TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models

Conference: ICLR 2026
arXiv: 2509.24803
Code: GitHub
Area: Human Understanding/Time Series
Keywords: Time Series Reasoning, LLM, Reinforcement Learning, Multi-task Joint Training, Causal Discovery

TL;DR

TimeOmni-1 proposes the first unified time series reasoning model. Through TSR-Suite (the first reasoning-oriented time series dataset suite) and a two-stage training process (SFT for injecting time series priors + RL for refining reasoning), it significantly outperforms GPT-4.1 across multiple time series reasoning tasks.

Background & Motivation

Time series understanding is shifting from basic pattern analysis to advanced reasoning, yet two major bottlenecks persist:

Scarcity of High-Quality Data: Existing time series QA datasets (e.g., Time-MQA) remain at a superficial question-answering level. They suffer from: (a) Overly simple questions where the gap between reasoning and non-reasoning models is minimal; (b) Insufficient context, where missing key information forces models to guess rather than reason.

Lack of Viable Reasoning Paths: It remains unclear which tasks truly necessitate time series reasoning capabilities. Existing methods are restricted to narrow tasks (e.g., TimeMaster training six models for six datasets), lacking cross-task transferability.

The authors propose two core Design Principles: - Principle 1 — QA Must Reward Reasoning: Reasoning models should significantly outperform non-reasoning models \(\bar{S}(M_{RM}) \gg \bar{S}(M_{NRM})\). - Principle 2 — Context Must Be Sufficient: Sufficient time series input \(X\) and auxiliary context \(C\) must be provided to avoid ambiguity.

Method

Overall Architecture

TimeOmni-1 decomposes the objective of "enabling LLMs to truly reason with time series" into two complementary phases. First, using TSR-Suite, three types of temporal capabilities—perception, extrapolation, and decision-making—are fed into the model in the form of human-guided Chain-of-Thought (CoT). These reasoning chains are produced via a hierarchical annotation workflow: "LLM drafting → Human expert verification → LLM normalization." Second, a two-stage curriculum learning process solidifies these priors: Stage 1 utilizes Supervised Fine-Tuning (SFT) to teach the model "how to think," while Stage 2 employs GRPO with task-customized rewards to make the model "think more accurately." Throughout this process, four tasks are integrated into a single model for joint training, ultimately converging into the unified TimeOmni-1.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
    D0["Raw Time Series Data<br/>10 Domains · 23K+ Samples"]
    subgraph DATA["TSR-Suite Dataset + Hierarchical CoT Annotation"]
        direction TB
        A["LLM Analyzer<br/>Draft reasoning chains under human templates"] --> H["Human Experts<br/>Verify context sufficiency + Write expert chains"]
        H --> R["LLM Rewriter<br/>Uniform format"]
        R --> T["4 Atomic Tasks<br/>Perception→Extrapolation→Decision"]
    end
    D0 --> DATA
    DATA --> S1["Stage 1: SFT<br/>Injecting TS reasoning priors"]
    S1 --> S2["Stage 2: GRPO<br/>Task-customized rewards"]
    S2 --> M["Unified Model TimeOmni-1<br/>Multi-task joint training"]

Key Designs

1. TSR-Suite: Embedding "Reasoning Necessity" and "Context Sufficiency" into the Dataset

Existing time series QA datasets often have shallow questions and sparse contexts, offering reasoning models little advantage over non-reasoning ones and forcing models to rely on guessing. TSR-Suite directly embeds two design principles into the data: first, reasoning models must significantly outperform non-reasoning models \(\bar{S}(M_{RM}) \gg \bar{S}(M_{NRM})\); second, sufficient time series input \(X\) and auxiliary context \(C\) are provided to eliminate ambiguity. Based on these, it constructs 4 atomic tasks covering three layers of capability: Perception (Task 1: Scene Understanding/Single-sequence Attribution, Task 2: Causal Discovery/Multi-sequence Causal Relationships), Extrapolation (Task 3: Event-aware Prediction/Inferring future trends under event perturbations), and Decision-making (Task 4: Decision Formulation/Integrating previous capabilities for action selection). The suite contains 23K+ samples across 10 domains, with 2.3K meticulously curated via human guidance. This pipeline encodes the cognitive logic of "understand first, then predict, then act," carving a genuine reasoning path for the model.

2. Hierarchical CoT Annotation Workflow: Reducing Noise and Cost via Human-AI Collaboration

High-quality reasoning chains cannot rely solely on humans (too expensive) or LLMs (prone to error). TimeOmni-1 addresses this gap with a three-step division of labor: the LLM Analyzer generates initial versions (Step-1 CoT) under human-guided templates; human experts then verify context sufficiency and write expert chains (Step-2 CoT) specifically for cases where the LLM failed; finally, the LLM Rewriter normalizes these into a uniform format. Human-guided templates are critical—GPT-4.1's zero-shot causal discovery improved from 28.7% to 71.1% when using these templates, demonstrating that templates provide the necessary reasoning scaffolding rather than direct answers.

3. Task-Customized RL Rewards: Differentiated Scoring for Discrete and Sequential Tasks

Output formats vary significantly across the four tasks, meaning a single reward function would cause distortion. Therefore, the Stage 2 GRPO splits scoring by task type. All tasks undergo a format reward \(\mathcal{R}_{format}\), enforcing the <think></think><answer></answer> structure. Discrete tasks (Tasks 1, 2, 4) use an accuracy reward based on exact matching \(\mathcal{R}_{discrete} \in \{0,1\}\). Sequence prediction (Task 3) is split into two parts: a count reward \(\mathcal{R}_{count}=0.1\) for correct sequence length, plus a normalized reward derived from MAE via exponential decay, separately incentivizing length accuracy and numerical precision.

4. Multi-task Joint Training: Mutual Benefit Across Three Capability Layers

Prior narrow-task methods (e.g., TimeMaster training 6 models for 6 datasets) fail to transfer across tasks. TimeOmni-1 incorporates all tasks into a single model and demonstrates positive transfer through two progressive experiments: Progressive Capability Transfer—even without direct training on the decision task, training on perception and extrapolation alone raised decision accuracy from 25.5% to 31.3%; and Progressive Capability Supplementation—gradually adding prerequisite tasks to joint training further improved decision accuracy from 40.9% to 47.9%. Together, these support the paradigm of "train once, reuse across tasks."

Loss & Training

Stage 1 uses standard SFT cross-entropy loss on human-guided CoT data to inject time series reasoning priors. Stage 2 employs GRPO (Group Relative Policy Optimization) with rewards combined by task type: for discrete tasks, \(R = \mathcal{R}_{format} + \mathcal{R}_{discrete}\); for sequence prediction, \(R = \mathcal{R}_{format} + \mathcal{R}_{count} + \text{exp-decay}(\text{MAE})\).

Key Experimental Results

Main Results

ID/OOD Tests Across Four Tasks (ACC %, MAE↓ for Task 3):

Method	Scene Understanding (ID)	Causal Discovery (ID)	Event Prediction (ID/MAE)	Decision (ID)
GPT-4.1	85.5	28.7	13.79	25.5
Qwen2.5-7B	48.5	21.6	23.28	25.5
Time-R1	30.9	30.2	17.61	27.8
TimeOmni-1	90.7	69.3	14.30	47.9

TimeOmni-1 outperforms GPT-4.1 by 40.6% in causal discovery (ID) and by 22.4% in decision tasks.

Ablation Study

Configuration	Causal Discovery (ID)	Decision (ID)	Description
Base model	21.6	25.5	LLMs lack temporal priors
ANS-SFT (Answer Supervised)	30.5	51.0	Fits answer distribution only, no reasoning
CoT-SFT (Stage 1)	67.7	40.9	CoT injection significantly boosts causal discovery
CoT-SFT+RL (Stage 2)	69.3	47.9	RL refinement provides further gains
Single-task (CoT-SFT+RL)	67.5	40.9	Joint training outperforms single-task

Key Findings

• LLMs Inherently Lack Time Series Reasoning Priors: Base models achieved only 21.6% in causal discovery (near the 33.3% random baseline); RL alone cannot establish this capability.
• Human-Guided Templates are Crucial: GPT-4.1 zero-shot causal discovery rose from 28.7% to 71.1% when using human-guided templates.
• Mutual Benefits from Joint Training: Cross-task joint training outperformed single-task training across all categories, supporting the "train once, use across tasks" paradigm.
• General Reasoning Capabilities are Preserved: TimeOmni-1's average accuracy on general reasoning benchmarks (DROP, GPQA, ReClor) increased by 16.5% compared to the base model.
• SR (Successful Response Rate): TimeOmni-1 achieved SR \(\geq\) 93.8% across all tasks, far surpassing current specialized models (e.g., ChatTS had SR=0% on event prediction).

Highlights & Insights

1. Systematic Definition of Time Series Reasoning Tasks: First to explicitly propose "reasoning necessity" and "context sufficiency" as design principles, building a task system that genuinely requires reasoning.
2. Progressive Path: Perception → Extrapolation → Decision: Reflects the cognitive logic of "understanding before predicting before acting," providing depth in task design.
3. Complementary Relationship in Two-Stage Training: SFT handles "knowing how to think," while RL handles "thinking more accurately"—both are indispensable.
4. Empirical Evidence of Cross-Task Positive Transfer: Carefully designed progressive experiments prove the intrinsic links between the three core temporal reasoning capabilities.

Limitations & Future Work

1. Limited Data Scale: TSR-Suite contains only 23K samples (2.3K human-annotated), which is small compared to general NLP datasets.
2. Concentrated Task Types: The 4 tasks focus on classification and prediction, lacking diverse reasoning tasks like anomaly detection or trend interpretation.
3. OOD Generalization Gap: Event prediction OOD MAE is 145.53 (vs 14.30 ID), indicating cross-domain generalization needs improvement.
4. Base Model Constraints: Validated only on Qwen2.5-7B; scaling behavior of larger models remains unexplored.
5. Human-Template Dependency: Data construction relies heavily on human-guided templates, limiting scalability.

Related Work & Insights

• Time-R1 is the closest reasoning model but is limited to classical forecasting; TimeOmni-1 extends this to multi-task reasoning.
• DeepSeek-R1 proved that RL can enhance reasoning; TimeOmni-1 brings this paradigm to the time series domain.
• Time-MQA is large but suffers from simple tasks and insufficient context; TSR-Suite provides targeted improvements.
• Key insight for Time Series Intelligence: General temporal models require the injection of reasoning priors rather than just pattern fitting.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First systematic construction of a TS reasoning task system + first unified TS reasoning model.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across four tasks (ID/OOD), progressive experiments, ablations, and general capability assessments.
• Writing Quality: ⭐⭐⭐⭐ Clear structure driven by design principles, though some charts have high information density.
• Value: ⭐⭐⭐⭐⭐ Opens a new direction for time series reasoning; data, models, and code are all open-sourced.

Related Papers

• [ICLR 2026] Semantic-Enhanced Time-Series Forecasting via Large Language Models
• [ICLR 2026] Multi-Scale Hypergraph Meets LLMs: Aligning Large Language Models for Time Series Analysis
• [ICLR 2026] Reasoning on Time-Series for Financial Technical Analysis
• [ICML 2026] Building Social World Models with Large Language Models
• [ICLR 2026] TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale