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, leveraging TSR-Suite (the first reasoning-oriented time series dataset suite) and a two-stage training paradigm (SFT for injecting temporal priors + RL for refining reasoning), achieving significant improvements over GPT-4.1 across multiple time series reasoning tasks.

Background & Motivation

Time series understanding is transitioning 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, suffering from critical issues — (a) questions are overly simplistic, with negligible performance gaps between reasoning and non-reasoning models; (b) insufficient context forces models to guess rather than reason.

Lack of Viable Reasoning Pathways: It remains unclear which tasks genuinely require time series reasoning capabilities. Existing methods are confined to narrow tasks (e.g., TimeMaster trains six models on six datasets separately) without cross-task transferability.

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

Method

Overall Architecture

TimeOmni-1 adopts a two-stage curriculum learning framework:

• Stage 1 (SFT): Injects time series reasoning priors — supervised fine-tuning with human-guided chain-of-thought (CoT), establishing three core temporal reasoning capabilities in the LLM: perception, extrapolation, and decision-making.
• Stage 2 (RL): Refines reasoning — employs GRPO with task-customized reward functions to convert imitative priors into robust reasoning behavior.

Key Designs

1. TSR-Suite (Time Series Reasoning Suite): The first comprehensive time series reasoning dataset, covering 4 atomic tasks and 3 major reasoning capabilities:

   • Perception: Task 1 — Scene Understanding (attribution of a single series); Task 2 — Causal Discovery (causal relationships among multiple series)
   • Extrapolation: Task 3 — Event-Aware Forecasting (inferring future trends under event perturbations)
   • Decision-making: Task 4 — Decision Making (integrating perception and extrapolation for action selection)

The suite contains 23K+ samples, of which 2.3K are carefully curated through a human-guided hierarchical annotation pipeline, spanning 10 domains.

1. Hierarchical CoT Annotation Pipeline: A three-step annotation process — (a) an LLM Analyzer generates reasoning chains under human-guided templates (Step-1 CoT); (b) human experts verify contextual sufficiency and author expert reasoning chains for LLM error cases (Step-2 CoT); (c) an LLM Rewriter normalizes the expert reasoning chains.

2. Task-Customized RL Reward Design:

   • Format reward \(\mathcal{R}_{format}\): enforces the <think></think><answer></answer> structure
   • Discrete tasks (Tasks 1, 2, 4): exact-match accuracy \(\mathcal{R}_{discrete} \in \{0,1\}\)
   • Sequential task (Task 3): count reward \(\mathcal{R}_{count}=0.1\) (for correct prediction sequence length) + exponential-decay-mapped normalized MAE reward

3. Multi-task Joint Training: All tasks are trained within a single unified model. Two complementary experiments validate cross-task gains:

   • Progressive Capability Transfer: Without directly training on the decision-making task, training on perception + extrapolation alone raises decision accuracy from 25.5% to 31.3%.
   • Progressive Capability Supplementation: Incrementally incorporating prerequisite tasks in joint training raises decision accuracy from 40.9% to 47.9%.

Loss & Training

• Stage 1: Standard SFT loss (cross-entropy) using human-guided CoT data.
• Stage 2: GRPO (Group Relative Policy Optimization) with reward functions:
  • Tasks 1/2/4: \(R = \mathcal{R}_{format} + \mathcal{R}_{discrete}\)
  • Task 3: \(R = \mathcal{R}_{format} + \mathcal{R}_{count} + \text{exp-decay}(\text{MAE})\)

Key Experimental Results

Main Results

ID/OOD evaluation across four tasks (ACC %, Task 3 reports MAE↓):

Method	Scene Understanding (ID)	Causal Discovery (ID)	Event Forecasting (ID/MAE)	Decision-Making (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 surpasses GPT-4.1 by 40.6% on causal discovery (ID) and by 22.4% on the decision-making task.

Ablation Study

Configuration	Causal Discovery (ID)	Decision-Making (ID)	Note
Base model	21.6	25.5	LLM lacks temporal priors
ANS-SFT (answer supervision)	30.5	51.0	Fits answer distribution only, no reasoning
CoT-SFT (Stage 1)	67.7	40.9	CoT injection significantly improves causal discovery
CoT-SFT+RL (Stage 2)	69.3	47.9	RL refinement yields further gains
Single-task training (CoT-SFT+RL)	67.5	40.9	Joint training outperforms single-task

Key Findings

• LLMs Inherently Lack Temporal Reasoning Priors: The base model achieves only 21.6% on causal discovery (near random chance at 33.3%), and RL alone cannot establish this capability.
• Human-Guided Templates Are Critical: GPT-4.1 achieves 28.7% on zero-shot causal discovery, rising to 71.1% with human-guided templates.
• Joint Training Yields Mutual Benefits: Cross-task joint training outperforms single-task training on all tasks, supporting a "train once, transfer across tasks" paradigm.
• General Reasoning Capabilities Are Preserved: TimeOmni-1 achieves an average accuracy improvement of 16.5% over the base model on general reasoning benchmarks including DROP, GPQA, and ReClor.
• Success Rate (SR): TimeOmni-1 achieves SR ≥ 93.8% across all tasks, substantially outperforming existing time series-specific models (e.g., ChatTS achieves SR = 0% on event forecasting).

Highlights & Insights

1. Systematic Definition of Time Series Reasoning Tasks: The first work to explicitly articulate "reasoning necessity" and "context sufficiency" as design principles, establishing a task framework that genuinely demands reasoning.
2. Perception → Extrapolation → Decision-Making as a Progressive Capability Pathway: Reflects the cognitive logic of "understand first, then predict, then act," demonstrating depth in task design.
3. Complementary Roles of the Two-Stage Training: SFT instills "knowing how to think," while RL refines "thinking more accurately" — both stages are indispensable.
4. Empirical Evidence for Cross-Task Positive Transfer: Through carefully designed progressive experiments, the work demonstrates that the three core temporal reasoning capabilities are intrinsically interconnected.

Limitations & Future Work

1. Limited Data Scale: TSR-Suite contains only 23K samples (with 2.3K manually annotated), which is small compared to general NLP datasets.
2. Narrow Task Coverage: The 4 tasks focus predominantly on classification and forecasting, lacking more diverse reasoning tasks such as anomaly detection and trend interpretation.
3. Persistent OOD Generalization Gap: Event forecasting OOD MAE reaches 145.53 (vs. 14.30 in-domain), indicating that cross-domain generalization requires further improvement.
4. Base Model Constraints: Validation is limited to Qwen2.5-7B; scaling behavior with larger models remains unexplored.
5. Reasoning Chain Quality Depends on Human Templates: Data construction relies heavily on human-guided templates, limiting scalability.

Related Work & Insights

• Time-R1 is the closest existing time series reasoning model but is limited to classical forecasting; TimeOmni-1 extends the scope to multi-task reasoning.
• DeepSeek-R1 demonstrates that RL can enhance reasoning capabilities; TimeOmni-1 introduces this paradigm to the time series domain.
• Time-MQA is large in scale but suffers from overly simple tasks and insufficient context; TSR-Suite is designed to address these shortcomings directly.
• The work offers important insights for the field of time series intelligence: general-purpose temporal models require the injection of reasoning priors rather than relying solely on pattern fitting.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First systematic construction of a time series reasoning task framework + first unified time series reasoning model
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four-task ID/OOD evaluation + progressive experiments + ablation study + general capability assessment — highly comprehensive
• Writing Quality: ⭐⭐⭐⭐ Clear structure with design-principle-driven exposition; some figures and tables are overly dense in information
• Value: ⭐⭐⭐⭐⭐ Opens a new direction for time series reasoning; dataset, model, and code are fully open-sourced

Related Papers

• [ICLR 2026] Reasoning on Time-Series for Financial Technical Analysis
• [ACL 2026] Learning Uncertainty from Sequential Internal Dispersion in Large Language Models
• [NeurIPS 2025] PlanU: Large Language Model Reasoning through Planning under Uncertainty
• [ICLR 2026] VoT: Event-Driven Reasoning and Multi-Level Alignment Unlock the Value of Text for Time Series Forecasting
• [ICLR 2026] EDINET-Bench: Evaluating LLMs on Complex Financial Tasks using Japanese Financial Statements