TS-Reasoner: Aligning Time Series Foundation Models with LLM Reasoning

• TSFMs capture temporal patterns but generally lack textual understanding, while LLMs reason over text yet struggle with temporal dependencies and numerical values (Goswami et al., 2024; Das et al., 2024; Fons et al., 2024; Merrill et al., 2024).The paper frames their complementary strengths as motivation for combining numerical time-series representations with contextual information.Source: §1 · pp. 1–2, Figure 1 · p. 2
• TS-Reasoner extracts temporal representations with a TSFM and uses a TS-to-Text adapter to project them into the LLM’s input embedding space.This is intended to integrate the TSFM’s temporal understanding with the LLM’s linguistic and reasoning capabilities.Source: §1 · p. 2
• Training first aligns the models by generating textual captions of time series, using an LLM/VLM prompting strategy to curate diverse, high-quality caption data, and then fine-tunes for reasoning.The introduction presents caption curation as a response to a training-data bottleneck.Source: §1 · p. 2
• The paper’s stated contributions are a formulation connecting TSFMs and LLMs for contextual time-series reasoning, a captioning method to diversify alignment data, and empirical analysis of existing approaches.The evaluation covers understanding and reasoning on TimeSeriesExam (Cai et al., 2024a) and MTBench (Chen et al., 2025).Source: §1 · pp. 2–3
• TS-Reasoner significantly outperforms open-source LLMs, VLMs, and TSLLMs on TimeSeriesExam and MTBench, with analyses supporting advantages in generalization, scalability, and training-data efficiency.The introduction reports these comparisons but provides no numerical results.Source: Figure 2 · p. 2, §1 · pp. 2–3

LLMs for Time Series

• LLM approaches prompt forecasting, textualize or align time series, or use vision-based models, but reported reasoning difficulties motivate context-enabled methods and TS-Reasoner’s pretrained-TSFM grounding (Merrill et al., 2024; Chow et al., 2024; Zhang et al., 2025a; Xie et al., 2024).Prior work also uses statistical models and neural networks for forecasting, while some LLMs show non-trivial forecasting performance despite struggling with time-series reasoning.Source: §2 · p. 3

Modality Alignment

• Existing time-series/LLM alignment work enables understanding but targets narrow domains and tasks and trains encoders from scratch, whereas this work adapts VLM alignment using pretrained TSFMs (Xie et al., 2024; Zhang et al., 2025a).It identifies integrating time-series characteristics into LLMs and the shortage of time-series-text pairs as challenges.Source: §2 · p. 3

Time Series Foundation Models

• TSFM research includes masked pretraining (Zerveas et al., 2021; Nie et al., 2022), large-scale forecasting models, MOMENT’s multitask capabilities (Goswami et al., 2024), and ChronoSteer’s use of LLM revisions to improve TSFM forecasting (Wang et al., 2025a).The passage describes large-scale pretraining as improving forecasting performance.Source: §2 · p. 3

3.1 Problem Formulation

• Given natural-language context and corresponding time series, the model generates an output composed of an intermediate reasoning path and a final answer.The answer is conditioned on the reasoning path, which explains the derivation from the multimodal context.Source: §3.1 · pp. 3–4

3.2 Overview

• TS-Reasoner combines a pretrained TSFM, a pretrained LLM, and a trainable TS-to-Text adapter in a two-stage alignment-pretraining and instruction-tuning pipeline.The TSFM encodes time series, and the adapter projects its features into the LLM’s input space.Source: Figure 3 · p. 4, §3.2 · p. 4

3.3 Model Architecture

• Value-preserving normalization retains each series’ mean and scale as text; TSFM-encoded non-overlapping patches pass through an MLP adapter and are interleaved at placeholders in the LLM context.The fused input accommodates multiple time series and enables the LLM to model inter-series and text-series dependencies.Source: §3.3 · p. 5

3.4 Training Recipe

• Captions are generated from time-series plots using attribute-enriched prompts and paraphrases, with 10K series each sampled from synthetic Chronos (Ansari et al., 2024) and context-attributed data (Merrill et al., 2024).The attribute categories include trend, frequency, periodicity, noise, and local variation.Source: §3.4 · p. 6
• Instruction tuning on Xie et al. (2024) Q&A and instruction-following data develops complex-instruction adherence, structured responses, and context-driven time-series reasoning, while the TSFM remains frozen throughout training.The LLM remains trainable in both stages, preserving the TSFM’s pretrained temporal knowledge.Source: §3.4 · pp. 5–6

4.1 Main Results

• On TimeSeriesExam overall, TS-Reasoner exceeds the best comparable-scale open-source LLM, VLM, and TSLLM by 8.17, 5.82, and 4.11 points, respectively, and leads MTBench baselines by 1.52–3.28 points.These comparisons concern comparable-scale open-source models; GPT-4.1 (vision) remains ahead on TimeSeriesExam overall accuracy, while TS-Reasoner is competitive with larger ChatTS-14B.Source: Table 2 · p. 8, §4.1 · p. 8
• +3.33 points on Pattern Recognition and +7.28 points on Noise Understanding mark TS-Reasoner’s gains over the best comparable-scale open-source baseline; financial and weather reasoning gains are 1.52–3.28 points.The authors attribute the gains to robust alignment and better reasoning over numerical patterns, while noting that complex multi-step reasoning remains limited by the 7B backbone’s capacity.Source: §4.1 · p. 8

4.2 Analysis of Data Scaling and Efficiency

• 60K alignment samples achieve superior overall accuracy to ChatTS-7B using less than half its data, while 10K instruction-tuning samples suffice to outperform it.The paper attributes this data efficiency to the pretrained TSFM and effective alignment, which provide a robust temporal foundation.Source: §4.2 · pp. 8–9, Figure 5 · p. 9

4.3 Choices of Captioning Model for Alignment

• Caption quality tracks the captioning model’s capability: GPT-4.1 captions outperform InternVL3-8B captions, which outperform Qwen2.5-VL-7B-Instruct captions across both benchmarks.The authors attribute GPT-4.1’s stronger results to its advanced time-series understanding, and InternVL3-8B’s edge over Qwen2.5-VL-7B-Instruct to better time-series understanding.Source: §4.3 · p. 9, Figure 7 · p. 10

4.4 Choices of TSFM and LLM in TS-Reasoner

• On TimeSeriesExam, replacing TimesFM reduces overall accuracy from 54.83% to 45.74% with MOMENT or 53.21% with Chronos; TS-Reasoner also leads Qwen2.5 baselines at 0.5B, 3B, and 7B.The respective Overall Accuracy pairs are 29.49% vs. 26.34%, 47.71% vs. 38.01%, and 54.83% vs. 46.66%; the TSFM’s gain is smaller at 0.5B and grows at larger LLM scales.Source: Figure 6 · p. 9, Table 3 · p. 10, §4.4 · p. 10

4.5 Comparison between Textual and Visual Time Series for Captioning

• Visually derived captions outperform text-derived captions across most tasks, which the authors attribute to plots’ compact, holistic representation supporting higher-fidelity captions and more effective alignment.Appendix B’s qualitative case analysis also reports more accurate pattern recognition from visual time-series representations.Source: §4.5 · pp. 10–11, Figure 8 · p. 11

4.6 Caption Diversity Analysis

• Attribute-aware captions achieve an MTLD score of 133.30 versus 42.95 for template-based captions, while Self-BLEU-4 falls from 0.82 to 0.45.The study evaluates lexical diversity on random samples of 1K captions from each dataset, interpreting the results as greater caption expressiveness and diversity.Source: Table 4 · p. 11, §4.6 · p. 11

4.7 Open-ended Time Series Reasoning

• On open-ended inductive reasoning, TS-Reasoner-7B outperforms all listed baselines, exceeding GPT-4o and ChatTS-7B by 21.1 and 4.5 points, respectively.Evaluation uses RAGAS, with 98% agreement between the LLM evaluator and human judgments in a manual inspection of 100 sampled cases.Source: Table 5 · p. 11, §4.7 · pp. 11–12

5 Conclusion

• TS-Reasoner substantially outperforms comparable-scale open-source LLMs, VLMs, and TSLLMs on time-series understanding and reasoning benchmarks.The conclusion reports extensive experiments but gives no specific metric values here.Source: §5 · p. 12
• TS-Reasoner bridges TSFMs and LLMs, while attribute-aware captioning enriches alignment data to mitigate the intrinsic semantic gap.The framework targets LLM understanding and reasoning about time series.Source: §5 · p. 12
• As a locally deployable open-source model, TS-Reasoner suits privacy-sensitive settings and can cost less computationally than large proprietary multimodal models.The conclusion notes that finance and healthcare regulations may prohibit sending time series to external cloud services; deployment choice also depends on privacy, cost, and domain-specific adaptability.Source: §5 · pp. 12–13

A TimesFM for time series embedding

• Each time series is normalized to zero mean and unit variance, then divided into consecutive, non-overlapping fixed-length patches.The resulting sequence contains floor(L/P) patches.Source: A TimesFM for time series embedding · p. 18
• A two-layer MLP residual block independently projects each patch into the model dimension.Source: A TimesFM for time series embedding · p. 18
• Position-encoded patch representations pass through a stacked Transformer to produce one d_ts-dimensional feature for each patch.The output sequence has N patch features.Source: A TimesFM for time series embedding · p. 18

Attribute-Aware Captions Provide Semantically Richer Descriptions

• Attribute-aware captions combine temporal dynamics with the series’ overall trend, periodicity, and noise level, unlike basic captions’ superficial chronological narration.The added attributes provide a more holistic description for downstream reasoning.Source: B Qualitative Analysis of Captioning · p. 18

Visual Representation is Crucial for Capturing Global Temporal Patterns

• Visual plots reveal global periodicity that an LLM reading raw numerical data may miss, despite detecting local oscillations or noise.The authors hypothesize that long text sequences distract the LLM, while plots compress the series into a holistic view that makes global structure salient.Source: B Qualitative Analysis of Captioning · pp. 18–19

• ChronoSteer transfers LLM knowledge into a TSFM for forecasting, while TS-Reasoner transfers TSFM knowledge into an LLM to improve time-series understanding.Source: C Comparison with Related… · p. 19
• Unlike TempoGPT’s discrete quantization tokens and Time-VLM’s image-based inputs, TS-Reasoner directly aligns continuous pretrained-TSFM embeddings with the LLM, avoiding quantization loss and visualization.Source: C Comparison with Related… · p. 19

D.1 Scenarios Where TSFM is Most Beneficial

• Removing the TSFM consistently degrades performance, but by a smaller margin than removing instruction tuning.The paper reports this ablation comparison in Table 6.Source: §D.1 · p. 19
• In finance, removing the TSFM costs 3.58 points at an average input length of 375 steps, versus 2.57 points at 135 steps.The drop grows with input length within the domain.Source: §D.1 · p. 20
• In weather, removing the TSFM costs 1.79 points at an average input length of 336 steps, versus 0.93 points at 168 steps.The drop grows with input length within the domain.Source: §D.1 · p. 20
• The TSFM increasingly benefits longer series because its patch-level representations compactly summarize long-range temporal structure that the LLM struggles to recover from raw values.The MTBench short/long split refers to the question horizon (7-day versus 14-day prediction), not input length; input lengths vary independently across subsets.Source: §D.1 · pp. 20–21
• Removing the TSFM reduces accuracy by 3.07 points on TimeSeriesExam, versus 2.22 points on average on MTBench.TimeSeriesExam questions rely almost entirely on the series, while MTBench questions can also draw on contextual news or weather reports.Source: §D.1 · p. 21
• The TSFM contributes more when reasoning hinges primarily on time-series data and more modestly when complementary textual context is available.TimeSeriesExam depends almost entirely on the series, while MTBench questions can also draw on contextual news or weather reports.Source: §D.1 · p. 21

E Prompts

• Figures 12–15 show the captioning prompts, the TimeSeriesExam prompt, and the prompts for MTBench, respectively.Figure 12 covers captioning, Figure 13 TimeSeriesExam, and Figures 14 and 15 MTBench.Source: E Prompts · p. 22

Data Leakage Examination

• Alignment used synthetic sources (Ansari et al., 2024) and public archives (Merrill et al., 2024), while instruction tuning used Xie et al. (2024) data and was checked for numerical overlap.The numerical-overlap check covered MTBench and TimeSeriesExam.Source: F Data Leakage Examination · p. 22
• Removing the TSFM led to larger drops for longer input series within each domain.“Short” and “long” denote 7-day and 14-day question horizons, respectively.Source: F Data Leakage Examination · p. 24, Table 9 · p. 24
• Generic trend and periodicity captions were kept structurally and semantically distinct from MTBench’s domain-specific analyses and reports and TimeSeriesExam’s property questions.Source: F Data Leakage Examination · p. 24
• 0% of evaluation examples contained a 13-gram appearing in the Stage-2 instruction-tuning corpus.The authors interpret this as evidence that Stage 2 supports genuine reasoning rather than memorization of benchmark formats.Source: F Data Leakage Examination · p. 24, Table 10 · p. 24