TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models

1 Introduction

• Existing TS-MLLMs often learn temporal patterns implicitly, leaving prediction reasoning invisible even though downstream actions depend on the pattern behind an answer.Task-specific reasoning templates can limit transfer to unseen patterns, and models may struggle when questions compose multiple patterns in novel ways.Source: §1 · pp. 1–2
• Prior RL-based approaches make reasoning legible, but rely substantially on task-specific supervised reasoning templates, with RL refining format compliance and offering modest accuracy gains.The resulting models may remain tied to narrow training distributions and struggle with novel compositions.Source: §1 · pp. 1–2
• TIMETHINK combines a synthetic generator of atomic and composite questions with sequential SFT and reinforcement learning with verifiable rewards.Question-aware rewards vary by answer type, providing denser learning signals.Source: §1 · p. 2
• A programmatic generator produces domain-independent timeseries primitives and deterministic, question-aware ground truth, making RLVR tractable at scale.The framework also creates evaluation sets with varying compositions of primitives.Source: §1 · p. 2
• Across six tasks, TIMETHINK outperforms existing timeseries models on compositional questions, and TIMETHINK (RL) generalizes to extended temporal horizons.The evaluation includes diverse synthetic and real-world datasets.Source: §1 · p. 2

2 Related Work

• Timeseries models encode numerical signals as text, visual plots, or native embeddings, but often supervise only final answers, leaving temporal reasoning implicit (Kim et al., 2024; Jin et al., 2023; Xu et al., 2025; Zhang et al., 2025).Text approaches can suffer token exhaustion and degraded long-sequence fidelity (Pillai et al., 2025), while visual prompting may capture surface patterns rather than temporal dynamics.Source: §2 · p. 2
• Explicit-reasoning approaches use annotated chain-of-thought supervision or task-specific reinforcement learning, but are typically trained on narrow distributions and can struggle to compose temporal patterns (He et al., 2026).Deep forecasters have been shown to compose temporal structure under distribution shift, but that composition is implicit and not trained for (Potosnak et al., 2024).Source: §2 · pp. 2–3
• The identified gap is explicit reasoning over core temporal primitives such as trend, correlation, and seasonality, trained with verifiable rewards.The related work describes this as the target of the present work.Source: §2 · p. 3

3.1 Synthetic Data Generation

• Nine atom operators extract temporal facts, while nine sub-skill operators combine them into answers with deterministic ground truth.Composite questions apply atoms to a series or sub-series, then use one or more sub-skills to combine the resulting facts.Source: §3.1 · p. 3
• Composite QA families are split into demonstrated-and-rewarded Comp-D, reward-only Comp-R, and held-out OOD sets to test transfer to unseen compositions.Atomic and composite QAs are mixed in both training stages to reduce overfitting to one chain-of-thought decomposition and avoid primitive-skill regression.Source: §3.1 · pp. 4–5

3.1.1 LLM-driven QA Evolution

• LLM-driven evolution diversifies a small portion of templated QAs while preserving their reasoning structure and verified answers.Question paraphrasing also augments examples while retaining the same reasoning block and final answer.Source: §3.1.1 · p. 5

3.2 Training

• SFT teaches the model to emit a worked reasoning chain followed by the answer, using synthetic demonstrations and a log-likelihood objective.The training framework uses an SFT stage before RLVR.Source: §3.2 · p. 5
• RLVR scores generated responses on answer correctness and required format, letting the model discover reasoning chains beyond its SFT demonstrations.Answer verification covers binary and categorical exact matches, numerical proximity, and Set-F1; GRPO optimizes the reward.Source: §3.2 · pp. 5–6

4 Experiments

• Evaluation covers six task families, including multiple-choice, categorical, numerical, clustering, ranking or enumeration, and description questions.Datasets include TSEvol, MCQ2, SenTSR-Bench, and the TIMETHINK benchmark; metrics include accuracy, proximity accuracy, F1, and judge score.Source: §4 · p. 6
• Baselines include general-purpose LLMs, timeseries and temporal-reasoning models, and the timeseries encoder-based ChatTS approach.The named baselines include Qwen2.5-Instruct-7B, Mistral-Instruct-7B-v0.3, Time-R1-3B, Time-MQA-7B, and TimeOmni-1.Source: §4 · p. 6
• 90K SFT samples and 110K RL samples make up the entirely synthetic training corpus, with no real-world timeseries data included.The compared Qwen3-8B variants use answer-only targets, SFT chain-of-thought supervision, or SFT followed by RLVR.Source: §4 · p. 6

• RL raises Comp-D from 0.696 to 0.731 (+3.5 pp) and Comp-R from 0.532 to 0.733 (+20.1 pp), outperforming all baselines on synthetic compositional QA.Source: §5 · p. 6, Table 1 · p. 7
• Nearly 10 pp: TIMETHINK (RL) improves over SFT from 0.533 to 0.629 on OOD compositional performance, averaged across numerical, categorical, and enumeration tasks.Source: Figure 3 · p. 7, §5 · p. 7
• On real-world TSEvol, TIMETHINK (RL) leads causal and inductive reasoning and average performance, improving the average from 0.632 to 0.709 despite synthetic-only training.Source: Table 2 · p. 7, §5 · p. 7
• With atomic gain essentially constant, composition gains are +0.18 on Comp-R, +0.04 on Comp-D, and +0.09 on OOD, indicating learned composition skill transfers.Source: §5 · p. 7, Figure 2 · p. 7
• Figure 4 evaluates overall performance on unseen-length timeseries spanning [769, 1536] steps and defines Δ as TIMETHINK (RL)’s gain over the best baseline.Source: Figure 4 · p. 8

Length generalization

• On unseen lengths [769, 1536], TIMETHINK (RL) consistently outperforms SFT and ChatTS across categories, with especially large gains on compositional tasks.Source: §6 · p. 8

Data scaling: SFT vs RL

• 110K additional SFT samples improve SenTSR-Bench mean accuracy by 2.8 pp and regress MCQ2 by 8.0 pp; 10K RL samples yield +8.1 pp and +5.7 pp, respectively.Source: Table 3 · p. 8, §6 · p. 8

Sensitivity analysis w.r.t. the number of timeseries

• RL leads across all four variable-count groups, with its margin over SFT growing from approximately 8 pp on UTS to 10 pp for prompts with ≥5 variables.Source: §6 · p. 8, Figure 5 · p. 8

Reward design

• Proximity rewards beat exact-match rewards by +18.4 vs. +8.4 pp on numerical tasks; composite-only rewards leave indirectly rewarded primitives degraded on dependent OOD composites.Source: §6 · p. 8

7 Conclusion

• Domain-invariant primitives such as trend and seasonality provide deterministic ground truth for timeseries reasoning.The paper presents primitive-based reasoning as a promising route for high-stakes domains including healthcare and finance.Source: §7 · p. 8
• TimeThink composes primitives autonomously and outperforms text-based and native timeseries models on synthetic and real-world benchmarks.A curriculum over sequence length carries these gains to longer horizons.Source: §7 · p. 8
• The paper identifies primitive-based reasoning as a promising route for high-stakes domains such as healthcare and finance.Source: §7 · p. 8

8 Limitations

• TimeThink uses a fixed synthetic primitive set that may omit real-world temporal phenomena, and it does not target exhaustive sensor-domain knowledge.The authors identify mixed real-world and synthetic training as a possible direction if precise-ground-truth datasets become available.Source: §8 · p. 9

A.1 Comparison with a Coding Agent

• On SenTSR-Bench, the coding agent scores 25.8 versus 35.5 for TimeThink, while its synthetic advantage makes it an unlike-for-like baseline.The synthetic questions are simple programs over raw values, giving code execution an advantage unavailable to TimeThink; Table 4 compares the agent with TimeThink.Source: §A.1 · p. 11, Table 4 · p. 11

A.2 Additional SFT vs. RL

• Starting from the same SFT checkpoint with 10K additional samples, RL performs best on most metrics, suggesting benefits beyond reward-filtered additional SFT.The SFT comparison uses rejection-fine-tuning on responses selected by the same verifiable reward as GRPO.Source: §A.2 · p. 11

A.3 Clustering Task

• The clustering criteria are trend synchronization, based on matching or opposite trend segments, and local fluctuation synchronization near a target position.Source: §A.3 · p. 11
• SFT improves average clustering F1 over ChatTS by 25.5 points, while RLVR adds 1.1 points.Table 6 reports clustering performance by synchronization criterion using F1.Source: §A.3 · p. 11, Table 6 · p. 11
• The authors attribute the small RLVR gain to this task family being well covered by supervised reasoning traces during SFT.Source: §A.3 · p. 11

A.4 Numerical Task

• TimeThink (RL) beats the best text-only baseline by 23.1 points on position, 17.1 on count, and 5.9 on magnitude, and is strongest in every category.Table 7 reports relative accuracy for each numerical task.Source: §A.4 · p. 11, Table 7 · p. 12
• Relative to TimeThink (SFT), RLVR improves position by 14.9 points, count by 17.2 points, and magnitude by 10.7 points.Numerical tasks use proximity rewards, giving partial credit for predictions close to ground truth.Source: §A.4 · p. 11

A.5 Descriptive Task

• TimeThink (RL) exhibits superior timeseries-description performance under LLM-judge scoring.Table 8 reports the free-form descriptive-task judge score.Source: §A.5 · p. 11, Table 8 · p. 12

A.6 Full Results on SenTSR-Bench

• The section presents the complete SenTSR-Bench results in Table 9.Source: §A.6 · p. 12, Table 9 · p. 12
• TIMETHINK (RL), trained only on synthetic data, almost matches TimeOmni-1 in average performance on SenTSR-Bench.The comparison appears in the full results in Table 9.Source: §A.6 · p. 12
• Table 9 presents full SenTSR-Bench accuracy across the WHAT HAPPENED, HOW HAPPENED, and SUGGESTED FIX splits.The supplied passage identifies the table's split structure but gives no cell values.Source: Table 9 · p. 12

A.7 Full Results on MCQ2

• TIMETHINK (RL) outperforms all MCQ2 baselines except TimeOmni-1.The supplied passages do not provide the table's individual scores.Source: §A.7 · p. 12, Table 10 · p. 12

A.8 Reward Design Case Studies

• RL gains +18.4 pp on the proximity-rewarded periodicity task versus +8.4 pp on the exact-match task, while count-only supervision of trend segments leaves TIMETHINK (RL) 18.9 pp and 13.4 pp below SFT on two OOD composites.These composites require precise trend-segment positions; the authors conclude atomic primitives should be rewarded directly because count-only RL can degrade the primitive and composites built on it.Source: §A.8 · p. 12

B Qualitative Example

• Figure 6 compares ChatTS, TIMETHINK (SFT), and TIMETHINK (RL) responses to the same question.The supplied passages do not state the responses' comparative outcomes.Source: B Qualitative Example · p. 13, Figure 6 · p. 13

C LLM-as-Judge Prompt Templates

• Evaluations use task-family-specific LLM judge templates, including free-form, multiple-choice, structured-verdict, and description-extraction judges; the multiple-choice judge separates verdict correctness from reasoning quality.The structured-verdict judge also extracts numeric answers and provides graded correctness and explanation signals.Source: C LLM-as-Judge Prompt Templates · pp. 14–15

D.1 Atomic QA

• Atomic QA applies one of nine primitives directly to a timeseries, instantiated across 14 single-metric question variants.Primitives include summary statistics, extrema, event and trend-segment enumeration, periodicity, and trend classification.Source: §D.1 · p. 15

D.2 Sub-skills

• The nine sub-skills operate on atomic outputs or earlier sub-skills, including counting, filtering, locating, ranking, and arithmetic.Other listed operations include summation, comparison, thresholding, and argmax/argmin.Source: §D.2 · pp. 15–16

D.3 Comp-D compositions (36)

• Comp-D compositions combine atomic operators and sub-skills into questions spanning metric comparisons, trend/event analysis, periodicity, and multi-condition judgments.Examples include comparing half-series means and testing whether an instability gap or anticorrelated high-noise state holds.Source: §D.3 · p. 16
• Comp-D recipes compose statistics, event and segment operations, and trend aggregation, including seasonal-amplitude thresholds, longest-segment classification, and the share of decreasing points.Source: §D.3 · p. 17

D.4 Comp-R compositions (32)

• Comp-R recipes span event and segment enumeration, localized segment queries, distribution comparisons, and cross-metric aggregation of means and standard deviations.Source: §D.4 · pp. 17–18

D.5 OOD compositions (34)

• OOD recipes combine quarter and chunk statistics with event–segment reasoning and cross-metric relations, including whether maxima occur at similar positions or events precede one another.Source: §D.5 · pp. 18–19

E Compositional Performance Breakdown

• Comp-D gains are small on average because SFT chain-of-thought supervision already teaches the decomposition; Comp-R gains are large where RL learns compositions not demonstrated during SFT, while OOD compositions test transfer to shapes held out throughout training.The breakdown reports TIMETHINK (SFT) and TIMETHINK (RL) scores on constituent atoms (∆a) and compositions (∆c).Source: E Compositional Performance Breakdown · p. 19, Table 11 · p. 20, Table 12 · p. 21