- 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
- 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
-
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.
-
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
-
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?
-
What is the fundamental difference between the proposed supervision approach and Taco and SensorLM?
-
Can the authors demonstrate generalization on independent zero-shot benchmarks or downstream tasks?
-
Whhy does adding statistics based supervision reduce the TSQA performance relative to the TSQA-only training?
3 comments
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.
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.
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.