Enhancing LLM Reasoning for Time Series Classification by Tailored Thinking and Fused Decision

We sincerely appreciate the time and effort you have dedicated to providing insightful comments and valuable suggestions. We have carefully considered your comments and made the required improvements by conducting new experiments, providing clarifications, and performing additional analyses.

W1: Repeat experiments with multiple random seeds

We appreciate your valuable suggestion. Due to time constraints during the rebuttal period, we conducted five repeated trials of ReasonTSC with DeepSeek-R1 (using MOMENT as the plug-in model) on one subset (BME), recording both average accuracy and standard deviation in the table below. The results demonstrate stable LLM performance, with DeepSeek-R1 consistently outperforming MOMENT. We will conduct repeated trials for all experiments in the revised paper.

W2: Concerns for data leakage between training and test sets

We sincerely appreciate this important observation regarding potential data leakage. While using 3-shot samples from the test set could have introduced leakage and created an unfair comparison with MOMENT, we have addressed this concern by re-running all experiments using samples selected exclusively from the training set. The updated results below demonstrate that ReasonTSC with DeepSeek maintains its performance advantage over MOMENT, showing comparible results from our original results. This confirms the framework's capability to identify and correct plug-in model errors without test data dependency. Due to rebuttal time constraints, we will comprehensively update all relevant methodological descriptions and experimental results for each LLM in the final manuscript to reflect these corrections and ensure full experimental rigor.

W3: Clarify test set reporting method & evaluate performance of few-shot setting

We appreciate your thoughtful questions regarding the evaluation methodology. Figure 3 (Section 3.3) reports average performance across 9 UCR subsets using MOMENT as the plug-in model. For few-shot samples, we maintain consistency in the number of examples per category (e.g., in a 5-category subset, 1-shot means exactly one sample per category, totaling five examples). We will make these details more explicit in the revised paper.

We condcut new experiments within LLM context length constraints (using datasets with 80 time points per category). Under the setting of ReasonTSC with MOMENT and GPT-4o-mini, performance remains stable from 1-shot to 5-shot configurations, showing only minor degradation, which is consistent with the observation of Figure 9, where 1-shot and 3-shot results are nearly identical. The 2-shot configuration already achieves strong performance, demonstrating ReasonTSC's robustness in guiding LLMs for time series tasks. We will update the relevant figures and descriptions to better present these findings and ensure clarity.

S1 & S2: Bold font and reporting the percentage improvement over the reference model (MOMENT)

We appreciate these suggestions and will address related format issues in the revised paper.

Report Improvement over Plug-in Models： We will incorporate the improvement percentage metrics, as you suggested. The table below presents our preliminary results.

S3 Ablation study on decision correction

We have conducted a new ablation study comparing performance with and without the third reasoning stage on two subsets, as you suggested. We remove ReasonTSC's integrative reasoning in the third stage while retaining TSFM's knowledge and test cases. The results show performance declines in corrected predictions for both GPT and DeepSeek when omitting the third-stage reasoning, demonstrating the importance of reasoning strategy in comprehensive decision correction. We currently present preliminary results due to rebuttal time constraints, and we will expand this analysis to more datasets with repeated trials for robust average scores in the revised paper.

S4: Ablation study on reasoning stages

We have conducted a new ablation study examining the individual contributions of stages 2 and 3 in ReasonTSC, as you suggested. Our analysis reveals that removing either the second or third reasoning stage leads to noticeable performance degradation for both GPT and DeepSeek models. These findings confirm that each component of ReasonTSC's multi-stage architecture plays a crucial role in enhancing the model's reasoning capabilities. Due to time constraints during the rebuttal period, we completed this experiment on only one dataset. We will repeat this experiment across all 15 datasets used in the paper.

S5: More discussion of Vanilla CoT

We appreciate this constructive suggestion regarding vanilla CoT. The ICLR'25 paper "Can LLMs Understand Time Series Anomalies?" demonstrates that basic CoT (simply appending "please think step by step") yields limited performance for time series tasks. As shown in Tables 1 and 2 of the ICLR'25 paper, vanilla CoT in our framework refers to removing the TSC-tailored integrative reasoning in the third reasoning round, which also shows unsatisfactory results. To better highlight the differences, we will explicitly clarify vanilla CoT examples in the main paper. Further analysis in Figure 9 (Appendix D.2) reveals that ReasonTSC maintains substantial performance gains (average increase ratio on four subsets) even when removing key components like logits or TSFM predictions, achieving improvement ratios over twice those of vanilla CoT.

Q1: Selection criteria of datasets

The 15 datasets selected from the UCR/UEA Archive (detailed in Table 4, Appendix C.1) are chosen to provide comprehensive coverage across several dimensions: they span multiple application domains and exhibit substantial variation in key characteristics including class numbers (ranging from 3 to 15 categories) and time series lengths (from 80 to 288 time points per dimension).

Q2 Clarification of performance gain

The 10% improvement refers to the maximum performance gain observed across individual datasets, with the specific results for each subset provided below. We will clarify this point more explicitly in the revised manuscript.

Q3: Clarification of few-shot settings

Table 1 reports the best-case results for all models under a consistent 2-shot setting, where each category in a subset is provided with two samples. For example, a 5-category subset uses 10 samples in total within the ReasonTSC framework. We will clarify this in the manuscript to ensure clarity.

Q4: Analysis of plug-in model’s contribution

The performance decrease shown in Figure 6 (a) (Lines 311-312) reports the average accuracy across all 9 UCR subsets using MOMENT as the plug-in model. Specifically, removing the plug-in outputs leads to accuracy drops of 47.75% for GPT-4o-mini, 40.43% for LLaMA-3.3-70B-Instruct, and 25.65% for DeepSeek-R1. We will clarify this point in the manuscript to avoid ambiguity. Besides, we provide a detailed table below showing both the mean accuracy and standard deviation for each LLM in the ablation scenario.

We sincerely appreciate the time and effort you have dedicated to providing insightful comments and valuable suggestions. We have carefully considered your comments and made the required improvements by conducting new experiments, providing clarifications, and adding new references.

W1: Comparison with TimeCAP

Thank you for pointing us to the related work. TimeCAP presents a novel and effective framework that revolutionizes the role of LLMs in time-series event prediction. TimeCAP uniquely incorporates two independent LLM agents: the first generates rich textual summaries capturing domain-specific context, while the second leverages these summaries for highly informed predictions. A lightweight multi-modal encoder then produces joint embeddings of both the series and its textual summary, enabling the retrieval of highly relevant in-context examples from the training set to further refine the predictor’s prompt. Across seven real-world datasets, TimeCAP surpasses state-of-the-art competitors, achieving an impressive average F1-score uplift of 28.75%.

We will revise our main paper and cite TimeCAP in line 48:

TimeCAP [a] introduces a novel framework that enhances time-series event prediction by employing two specialized LLM agents. A multi-modal encoder further refines the process by retrieving relevant in-context examples to optimize the predictor’s prompts.

[a] TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI'25)

W2: Explanation of Eq.(2)-(4)

We appreciate your valuable suggestion. We will incoporate more explanations about these equations in the revised paper, as you suggested. In detial,

• Eq.(2) indicates the intermediate rationales generated by a reasoning LLM via a multi-run reasoning process. $J$ is the number of reasoning turns/steps, $X_j$ is the input training time series sample at the $j^{th}$ reasoning turn/step, and $\phi(X_j)$ is the tailored prompt based on the corresponding input sample. $P_\theta$ is a reasoning LLM that can be fine-tuned by the shown training samples. Note that the $j^{th}$ rationale $r_j$ is generated based on the $(j-1)^{th}$ rationale.

• Eq.(3) implies the final reasoning language model after the multi-run training process. This final model is obtained on the basis of all the intermediate rationales, input training samples, and tailored prompts.

• Eq.(4) indicates the enhanced time series classification based on the final reasoning language model depicted in Eq.(3). $x_t$ is the testing sample, and $\Phi(x_t)$ is the tailored prompt designed for the testing task.

W3: Reasoning outputs from LLMs

We appreciate your suggestion and will include more comprehensive and representative reasoning outputs from LLMs in the revised paper. In the current manuscript, we provide some typical reasoning outputs in Appendix B.2, showing the illustrative rationales generated by ReasonTSC with DeepSeek, Llama, and GPT in the third round. These illustrative generations cover three representative cases:

1. ReasonTSC with DeepSeek identifies and corrects the plug-in model’s biased prediction by analyzing its behavioral tendency;

2. ReasonTSC with Llama initially agrees with the plug-in model’s prediction but subsequently overrides it after detecting closer logit values and more representative temporal patterns in category 6;

3. ReasonTSC with GPT maitains consistency with the plug-in model’s final prediction after analysis of temporal characteristics and the category-wise logit distributions.

W4: Insights about Deepseek

In our results, it is observed that DeepSeek presents relatively better performance under the ReasonTSC framework compared to Llama and GPT. This might stem from DeepSeek's significantly larger parameter size (671B for DeepSeek-R1 vs. 8B for GPT-4o-mini and 70B for Llama) as well as its specialized reasoning-enhanced training. We also present additional details of LLMs we evaluated, including model parameters, release dates, and whether reasoning-focused post-training techniques were applied, in Table 5 (Appendix C.2).

W5: Include dataset statistics

We appreciate your suggestion and will include additional dataset statistics in the main paper. In the current version, we present the statistics of 15 representative subsets selected from the UCR/UEA Archive in Table 4 (Appendix C.1). These datasets cover diverse domains and vary in key characteristics (e.g., number of classes, time series length).

W6&7: Concerns about different input lengths

In our experiments, we evaluate time series samples with varying lengths (per category), ranging from 80 to 288 time points. This results in total input lengths (number of categories × length) between 160 and 4032 time points under the 2-shot setting, and token counts ranging from 1800 to 12000. Notably, on datasets with over 11k tokens (Arr.Hd, Eplp, and Dod.LD), ReasonTSC with DeepSeek achieves performance improvements of 26.03%, 43.98%, and 36.37% respectively, compared to vanilla CoT. Furthermore, Figure 12 (Appendix D.4 ) further demonstrates ReasonTSC ’s stability across different class counts, series lengths, and token counts. The results indicate that all three LLMs maintain stable performance as sequence length and token count increase. The slightly lower improvements observed on shorter series may stem from shorter time series samples containing fewer discernible patterns, which provide less information for LLM to understand patterns.

W8: Evaluate the performance of the few-shot setting

We appreciate your question and have conducted additional experiments to examine the impact of increasing the few-shot examples within LLM context length constraints (using datasets with 80 time points per category). Using MOMENT as the plug-in model with GPT-4o-mini under ReasonTSC, we observe relatively stable performance from 1-shot to 5-shot configurations, with only consistently slight degradations. These findings in consistent with Figure 9, where 1-shot and 3-shot performances are closely aligned. Notably, the 2-shot configuration already yields satisfactory results, demonstrating ReasonTSC's robustness in steering LLMs for time series reasoning tasks. We will revise the relevant figures and descriptions in the main paper to better reflect these findings and ensure clarity.

Q1 & W1.1: Generalization to multivariate time series

We have generalized our proposed method to multivariate time series cases and conducted a new experiment, as you suggested. The details of the generalization are as follows:

- We use <dim> tags to separate different dimensions within each sample; 

- We employ interval sampling on the time series data to mitigate extremely long sequences by concatenating all dimensions.;

- We adapt the prompt to guide the LLM in comparing dimensions separately across categories in the first round of pattern reasoning.

Experimental results: on Epilepsy (60 samples, 4 classes, 3 dimensions, 206 time points per dimension) with 1:4 interval sampling, it yields accuracies of 80.70% (DeepSeek-R1) and 73.68% (GPT-4o-mini), both improving upon the plug-in model MOMENT (71.93%). These new results demonstrate that our method generalizes to reasoning over multivariate time series and improves accuracy.

Cross-variable dependencies: We appreciate your insightful suggestion, which indeed has the potential to further enhance performance when effectively exploited. Due to the limited time during the rebuttal period, we only managed to complete new experiments based on the preliminary generalization strategy described above, which already demonstrate the effectiveness of our method. We plan to explore more in-depth strategies for exploiting cross-variable dependencies, for example:

- Concatenate all the variables with dimension ids;

- Claim the multivariate analysis requirement, e.g., Each sample consists of #Count-dimensional time series data, where each dimension is labeled with its dimension ID in the format. The time series for each dimension can be compared separately across categories.

It draws inspiration from existing literature (e.g., [a]) showing that attaching spatial (variable) and temporal (time-series pattern) identity information can help simple neural networks (MLPs) effectively capture the cross-dimensional spatiotemporal correlations in multivariate time series data.

[a] Separable spatial-temporal residual graph for cloth-changing group re-identification (TPAMI'24)

W2.2: Longer time series

Thank you for pointing out this valuable aspect. In our experiments, we evaluate time series samples with varying lengths (per category), ranging from 80 to 288 time points. This results in total input lengths (number of categories × length) ranging between 160 and 4032 time points under the 2-shot setting. Notably, on datasets with over 11k tokens (Arr.Hd, Eplp, and Dod.LD), ReasonTSC with DeepSeek achieves performance improvements of 26.03%, 43.98%, and 36.37% respectively, compared to vanilla CoT. Furthermore, Figure 12 (Appendix D.4) further demonstrates ReasonTSC’s stability across different class counts, series lengths, and token counts. The results indicate that all three reasoning LLMs maintain stable performance as sequence length and token count increase.

Q2.2 & W4.3 & C1: Further ablation studies to assess pattern importance

We conduct two new ablation studies to assess the importance of fundamental patterns, as you suggested.

First, we remove three top patterns identified by DeepSeek for each subset (Figure 5) and test with ReasonTSC using Chronos as the plug-in model. The results demonstrate that removing these typical pattern reasoning guidance leads to noticeable performance degradation across all subsets, confirming their critical role in the reasoning process. Second, we remove three other random patterns and observe a less pronounced performance drop.

These findings show that important fundamental pattern identification could potentially further enhance performance for certain datasets.

Q2.1 & W1.2: Further clarification of pattern selection rationale

Our pattern selection strategy is grounded in existing time series literature, which has proven the fundamental role and effectiveness of these patterns. Our selected patterns are consistent with existing literature, e.g., [b-c]:

[b] ChatTS: Understanding, Chat, Reasoning about Time Series with TS-MLLM (VLDB'25)

[c] Language Models Still Struggle to Zero-shot Reason about Time Series (EMNLP'24)

Q2.3: Discussion about domain-specific patterns

We appreciate your insightful suggestion, which is definitely an interesting future direction and could help our work be better implemented in related application areas. We will add the following discussion:

In this work, we propose a general reasoning framework for enhancing time series classification, where most fundamental time series patterns are included in the reasoning process. It is a promising future direction to further exploit domain-specific patterns during the reasoning process to improve domain-specific tasks.

Q3 & W1.4: Cost measurement

We measure the time and API costs of ReasonTSC and Vanilla CoT. Since time series data dominates the prompt length, the computational overhead and API costs between ReasonTSC and Vanilla CoT are comparable. However, ReasonTSC achieves significantly better performance and provides explainable rationales.

Q4 & W2.3: Robustness and failure analysis

We appreciate this insightful question. We have observed cases where LLMs failed despite the plug-in prediction being correct, often due to either misweighting patterns between categories or reasoning errors. For instance, GPT-4o-mini occasionally misaligned labels (e.g., some subsets use labels starting from 1, while GPT assumed labels starting from 0 during inference), leading to incorrect predictions.

Additionally, Table 3 shows cases where ReasonTSC successfully corrects plug-in model errors (override accuracy). Below, we also provide the fail override accuracy (instances where ReasonTSC contradicted the plug-in's prediction but produced the wrong answer). These metrics represent averages across 15 UCR/UEA subsets.

W3.2: Prompt sensitivity

We conduct a new experiment for prompt sensitivity analysis by asking DeepSeek-R1 to paraphrase ReasonTSC's original prompt while preserving its core meaning. Using this modified prompt, ReasonTSC with GPT-4o-mini achieves 62.59% (vs original 63.31%) on Dist.TW and 32.47% (vs original 31.17%) on Dod.LD, demonstrating strong robustness to prompt variations.

Q5&W2.1 Comparisons with Visual Reasoning Approaches

We generalize ReasonTSC to VLM and conduct new experiments, as you suggested. We compare with a latest work in VLM for TS[d], which proposes VL-Time by combining visualized time-series data with vanilla CoT. We integrate ReasonTSC's reasoning strategy into VL-Time by adapting pattern-based guidance for visual analysis. The results on the EMG dataset show that ReasonTSC substantially enhances VL-Time's reasoning capabilities, validating its flexibility across different time-series domains and tasks.

[d] A Picture is Worth A Thousand Numbers: Enabling LLMs to Reason About Time Series via Visualization.

W1.3: Discussion of n-shot selection

We acknowledge your insightful observation. We achieve current improved performance by random selection and plan to explore more sophisticated strategies in future work, e.g., based on embedding similarity (both most and least relevant samples) or choosing the class center of each category.

C2: Sensitivity to the quality of plug-in models

We appreciate your insightful suggestion. In this work, we propose ReasonTSC as a reasoning framework that enhances time series classification, accommodating any plug-in models and improving their performance, as shown in Tables 1&2 (using MOMENT and Chronos as plug-in models).

C3: Clarification about domain-specific knowledge in prompt

The domain-specific knowledge incorporated in our prompts is derived from the UCR/UEA Archive's documentation, which provides real-world brief descriptions of each dataset's domain along with explanations of category labels.

W3.1 & W3.3" Backtracking mechanism

We attribute the backtracking effectiveness primarily to fundamental differences in model capabilities, particularly parameter scale (671B for DeepSeek-R1 compared to 8B for GPT-4o-mini and 70B for Llama) and inherent reasoning capacity. We recommend using LLMs with stronger reasoning capabilities when implementing ReasonTSC.

W3.4: Datasets with many more classes

Our experiment covers a class range of 3 to 15 (Table 4 in Appendix C.1). We explicitly analyze the impact of category count on performance, as shown in Figure 12(a) (Appendix D.5), which illustrates the average performance increase ratio across three LLMs.

W4.4: Discussion about less obvious performance improvement

The observed differences likely stem from inherent data characteristics. Datasets with clear, simple patterns show more pronounced gains under ReasonTSC, whereas those with complex or subtle patterns (e.g., medical image contours) exhibit more limited improvements due to their inherent interpretability challenges.

W4.1 & 4.2: Presentation issues

We will thoroughly address these issues in the revised paper.

Q1 & W1: Qualitative and quantitative comparisons to existing TS reasoning methods

We appreciate the reviewer for raising this important point. In the main paper, we compare ReasonTSC with such a baseline approach from the ICLR'25 paper 'Can LLMs Understand Time Series Anomalies?', which utilizes vanilla Chain-of-Thought (CoT). While this compared work concluded that explicit reasoning fails to improve LLMs' time series anomaly detection capabilities, our results demonstrate that proper reasoning frameworks can indeed enhance LLMs' time series reasoning ability.

During rebuttal, we conduct new experiments to compare with the self-consistency and self-correction reasoning techniques from a more recent paper 'Evaluating System 1 vs. 2 Reasoning Approaches for Zero-Shot Time Series Forecasting', which is posted to arXiv on 14 March 2025, just Two months before NeurIPS Deadline. Our experiments with GPT-4o-mini show that naively applying these reasoning techniques, which rely solely on the LLM’s inherent reasoning, does not substantially improve performance. In contrast, ReasonTSC enhances time-series understanding by integrating TSFM’s knowledge as a plug-in module, demonstrating clear advantages over these baseline approaches.

Q2 Clarification about classification task considered in this work

The time series classification tasks are prevalent and significant in many real-world applications. With the advent of reasoning LLMs that present impressive performance in NLP and vision reasoning tasks, LLMs' reasoning ability should be naturally applied to the time series domain. Tasks like forecasting are limited by their evaluation metric (MSE), which leads to overlooking deeper model understanding. A model that simply outputs a nearly constant line might still achieve a passable MSE but reveals little about its capacity to interpret dynamics. On the contrary, classification tasks are suitable to evaluate a model's time series reasoning capabilities since models need to reason across categories and produce definitive decisions.

Re W3 Scalability of ReasonTSC

The ReasonTSC framework can be easily generalized to diverse time-series tasks, including vision-based datasets. For example, in the paper "A Picture is Worth A Thousand Numbers: Enabling LLMs to Reason About Time Series via Visualization", the authors propose VL-Time, a prompt-based approach combining visualized time-series data with vanilla CoT reasoning. To demonstrate generalizability, we integrate ReasonTSC's tailored step-by-step reasoning in the third round into VL-Time and evaluate it on the EMG dataset under zero-shot settings. The results (shown below) reveal that ReasonTSC substantially enhances LLMs' reasoning capabilities compared to baseline methods. This validates its flexibility across different time-series domains and tasks.

Re W2 Generalizability to other TS tasks

We conduct new experiments on the anomaly detection task, as you suggested. We compare ReasonTSC against the baseline (AnomLLM) across diverse anomaly-specific datasets from the study “Can LLMs Understand Time Series Anomalies?” . The results demonstrate how our work achieves robust generalization. The compared method (AnomLLM) relies on the basic prompt “Let's think step by step.” In contrast, under zero-shot settings, we replace this vanilla chain-of-thought (CoT) with ReasonTSC’s tailored step-by-step reasoning strategy, evaluating it on five critical anomaly-type datasets (range, freq, point, noisy-freq, and noisy-point) defined by AnomLLM. For each dataset (representing one anomaly type), we measured precision, recall, and F1-scores, both for standard detection and affine-transformed (Affi) scenarios. As shown in the results below, ReasonTSC with GPT-4o-mini outperforms AnomLLM across datasets, including Range, Freq, Noisy-freq, and Noisy-point, demonstrating its stronger generalization and underscoring the value of task-adaptive reasoning in time series anomaly detection.

We sincerely appreciate the time and effort you have dedicated to providing insightful comments and valuable suggestions. We have carefully addressed each point raised, and we hope our revisions and additional experimental results have adequately addressed your concerns.