TimeOmni-1: Unified Multimodal Time Series Reasoning

• TimeOmni-1 is a unified reasoning model that decomposes time series analysis into four atomic tasks across perception, extrapolation, and decision-making.
• It integrates a lightweight multimodal front end with a pretrained LLM, leveraging chain-of-thought supervision and reinforcement learning with task-grounded rewards.
• The framework supports robust evaluations in both in-distribution and out-of-distribution settings, outperforming models like GPT-4.1 in key tasks.

TimeOmni-1 is a unified reasoning model for multimodal time series understanding, introduced together with the Time Series Reasoning Suite (TSR-Suite) to formalize and operationalize time series reasoning in LLMs (Guan et al., 29 Sep 2025). In the formulation associated with the model, time series reasoning is decomposed into four atomic tasks spanning three fundamental capabilities: perception, extrapolation, and decision-making. The framework couples a lightweight multimodal front end to a pretrained base LLM, uses chain-of-thought supervision and reinforcement learning with task-grounded rewards, and is evaluated under both in-distribution and out-of-distribution settings. The stated motivation is that existing multimodal time series datasets mostly remain at the level of surface alignment and question answering, whereas TSR-Suite and TimeOmni-1 are designed around tasks that genuinely require time series reasoning (Guan et al., 29 Sep 2025).

1. Conceptual scope and problem setting

TSR-Suite defines four atomic tasks that together span three fundamental reasoning capabilities over time series: perception through scenario understanding and causality discovery, extrapolation through event-aware forecasting, and decision-making through deliberation over perception and extrapolation (Guan et al., 29 Sep 2025). The underlying problem setting is expressed with observed series $X=\{x^{(m)}_{1:T}\}_{m=1}^{M}$, auxiliary textual context $C$, a reasoning trace $R$, and final answer $y$, under the schema

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$

This formulation places reasoning traces inside the model’s output space rather than treating them as auxiliary annotations only. A plausible implication is that the framework is intended not merely to classify or regress from time series, but to model an explicit reasoning process over numerical observations and textual context.

The four tasks are specified as follows.

Capability	Task	Formal output
Perception	Scenario Understanding	$y\in\{1,\dots,K\}$
Perception	Causality Discovery	$$y\in\{\text{“1→2”},\text{“2→1”},\text{“no-causal”}\}$$
Extrapolation	Event-aware Forecasting	$\hat{x}_{t+1:t+H}$
Decision-making	Deliberation	$a^*=\arg\max_{a\in A}U(a\mid X,C)$

For scenario understanding, the model receives a single series $x_{1:T}$ and candidate labels $C$0, and selects

$C$1

For causality discovery, the task is framed as binary classification or directional prediction over two or more series, with the model inferring $C$2 if $C$3 causes $C$4, or $C$5 if there is no direct causal link. For event-aware forecasting, the model predicts the next $C$6 steps from history $C$7 and an event sequence $C$8 described in $C$9:

$R$0

For deliberation, the task combines series, textual constraints, and a discrete action space $R$1 to select

$R$2

where $R$3 is an expected utility such as profit derived via internal forecasting and causal reasoning (Guan et al., 29 Sep 2025).

A common misconception is that TimeOmni-1 is primarily a forecasting model. The task design indicates otherwise: forecasting is only one of four atomic tasks, and the suite explicitly treats perception and decision-making as coequal targets of training and evaluation (Guan et al., 29 Sep 2025).

2. TSR-Suite dataset and annotation pipeline

TSR-Suite is described as the first comprehensive time series reasoning suite that supports not only thorough evaluation but also the data pipeline and training of time series reasoning models (Guan et al., 29 Sep 2025). It contains more than 23K samples; the detailed overview gives the more specific figure of 23 605 total QA pairs assembled across 10 domains, with 2.3 K high-quality, human-guided chains and 23 K overall samples for training and evaluation (Guan et al., 29 Sep 2025).

The data curation process is hierarchical. Raw-data assembly across 10 domains yields 23 605 total QA pairs. In Step 1, an LLM Analyzer plus structured templates produces approximately 1 000 “easy” chain-of-thought seeds. In Step 2, human review of hard cases adds approximately 1 300 expert-verified chain-of-thought traces. The result is a 2.3 K human-verified subset, augmented with the remaining 21 K auto-labeled pairs for model development (Guan et al., 29 Sep 2025).

The design of the suite is significant because it ties task formalization, annotation, training, and evaluation into a single framework. This suggests that TSR-Suite is not only a benchmark but also an end-to-end data substrate for supervised and reinforcement learning over time series reasoning traces. The emphasis on “human-guided hierarchical annotation” indicates that the curated subset is intended to provide higher-fidelity reasoning supervision than purely synthetic chain-of-thought generation.

Another common misunderstanding is that the benchmark measures only answer accuracy. In fact, the framework also tracks whether a response is structurally valid. Evaluation first filters for valid responses, defined through Success Rate (SR), the fraction of outputs from which a well-formed <answer> can be extracted; only then are task-specific metrics computed on valid cases (Guan et al., 29 Sep 2025).

3. Model architecture and input–output interface

TimeOmni-1 extends a pretrained base LLM, with Qwen2.5-Instruct given as an example, using a lightweight multimodal front end (Guan et al., 29 Sep 2025). The time-series and text modalities are embedded separately:

$R$4

with a simple instantiation

$R$5

Fusion is implemented by prepending the series embedding as a soft prefix to the text embedding and passing the concatenation through $R$6 shared Transformer cross-attention layers:

$R$7

$R$8

The final $R$9 tokens are then fed into the original LLM decoder stack (Guan et al., 29 Sep 2025).

The output interface is standardized across tasks. TimeOmni-1 is trained to produce $$y\in\{\text{“1→2”},\text{“2→1”},\text{“no-causal”}\}$$1 For Tasks 1, 2, and 4, the answer is a single token such as option A/B/C. For Task 3, the answer is a list $y$0. These templates are enforced during both supervised fine-tuning and reinforcement learning via the format reward (Guan et al., 29 Sep 2025).

This unified decoder is methodologically important. Rather than using separate heads per task, all four tasks share the same decoder and reward. The detailed overview states explicitly that the tasks are interleaved within each batch, encouraging the model to learn transferable temporal priors across perception, extrapolation, and decision-making. A plausible implication is that the architecture treats task transfer as a core design goal rather than a byproduct.

4. Optimization pipeline and reward design

TimeOmni-1 is trained in two major stages atop a pretrained LLM: supervised fine-tuning (SFT) followed by reinforcement learning with task-grounded rewards (Guan et al., 29 Sep 2025).

In Stage I, the objective is to inject time-series reasoning priors via high-quality chain-of-thought traces. The loss is

$y$1

where $y$2 come from the 2.3 K human-verified subset, augmented with the remaining 21 K auto-labeled pairs. The reported hyperparameters are learning rate $y$3, batch size $y$4 with accumulation $y$5, and $y$6 epoch, using BF16, FlashAttention, and ZeRO-3 fine-tuning (Guan et al., 29 Sep 2025).

In Stage II, the model is optimized with group relative policy optimization (GRPO). From each $y$7, the method samples $y$8 trajectories $y$9 under $$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$0, groups them, and assigns the advantage

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$1

The policy update uses a clipped PPO-style loss plus a KL penalty to the reference SFT policy $$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$2 (Guan et al., 29 Sep 2025).

The reward design combines structural and task-grounded components. The format reward is

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$3

For discrete tasks, the correctness reward is

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$4

For forecasting, the overview defines

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$5

and

$$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$6

where $$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$7 if $$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$8 is correct and $$(R,y)\sim p_\theta(R,y\mid X,C)=p_\theta(R\mid X,C)\,p_\theta(y\mid R,X,C).$$9. The overall reward is

$y\in\{1,\dots,K\}$0

The Stage II optimization details are GRPO with clip $y\in\{1,\dots,K\}$1, KL weight $y\in\{1,\dots,K\}$2, learning rate $y\in\{1,\dots,K\}$3, batch size $y\in\{1,\dots,K\}$4 with minibatches $y\in\{1,\dots,K\}$5, $y\in\{1,\dots,K\}$6 epochs, $y\in\{1,\dots,K\}$7 trajectories per update, and sampling temperature $y\in\{1,\dots,K\}$8 (Guan et al., 29 Sep 2025).

The training pipeline encodes a specific view of reasoning: chain-of-thought supervision is used to bootstrap temporal priors, and RL is used to refine “genuine reasoning.” That phrasing comes from the ablation summary and should be read as a claim about the observed effect of the training stages rather than a general definition of reasoning (Guan et al., 29 Sep 2025).

5. Evaluation protocol, baselines, and reported results

Evaluation is carried out under both in-distribution and out-of-distribution splits (Guan et al., 29 Sep 2025). After filtering for valid responses using Success Rate, the framework computes Accuracy for Tasks 1, 2, and 4 as

$y\in\{1,\dots,K\}$9

and Mean Absolute Error for Task 3 as

$$y\in\{\text{“1→2”},\text{“2→1”},\text{“no-causal”}\}$$0

The reported baselines include proprietary LLMs such as GPT-4.1, open-source LLMs such as Llama-3.1, Mistral, and Qwen2.5, and time-series-specialized systems including Time-MQA, ChatTS, and Time-R1 (Guan et al., 29 Sep 2025).

The detailed overview lists the following key quantitative results. For Causality Discovery (Task 2, ID), TimeOmni-1 ACC is 69.3% versus 28.7% for GPT-4.1. For Scenario Understanding (Task 1, ID), TimeOmni-1 ACC is 90.7% versus 85.5% for GPT-4.1. For Event-aware Forecasting (Task 3, OOD SR), TimeOmni-1 SR is 82.3% versus 76.1% for GPT-4.1, corresponding to a gain of 6.2 percentage points. For Decision Making (Task 4, OOD), TimeOmni-1 ACC is 58.9% versus 27.8% for GPT-4.1 (Guan et al., 29 Sep 2025).

The abstract reports a partially different set of headline figures: TimeOmni-1 “significantly improves causality discovery accuracy (64.0% vs. 35.9% with GPT-4.1)” and “raises the valid response rate by over 6% compared to GPT-4.1 on the event-aware forecasting task” (Guan et al., 29 Sep 2025). The discrepancy reflects different reported evaluation slices in the supplied materials rather than a single unified table. An objective reading is therefore that the paper consistently claims improvement over GPT-4.1, while the exact comparison point depends on the split and metric summary being quoted.

The out-of-distribution generalization claim is explicit. The overview states that TimeOmni-1 holds strong OOD performance across all four tasks, often improving over GPT-4.1 by 20–30 percentage points in causality and decision-making, and raising valid-response rates by 5–10 points on forecasting. This should be interpreted as a summary statement of the reported empirical pattern rather than a replacement for the task-specific metrics above (Guan et al., 29 Sep 2025).

6. Ablations, interpretation, and naming ambiguity

The ablation studies compare Stage 1 SFT, answer-only SFT, and joint RL, as well as single-task versus multi-task joint training (Guan et al., 29 Sep 2025). The reported conclusions are that CoT-based SFT is essential to bootstrap temporal priors, RL further refines genuine reasoning, and joint multi-task training yields consistent gains, summarized as “train-once, use-across-tasks.” These findings position the model as a unified temporal reasoning system rather than a collection of task-specific adapters.

In interpretive terms, TimeOmni-1 embodies a specific research program in which reasoning with time series is treated as multimodal inference over numerical observations, textual context, and latent reasoning traces. The use of a shared decoder, shared reward, and interleaved training batches suggests an attempt to learn transferable temporal priors that can be reused across classification, directional inference, forecasting, and utility-based action selection. This suggests that the system is meant to generalize across task families, not merely across datasets within a single task.

There is, however, a naming ambiguity in the supplied arXiv materials. The paper explicitly titled “TimeOmni-1: Incentivizing Complex Reasoning with Time Series in LLMs” concerns time-series reasoning (Guan et al., 29 Sep 2025). By contrast, detailed summaries attached to two lunar timing papers use the designation “TimeOmni-1” for a cislunar timekeeping or Lunar Coordinate Time framework, including references to TCL, LCRS, six time scales, sub-picosecond synchronization, and lunar Positioning, Navigation and Timing architectures (Seyffert, 10 Sep 2025); (Turyshev, 29 Jul 2025). Those usages belong to a different technical context. A reasonable editorial distinction is therefore to reserve the primary sense of TimeOmni-1 for the time-series reasoning model of (Guan et al., 29 Sep 2025), while noting that the same label appears elsewhere in the supplied material in connection with relativistic cislunar timing systems.

Within the time-series reasoning literature represented here, the central significance of TimeOmni-1 is that it links benchmark design, hierarchical chain-of-thought annotation, multimodal prefix-style fusion, and task-grounded RL into a single framework for perception, extrapolation, and decision-making over time series (Guan et al., 29 Sep 2025).