponding key statistics are shown below; the permutation column gives the range of numerical values in the five captions:

StatisticComputed referenceOriginal captionCaptions across five permutations
PC1 explained variance ratio90.99%95.7%95.5% to 95.7%
Effective rank1.001.001.00 in all five
Mean synchronization R²0.90990.960.96 in all five
Mean volatility correlation0.62970.6560.650 to 0.667
Proportion of pairs with high volatility correlation100%100%100% in all five

This case shows that the model's judgments about the main multivariate structure remain consistent after channel reordering, with similar numerical estimates.

These results are coherent with the model's non-causal cross-channel attention in our alignment module, which allows each channel to attend to all other channels.

Thank you for the clarifications and additional experiments. The prompt-matched GRPO comparison substantially addresses Q1 and strengthens the evidence for the benefit of computed-statistics supervision.

However, Q2 remains unresolved. My concern was whether CGTime’s predictions track changes in cross-channel relationships when per-channel marginal statistics are preserved. The reported experiment only reorders channels, leaving the underlying multivariate structure and the reported reference statistics unchanged. Stable predictions therefore support robustness to channel presentation order, but cannot distinguish input-grounded analysis from reliance on learned response patterns or statistical priors. The requested mismatched-input evaluation of CGTime is also absent.

The zero-shot results provide useful evidence of transfer beyond the constructed tasks, but do not directly establish that predictions depend on the relevant structure of the actual time-series inputs. Likewise, non-causal cross-channel attention provides a mechanism for accessing other channels, rather than evidence that the model uses this information to recover the queried statistics. My central concern about the evidence for learned multivariate analysis therefore remains.

Thank you for your clarification.

We additionaly tested two variations: replacing the whole series and applying independent shifts on each channel, and using mismatched inputs for the same question. In each case, we compared the answer CGTime generated from the new input with reusing the answer it had given for the original series.

We kept the question fixed and either replaced the entire series with another of the same shape or shifted each channel independently by up to 10% of the series length. The shifts preserve each channel’s empirical marginal distribution while changing cross-channel alignment. In each comparison, both answers were scored against the same references recomputed from the new input.

ExperimentMetric-QACaption PCaption R
Whole-series mismatch
Reuse the original answer0.29610.29710.1919
Generate a new answer from the new input0.29910.35480.2252
Independent circular shifts
Reuse the original answer0.21940.20350.1318
Generate a new answer from the new input0.23820.22680.1455

For both variations, answering from the new input yields higher Metric-QA scores and higher Caption precision and recall than reusing the original answer. The new answers better reflect the new time series CGTime receives. This shows that CGTime is using the time-series information to generate its answers rather than relying solely on memorized patterns from previous inputs.

Caption example: independent shifts.

InputCGTime captionComputed mean cross-channel volatility correlation
Original series"The mean volatility correlation is 0.639, with a standard deviation of 0.150. [...] When one variable fluctuates sharply, the others tend to respond in sync."0.682
After independent circular shifts"The mean volatility correlation is 0.211, with a standard deviation as high as 0.259. [...] Some variables do not fluctuate in sync."0.117

All 16 statistics for this example. The prompt was the same for both inputs. Caption values are the numbers explicitly stated in the original Chinese text; a dash means no numerical value was given for that statistic. Reference values are rounded to three decimals.

StatisticOriginal referenceOriginal captionShifted referenceShifted caption
Mean cross-channel volatility correlation0.6820.6390.1170.211
SD of cross-channel volatility correlations0.1140.1500.2600.259
Volatility synchronization index0.4780.4120.0820.080
Variance fraction explained by PC10.9240.9630.5950.857
Effective rank1.0001.001.7771.86
Mean synchronization R²0.9240.960.8060.86
SD of synchronization R²0.036-0.1580.07
Stability of rolling correlations0.9620.9740.8260.883
Mean temporal SD of correlations0.0260.0300.112-
Mean change in correlation matrix (Frobenius norm)0.0900.1300.3870.392
Stability of synchronization over time0.811-0.834-
PC1 trend R²0.1310.000.0120.01
PC1 regime shift magnitude0.5580.000.315-
PC1 periodicity strength0.1440.800.1350.706
PC1 seasonality strength0.753-0.6000.000
Seasonality consistency0.8120.8510.842-

Caption scores for this example. Both captions are scored against the same 16 reference values from the shifted series. The checked columns use the same scoring formula after verifying the extracted statistics and values against each caption.

CaptionAutomatic PChecked PAutomatic RChecked R
Reuse the original caption0.19060.17590.14290.1429
Generate a new caption from the shifted series0.40330.47050.35290.3529
Reviewer_QQbKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper decouples perception from description to support the construction of higher-quality training data and provide verifiable supervision. Following this principle, the authors construct a dataset and train CGTime through multiple stages. They evaluate the model against proprietary, open-source, and specialized time-series models on internal and external benchmarks. CGTime achieves strong results on the internal statistical tasks and competitive performance on TSQA after task-specific training.

Critical strengths

Separating statistical computation from language generation offers a practical way to construct large-scale training data with explicit, verifiable supervision. This approach has potential for scaling time-series–language models while reducing dependence on the perceptual capabilities of the labeling model.

Critical weaknesses
  1. The paper requires substantial improvement in presentation. Key information about the dataset, training procedure, and ablations is deferred to the appendix, while some details remain insufficiently explained. This makes the main text difficult to follow and its contributions difficult to assess.
  2. The experiments do not fully substantiate the motivation and claims presented in the abstract and introduction.
Questions for authors
  1. The paper states that “existing methods are reliable, realistic, or scalable, but none achieves all three.” What experimental evidence demonstrates that the proposed pipeline simultaneously satisfies these three objectives?
  2. In the external evaluation. CGTime uses a substantial amount of TSQA training data during both SFT and RL, yet its accuracy remains below GPT-4o. How should this comparison be interpreted given the differences in task-specific training? To what extent do the TSQA results demonstrate the effectiveness of the proposed pipeline beyond adaptation to the target benchmark?
  3. PCA reduces the complexity of describing multivariate relationships, but its contribution is not sufficiently analyzed.
  4. The proprietary models used in the internal evaluations are primarily mini or nano variants, and the supervision-source comparison in Table 15 also relies on a nano teacher. It remains unclear whether the conclusions hold with more capable general-purpose models as baselines and teachers.
4 comments

Thank you for your careful review and suggestions for improving the presentation and evaluation.

To make the contribution and experimental evidence easier to assess, we have clarified the data pipeline and training setup in the revised manuscript by adding Training Data and Training Procedure paragraphs at the beginning of Section 3.1. The main text also summarizes the ablations of supervision source, Joint-SFT, Joint-GRPO, data mixing, and curriculum learning. These ablations evaluate how computed supervision and the training recipe contribute to downstream performance.

Q1. Reliability, Realism, and Scalability

CGTime separates statistical computation from language generation: computed statistics are verbalized into QA pairs and captions, providing numerical targets for SFT and verifiable reward references for RL.

Reliable means that CGTime's supervision has a verifiable numerical basis unlike those generated by generic LLMs. As an example, for a 10-channel, 691-step series, we compare Astra's description generated from the raw signal with an existing training label, reproduced in Appendix R.1 of the revised manuscript. Of Astra's 28 numerical targets, 16/28 fall outside the specified tolerance.

Realistic means that both the statistics and labels are based on real observed series rather than signals synthesized from predefined distributions.

Scalable means that computation and label generation are automated and batched, without manual labeling. We construct approximately three million observations, with one million each for univariate, bivariate, and multivariate data. Each pairs a real time series with computed statistics and a generated description.

Q2. TSQA Performance and Zero-Shot Generalization

We hypothesize that GPT-4o's strong general-purpose language and reasoning capabilities may help it extract cues from time-series values presented as text and make TSQA-relevant judgments, enabling it to perform well on TSQA even without a dedicated time-series encoder. We include GPT-4o as a strong external performance reference for smaller time-series models. CGTime's central contribution is a data-generation and computed-supervision pipeline for teaching models to analyze time series. Numerical accuracy is an important proxy for evaluation, but not the endpoint: through CoT-guided training, we aim to help models connect statistical evidence with coherent, step-by-step analysis and conclusions.

CGTime is trained with TSQA data, so to further evaluate CGTime's generalization capabilities beyond adaptation, we conduct independent zero-shot assessments on TemporalBench and TSRBench, detailed results are presented in Table 4 of the revised manuscript.

CGTime leads the local TemporalBench T1 comparison and exceeds published time-series-specific models' results on TSRBench event prediction task, supporting transfer to feature understanding and event prediction beyond TSQA.

Q3. The Role of PCA

We use PCA here because it helps reduce the complexity of constructing multivariate time-series descriptions.

PCA organizes shared variation across C channels into dominant factors rather than describing all C(C-1)/2 variable pairs. In a 10-channel example (shown in Appendix R.2 of the revised manuscript), the first three components explain 84.46% of the variance, and the original training label uses 845 tokens to describe factor structure, synchronization, and system dynamics. The non-PCA description generated by Astra needs to cover 45 pairs with 7 metrics each (315 numerical values) and uses 4,248 tokens, approximately 5.0 times the original length.

Q4. Stronger Baselines and Teachers

To test if our conclusion still holds under a stronger general-purpose model, we evaluated GPT-6.1 Terra as a stronger baseline on the paper's 2,000 Metric-QA questions and 2,000 Caption requests. Evaluation uses Gaussian scoring with bandwidth multiplier m=0.5 for Metric-QA. Appendix S reports the full-set and filtered QA results, and Caption results.

Terra usually refuses to produce an approximation when it can't give an accurate answer. Both models use the same 1,424-question QA subset that Terra was able to answer. Caption results use all 2,000 requests.

ModelQA overallQA multivariateCaption PCaption R
CGTime-4B0.3050.3490.3990.186
GPT-6.1 Terra0.5090.2600.3800.163

As shown above, CGTime still maintains competitive performance on multivariate QA and Caption precision, even when compared against a stronger baseline like GPT-6.1 Terra on the filtered QA set. Detailed results and analysis can be found in Appendix S, Tables 31 and 32 of the revised manuscript.

A stronger teacher model may not change this outcome. Consider the Astra generated inaccurate supervision in the Q1 case (Appendix R.1), if trained on such supervision, the model would inherit these errors, harming the accuracy and may set an upper bound.

Dear authors,

Thank you for the revisions and additional experiments. However, three concerns remain:

  1. The explanations of realism and scalability are reasonable but lack direct experimental validation.
  2. PCA reduces description complexity, but its trade-off with downstream performance remains unclear.
  3. A single Astra example is insufficient to assess whether a stronger teacher could provide better supervision.

I will therefore maintain my current score.

Best Regards,

Reviewer

Thank you for your clarifying your concerns.

  1. Realism and scalability. It's hard to find a quantitative standard for either in this setting. We assess realism by considering how the data were obtained, it was generated from real, open source time series datasets. For scalability, a useful point of comparison is the time required for manual annotation. In a similar univariate time-series annotation task[2], each series took about 3 minutes to annotate, or about 20 series per person per hour. At that rate, one person would need 6 years of continuous work to annotate a dataset of 1 million. Our dataset also includes multivariate series, which are harder and more costly to annotate by hand.

  2. The role of PCA. Without PCA, the number of pairwise relationships to describe grows quadratically with the number of channels. One way to illustrate the cost would be to compare token counts for the same 8-channel series: one description using PCA and another covering every channel pair. We would expect the pairwise description to be much longer and to exceed the proper context window for training.

    The trade-off is between preserving the details of each pairwise relationship and describing how the channels vary together. To analyze a complex multivariate system, we can examine the eigenvalues and eigenvector directions of its covariance matrix without inspecting every pair. This is why we use PCA: each eigenvalue measures the variance explained by a mode, and the direction of its eigenvector shows how the channels contribute to that mode [1]. In analyses of stock returns, for example, this approach distinguishes market-wide fluctuations from sector-specific patterns [3]. In CGTime, it lets us describe a shared pattern across channels without listing every pairwise correlation.

  3. Accuracy of the training labels. Our example shows that even a flagship model such as Astra struggles to report accurate statistics from a high-dimensional time series, thus it does not provide better supervision in terms of numerical accuracy. Computing those statistics directly preserves their accuracy. This difference in the training labels sets the upper bound on model performance.

[1] Hotelling (1933), Analysis of a Complex of Statistical Variables into Principal Components (https://doi.org/10.1037/h0071325) [2] Tang et al. (2023), Vistext: A benchmark for semantically rich chart captioning (https://vis.mit.edu/pubs/vistext.pdf) [3] Plerou et al. (2002), Random Matrix Approach to Cross Correlations in Financial Data (https://journals.aps.org/pre/abstract/10.1103/PhysRevE.65.066126)

Dear authors,

Thank you for the further clarification. However,

  • Reliability, realism, and scalability. Thank you for the additional examples. I suggest adding a comparison table showing which of these three aspects are covered by prior methods and how CGTime addresses all three, with clear criteria for each aspect.

  • PCA. PCA’s trade-off between description complexity and downstream performance still lacks experimental validation.

  • Stronger teachers. Only single example‾\underline{\text{Only single example}} is insufficient to conclude that stronger teachers cannot improve supervision quality.

Best Regards,

Reviewer

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.