Organizations: University of Auckland · Shandong University · Northeast Normal University · University of Macau · The Hong Kong University of Science and Technology (Guangzhou)
Large language models (LLMs) have recently shown strong potential for time series reasoning, while models that reason over time series are usually with long chains of thought. However, this process often consumes thousands of tokens, directly leading to substantial latency during inference. In contrast, System 1 models such as Jev make decisions instantly and can naturally be an alternative, but raw time series are not compatible with the embedding space of these models, thus hindering the reasoning performance, especially when input series contain thousands of numeric points. In this paper, we make a very early attempt to propose a time series System 1 model, namely TimesJev. Specifically, we first formalize System 1 sufficiency in time series reasoning, which motivates a compressed representation for time series with key frequency patterns preserved. To construct this representation, we derive a dimension reduction objective that combines (i) a sufficiency divergence that measures the core information lost during compression, and (ii) a realizability divergence that quantifies the incompatibility of the System 1 model to the optimal prediction given this representation. We optimize this objective by learning time series compression with a gating mechanism for sufficiency, and then aligning the selected representation with the embedding space for realizability. Finally, experiments on 13 datasets demonstrate that TimesJev outperforms reasoning models by over 62% with 62.9% less inference time than their fine-tuned versions on average.
Figures & tables
Figure 1: Accuracy versus inference latency on time series tasks, with bubble size denoting generated tokens.
Figure 2: Overview of TimesJev. (a) Inference with a System 1 model in one forward pass. (b) Compressing the time series through stochastic gates to reduce the sufficiency divergence. (c) Aligning the compact representation with the model through learned embeddings to reduce the realizability divergence.
Method
Anomaly Detection
Classification
MCQA
True/False QA
Overall
Environment
AIOps
Healthcare
Web
Human Activity
Healthcare
Nature
Web
Other
Healthcare
Nature
Web
Other
General-Purpose LLMs
Qwen3.5-9B
Zero-shot
49.60
63.00
44.80
58.20
22.80
61.96
77.50
83.61
55.81
71.68
74.59
80.95
79.31
54.31
LoRA SFT
79.20
98.00
99.20
91.80
76.80
71.74
86.67
86.89
72.09
84.27
88.52
87.30
89.66
86.91
Qwen3-8B
Zero-shot
50.00
50.40
54.00
59.80
21.20
63.04
66.67
90.16
53.49
70.98
80.33
88.89
72.41
53.91
LoRA SFT
69.20
97.20
99.00
85.00
78.40
68.84
78.33
86.89
67.44
77.27
88.52
88.89
89.66
83.51
Table 1: Main results on TSQA in accuracy (%). LoRA SFT denotes LoRA fine-tuning on the TSQA training set. The best result is in bold and the second best is underlined . † Official checkpoints fine-tuned on TSQA.
Method
Tokens
TFLOPs
Latency (ms)
Overall ( ↑ )
General-Purpose LLMs
Qwen3.5-9B
Zero-shot
1535
35.97
12528
54.31
LoRA SFT
7
6.20
102
86.91
Qwen3-8B
Zero-shot
577
14.79
4835
53.91
LoRA SFT
7
5.22
103
83.51
Time Series Reasoning Models
Table 2: Test-time cost and accuracy on TSQA. ≈ 0 indicates that the answer is read from option probabilities without decoding. Overall is the accuracy (%) in Table 1 . † Models fine-tuned on TSQA.
Figure 3: Accuracy by domain when the feature values are kept, the most important one is permuted, or all are shuffled across questions.
Figure 4: Top : t-SNE embeddings of the raw series and their compact representations on Web anomaly detection. Bottom : sufficient summary plots along the first sliced inverse regression (SIR) direction ( Li, 1991 ) .
Figure 5: Confusion matrices on classification for Qwen3-8B and TimesJev. Each row shows the share of a true activity assigned to each predicted activity.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Dataset
Opt.
Len.
M
Train
Test
Time
Anomaly Detection
4 domains
2
8–64
69
13,787
7,000
87.8
Classification
HAR
6
30
240
4,000
2,000
28.7
MCQA
Mixed
2–4
4–48
69
4,000
2,000
13.1
True/False QA
Mixed
2
4–48
69
3,703
2,000
10.5
Appendix
Table 3: Datasets and training statistics. All datasets come from TSQA; Mixed covers its five domains. Anomaly detection trains one model per domain, so its counts and time are summed over the four domains. Opt. is the number of options. Len. is the number of values in a question’s series. M is the size of the feature library. Time is the training time in minutes.
Setting
Value
Backbone
Decider-2B, frozen
Trainable module
Value-token MLP (0.54M)
Features
K=32 , stochastic gates
Gates
σ=0.5 , λ∈{0.003,0.01,0.03,0.1}
Optimizer
AdamW, lr 10−3 , weight decay 0
Schedule
5% warm-up, cosine decay, 3 epochs
Appendix
Table 4: Training settings, shared by all datasets.
Figure 6: Accuracy of TimesJev with the Decider-2B versus the number of selected features K .
Figure 7: Comparison of stochastic gates with other dimension reduction methods.
Figure 8: Share of the features selected by stochastic gates that come from each family of the library.
Variant
Anomaly Detection
Classification
MCQA
True/False QA
Overall
Environment
AIOps
Healthcare
Web
Human Activity
Healthcare
Nature
Web
Other
Healthcare
Nature
Web
Other
Ours (Full)
84.20
99.00
99.60
97.60
80.00
71.74
84.16
91.80
74.42
81.81
82.79
90.98
93.10
87.02
Compression
w/o STG selection
82.00
98.60
94.60
87.40
74.40
57.97
55.83
73.77
53.49
61.89
57.38
61.90
58.62
70.60
w/o subset constraint
82.20
98.80
96.60
92.40
71.20
50.00
50.00
62.30
60.47
67.13
82.25
90.48
79.31
75.63
w/o compression
73.60
98.60
99.40
78.80
71.80
57.97
65.83
85.25
72.09
72.03
80.70
88.89
89.66
79.59
Appendix
Table 5: Ablation study of TimesJev with Decider-2B on TSQA in accuracy (%). Bold denotes the best results.
Figure 9: Accuracy drop when each of the top two features is permuted across questions.
Figure 10: Case study on human activity classification.
Figure 11: Case study on anomaly detection in the environment domain.
Figure 12: Case study on anomaly detection in the web domain.
Large language models (LLMs) have demonstrated strong reasoning and instruction-following capabilities, making them potentially powerful tools for time-series analysis. However, time series lie outside their native textual modality, raising a fundamental question: how should time series be represented so that LLMs can reason about them effectively? Existing work typically serializes raw numerical sequences or fine-tunes pre-trained LLMs on time-series data. These approaches place the burden of extracting temporal structure directly on the LLM, creating a modality mismatch that often degrades performance on long sequences and introduces substantial computational overhead. In this work, we introduce Time-Series-to-Structured-Program representation (T2SP), a deterministic, training-free method that represents a time series as a structured symbolic program. T2SP decomposes time series into trends, periods, and salient events, expressing them in a program-friendly format aligned with the textual and code-like modalities on which LLMs are natively trained. By shifting temporal-structure extraction from the model to the representation itself, T2SP enables off-the-shelf LLMs to leverage their existing reasoning capabilities for time-series understanding. We evaluate T2SP on three reasoning tasks -- editing, captioning, and question answering -- where it consistently improves performance, reduces reasoning time, and lowers failure rates compared with raw-string representations. Our results demonstrate that T2SP provides an effective interface between time series and LLMs.
Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of whether or not the entire sequence is relevant. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection. These results show that selective data use drives effective time-series reasoning.
Shvat Messica, Jiawen Zhang, Kevin Li +2
Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA · The Hong Kong University of Science and Technology (Guangzhou) · Massachusetts Institute of Technology +1
Time series analysis has recently been coupled with Large Language Models (LLMs) to leverage their reasoning and world knowledge capabilities, yet gains remain limited. We attribute this to a fundamental mismatch between existing task formulations and LLM strengths: most settings reduce time series understanding to curve-fitting systems, focusing on low-level prediction while ignoring the semantic, contextual, and reasoning-intensive nature of real-world temporal decision-making.To address these limitations, we introduce TSCognition, a multimodal benchmark for multi-dimensional time series reasoning. It collects real-world time series and textual information from 15 public sources and constructs approximately 41K QA samples around five cognitive reasoning tasks: Decoding, Grounding, Inferring, Extrapolating, and Acting. Building on this, we further propose TSAlign, a unified framework that encodes time series into compact patch-level representations and aligns them with semantic directions in the LLM embedding space via gated residual injection and multivariate fusion.Experiments show that TSAlign outperforms existing LLM, VLM, and time series QA baselines on TSCognition and the publicly available TimerBed benchmark while substantially reducing computational cost.Code is available at: https://github.com/EIT-NLP/CognitiveTSR
Xin Qiu, Junlong Tong, Yao Zhang +3
Institute of Digital Twin, Eastern Institute of Technology, Ningbo · Zhejiang University · Munich Center for Machine Learning, LMU