TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models
Authors: Sudarshan Regmi, Arvind Pillai, Yu Yvonne Wu, Yuliang Chen, Bibek Panthi, Tess Z. Griffin, Michael V. Heinz, Lisa Marsch, Nicholas C. Jacobson, Andrew Campbell
Timeseries multimodal large language models (TS-MLLMs) have recently begun leveraging the reasoning capabilities of large language models (LLMs) for question-answering tasks. However, these models often fail to capture dynamic temporal patterns, providing only implicit reasoning that lacks the underlying explanations critical for high-stakes applications like healthcare. While reinforcement learning (RL)-based timeseries language models aim to address this, they often fall short because they are trained on narrow, in-distribution data and struggle with out-of-distribution compositional questions. To address these challenges, we present TimeThink, a synthetic framework for eliciting compositional timeseries reasoning. Core timeseries primitives (e.g., trend, seasonality) are domain-independent and can be deterministically generated. Guided by this premise, TimeThink first designs a synthetic data generator that produces atomic and composite question-answer pairs, providing objective ground truth with reasoning traces. Building on this framework, TimeThink employs a reinforcement learning with verifiable rewards (RLVR) training strategy that encourages explicit reasoning. Unlike template-reliant methods, this approach enables the model to learn the underlying logic of composition rather than simply imitating traces. Extensive experiments show that TimeThink, trained only on synthetic data, significantly outperforms strong baselines on both synthetic and real-world benchmarks.
Figures & tables
Figure 1: Overview of TimeThink . (A, B) Synthetic Data Generation: a programmatic generator builds Atomic QA targeting single primitives, Composite QA chaining multiple operators, and out-of-distribution (OOD) QA over unseen task axes. (C) LLM-Driven QA Evolution: templated QAs are evolved into more complex, diverse variants, with the reasoning trace ( <think> block) preserved so that responses stay grounded in timeseries properties. (D) Training Stage Optimization: Stage 1 (SFT) fine-tunes on programmatic reasoning traces at an anchor length of T≤256 ; Stage 2 (RLVR) applies GRPO to incentivize autonomous composition and extend the horizon to T≤768 , using a verifiable reward ( rfmt+rans ) over four answer categories (binary, proximity, categorical, Set- F1 ) besides the format reward.
Figure 2: RL gains come from both the atomic skills ( Δa ) and the sub-skill that composes them ( Δc ).
Figure 3: Overall OOD performance across tasks. TimeThink (RL) leads all native-TS baselines.
Figure 4: Overall performance on unseen-length timeseries ( [769,1536] steps). Δ is TimeThink (RL)’s gain over the best baseline.
Figure 5: Overall performance as the number of TS variables in the prompt grows.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Synthetic
Real-world
Model
Atomic
Comp-D
Comp-R
OOD
SenTSR-Bench
MCQ2
Code agent
83.7
66.8
68.7
72.3
25.8
42.8
TimeThink (RL)
72.8
73.1
73.4
62.9
35.5
45.0
Appendix
Table 4: Comparison with a code-interpreter agent: Qwen2.5-32B-Instruct given the raw values as text plus a Python interpreter. TimeThink reads the series through its encoder, without tools.
Model
Atomic
Comp-D
Comp-R
OOD
SenTSR-Bench
MCQ2
TimeThink (SFT)
0.621
0.693
0.528
0.533
0.264
0.432
+ 10K SFT
0.611
0.681
0.567
0.529
0.362
0.458
+ 10K RL
0.648
0.711
0.597
0.564
0.345
0.490
Appendix
Table 5: Additional SFT vs. RL from the same TimeThink (SFT) checkpoint, both using 10K additional training samples.
Model
Trend
Local
Avg
(#500)
(#193)
(#693)
ChatTS
46.2
23.0
39.7
TimeThink (SFT)
70.0
52.8
65.2
TimeThink (RL)
71.3
53.3
66.3
Appendix
Table 6: Clustering by synchronization criterion (F1 metric).
Model
Description score
Qwen2.5-7B-Inst
0.456
Mistral-7B-Inst-v0.3
0.464
Time-MQA
0.401
Time-R1
0.414
TimeOmni-1
0.476
ChatTS
0.542
Appendix
Table 8: Free-form descriptive task, judge score.
Model
Accuracy
Qwen2.5-7B-Inst
0.343
Mistral-7B-Inst-v0.3
0.270
Time-MQA
0.366
Time-R1
0.305
TimeOmni-1
0.538
ChatTS
0.420
Appendix
Table 10: Full MCQ2 evaluation.
Figure 6: Comparison of responses from ChatTS, TimeThink (SFT) and TimeThink (RL) on the same question.
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
Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitations: they often treat time series merely as text or images, failing to capture the patterns like trends and seasonalities needed to answer specific questions; and when trained on a mix of simple and complex tasks, simpler objectives often dominate the learning process, hindering the development of deep reasoning capabilities. To address these limitations, we propose the Pattern-Aware Alignment and Balanced Reasoning model (PATRA), introducing a pattern-aware mechanism that extracts trend and seasonality patterns from time series to achieve deep alignment. Furthermore, we design a task-aware balanced reward to harmonize learning across tasks of varying difficulty, incentivizing the generation of coherent Chains of Thought. Extensive experiments show that PATRA outperforms strong baselines across diverse Time Series Question Answering (TSQA) tasks, demonstrating superior cross-modal understanding and reasoning capability.
Junkai Lu, Peng Chen, Xingjian Wu +4
East China Normal University, Shanghai, China · Aalborg University, Aalborg, Denmark
Large Language Models (LLMs) have shown promising performance in time series modeling tasks, but do they truly understand time series data? While multiple benchmarks have been proposed to answer this fundamental question, most are manually curated and focus on narrow domains or specific skill sets. To address this limitation, we propose scalable methods for creating comprehensive time series reasoning benchmarks that combine the flexibility of templates with the creativity of LLM agents. We first develop TimeSeriesExam, a multiple-choice benchmark using synthetic time series to evaluate LLMs across five core reasoning categories: pattern recognitionnoise understandingsimilarity analysisanomaly detection, and causality. Then, with TimeSeriesExamAgent, we scale our approach by automatically generating benchmarks from real-world datasets spanning healthcare, finance and weather domains. Through multi-dimensional quality evaluation, we demonstrate that our automatically generated benchmarks achieve diversity comparable to manually curated alternatives. However, our experiments reveal that LLM performance remains limited in both abstract time series reasoning and domain-specific applications, highlighting ongoing challenges in enabling effective time series understanding in these models. TimeSeriesExamAgent is available at https://github.com/magwiazda/TimeSeriesExamAgent.
Malgorzata Gwiazda, Yifu Cai, Mononito Goswami +2
Technical University of Munich, Munich, Germany · Auton Lab, School of Computer Science, Carnegie Mellon University, Pittsburgh, US