TS-Reasoner Multimodal Reasoning System

• TS-Reasoner is a multimodal system that integrates frozen time series encoders with LLMs to jointly handle temporal patterns and textual context.
• It employs a two-stage training process with synthetic caption alignment and instruction finetuning, ensuring improved data efficiency and robust performance.
• Experiments show TS-Reasoner outperforms LLM, VLM, and TSLLM baselines on benchmarks like TimeSeriesExam and MTBench in accuracy and scalability.

Searching arXiv for the exact TS-Reasoner paper and related uses of the term. Searching arXiv for related papers that use “TS-Reasoner” in broader or alternate senses. TS-Reasoner is a multimodal reasoning system that aligns the latent representations of time series foundation models (TSFMs) with the textual inputs of LLMs for downstream understanding and reasoning tasks. Its reported design combines a pretrained, frozen TSFM encoder, a trainable TS-to-text adapter, and a pretrained LLM backbone, and it is trained in two stages: alignment pretraining through caption generation and instruction finetuning for time series reasoning. The implementation described in the literature uses TimesFM-1.0-200M, Qwen2.5-7B-Instruct, and an MLP adapter, and the reported experiments show that TS-Reasoner outperforms a range of LLM, VLM, and TSLLM baselines while achieving notable data efficiency, including superior overall accuracy with 60K alignment samples and instruction tuning with 10K samples that suffices to outperform ChatTS-7B (Yu et al., 3 Oct 2025).

1. Motivation and problem formulation

TS-Reasoner is motivated by the gap between temporal representation learning and language-based reasoning. Time series reasoning is presented as a task that extends beyond forecasting: applications in finance, energy usage, traffic, weather, and scientific discovery require recognition of temporal patterns together with interpretation grounded in domain knowledge, contextual information, and higher-level reasoning. In this setting, TSFMs are described as strong at modeling low-level dynamics and forecasting but limited in high-level reasoning and direct text integration, whereas LLMs are described as strong at knowledge use and reasoning but weak on the numerical and temporal dependencies of time series without expensive post-training and specialized data (Yu et al., 3 Oct 2025).

The system is therefore framed as an alignment method that grounds LLM reasoning in robust temporal features learned by a pretrained TSFM, while mapping those features into the LLM’s input space. Existing approaches based on converting time series into strings or plots are characterized as only partial solutions: LLMs lack temporal inductive bias and numerical precision, and they often miss periodicity, trends, and noise types while failing to couple text context with numeric signals. TS-Reasoner addresses this by preserving the TSFM as a frozen temporal encoder and learning a bridge into the LLM’s embedding space, so that downstream understanding and reasoning can use both modalities jointly (Yu et al., 3 Oct 2025).

2. Architecture and input representation

The reported TS-Reasoner architecture has three principal components: a pretrained, frozen TSFM encoder, a TS-to-Text adapter, and a pretrained LLM backbone. The TSFM is TimesFM-1.0-200M, described as a decoder-only TSFM with latent dimension $d_{ts} = 1080$. The LLM is Qwen2.5-7B-Instruct, with input embedding dimension $d_{LLM} = 5120$. The adapter is an MLP that maps TSFM latent vectors into the LLM’s input embedding space, enabling the LLM to consume time-series features as if they were language tokens (Yu et al., 3 Oct 2025).

Component	Role	Configuration
TimesFM-1.0-200M	Pretrained temporal encoder	Frozen, $d_{ts}=1080$
MLP adapter	TS-to-text projection	$$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$
Qwen2.5-7B-Instruct	Reasoning backbone	Trainable, $d_{LLM}=5120$

The data flow begins with a natural language instruction $X$ and a set of $K$ time series $S = \{T_0, \ldots, T_K\}$. Each series is instance-normalized to zero mean and unit variance, then partitioned into non-overlapping patches of fixed size $P = 32$, giving $N_i = \lfloor L_i / P \rfloor$ patches for series $d_{LLM} = 5120$0. The TSFM encodes each patched series into a sequence $d_{LLM} = 5120$1, which is projected by the MLP into $d_{LLM} = 5120$2. In parallel, the instruction is tokenized into text embeddings $d_{LLM} = 5120$3. Each aligned time-series sequence is inserted into the text embedding sequence at positions marked by the placeholders "<ts> <ts/>", and the fused sequence $d_{LLM} = 5120$4 is then processed by the LLM Transformer (Yu et al., 3 Oct 2025).

The formal presentation is:

$d_{LLM} = 5120$5

$d_{LLM} = 5120$6

$d_{LLM} = 5120$7

$d_{LLM} = 5120$8

$d_{LLM} = 5120$9

This design allows multiple series to be inserted at user-controlled positions, and the paper states that multivariate streams can be passed either as multiple series, with one placeholder per variable, or encoded jointly by the TSFM, depending on the TSFM’s interface. Irregular sampling and missing data are normalized before patching, but explicit imputation strategies are stated not to be part of the recipe (Yu et al., 3 Oct 2025).

3. Synthetic time series–text alignment corpus

A central component of TS-Reasoner is its synthetic alignment corpus. The stated challenge is the lack of diverse, high-quality time series–text pairs for alignment: template-based descriptions are numerically correct but structurally limited, which encourages spurious template learning and weak generalization. TS-Reasoner addresses this with attribute-aware caption synthesis using advanced LLMs and VLMs (Yu et al., 3 Oct 2025).

Two data sources are reported. The first is synthetic numeric series from Chronos, used to provide diverse temporal patterns without text context. The second is text-attributed series collected from Merrill et al., used to provide time series paired with real-world textual context. Numeric series are converted into plots via Python, denoted as $d_{ts}=1080$0, and the image is provided to a multimodal captioning model together with a captioning instruction enriched with attributes. The attributes include trend, frequency content, periodicity, noise level, local variations, anomalies, and other salient features. To increase lexical and structural diversity, the system generates $d_{ts}=1080$1 paraphrases of the base instruction and samples one uniformly for each series (Yu et al., 3 Oct 2025).

The generation procedure is stated as

$d_{ts}=1080$2

with

$d_{ts}=1080$3

GPT-4.1 (vision) is reported as the primary captioner, while InternVL3-8B and Qwen2.5-VL-7B are used to study fidelity effects. The paper states that higher-quality captioners yield better downstream performance. Diversity is quantified by reporting MTLD $d_{ts}=1080$4 versus $d_{ts}=1080$5 for templates, and Self-BLEU-4 $d_{ts}=1080$6 versus $d_{ts}=1080$7, which is presented as evidence that attribute-aware captions reduce the risk of spurious correlations. The captions are also described as covering pattern recognition, noise understanding, anomaly description, similarity analysis, and causal cues, thereby inducing both intrinsic temporal understanding and cross-modal grounding (Yu et al., 3 Oct 2025).

4. Two-stage training recipe

Training proceeds in two separate stages. Stage 1 performs alignment through caption generation, and Stage 2 performs instruction finetuning for time series reasoning. The TSFM remains frozen throughout training, while the LLM and adapter are trainable in both stages. The reported rationale is that stage 1 emphasizes grounding and alignment, whereas stage 2 elicits instruction following and reasoning behavior (Yu et al., 3 Oct 2025).

In Stage 1, the objective is autoregressive language modeling with teacher forcing over caption tokens $d_{ts}=1080$8:

$d_{ts}=1080$9

The data mix combines template-based captions from ChatTS with LLM-generated attribute-aware captions. The stated effect is a combination of numerically faithful structured captions and rich, diverse descriptions (Yu et al., 3 Oct 2025).

In Stage 2, instruction finetuning uses supervised cross-entropy on instruction-response pairs:

$$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$0

The instruction-following data come from ChatTS and cover time series understanding and reasoning tasks. No explicit chain-of-thought supervision is described. The paper notes that a combined objective

$$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$1

could be used optionally, but the reported TS-Reasoner training proceeds sequentially through the two stages rather than jointly (Yu et al., 3 Oct 2025).

The implementation details are explicit. Stage 1 uses learning rate $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$2, batch size $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$3, and $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$4 epoch over $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$5K caption samples. Stage 2 uses learning rate $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$6, batch size $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$7, and $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$8 epochs over $$f:\mathbb{R}^{d_{ts}} \rightarrow \mathbb{R}^{d_{LLM}}$$9K instruction samples. Training is reported on $d_{LLM}=5120$0 NVIDIA L40s GPUs. The number of trainable parameters is approximately $d_{LLM}=5120$1B, corresponding to the LLM plus adapter, while the TSFM remains frozen (Yu et al., 3 Oct 2025).

5. Benchmarks, results, and ablations

TS-Reasoner is evaluated on two principal benchmarks. TimeSeriesExam is a multiple-choice benchmark probing time series understanding and reasoning, with subtasks Pattern Recognition (PR), Noise Understanding (NU), Anomaly Detection (AD), Similarity Analysis (SA), and Causality Analysis (CA), using accuracy as the metric. MTBench covers finance and weather domains, with short-term (7-day) and long-term (14-day) horizons, also evaluated by accuracy (Yu et al., 3 Oct 2025).

On TimeSeriesExam, TS-Reasoner-7B reports overall accuracy $d_{LLM}=5120$2. The best reported 7B LLM baseline, Qwen2.5-7B-Instruct, is approximately $d_{LLM}=5120$3; the best reported VLM baseline, InternVL3-8B, is approximately $d_{LLM}=5120$4; and the best reported 7B TSLLM baseline, ChatTS-7B, is $d_{LLM}=5120$5. ChatTS-14B reaches $d_{LLM}=5120$6, and TS-Reasoner-7B is described as competitive despite being smaller. Subtask gains over the second-best baseline are reported as PR $d_{LLM}=5120$7, NU $d_{LLM}=5120$8, AD $d_{LLM}=5120$9, and SA $X$0, while CA is tied (Yu et al., 3 Oct 2025).

On MTBench, TS-Reasoner-7B reports Finance long $X$1, Finance short $X$2, Weather long $X$3, and Weather short $X$4. Gains over the best 7B baselines are reported to range from $X$5 to $X$6 points. Data efficiency is emphasized separately: alignment achieves superior overall accuracy with $X$7K samples, described as less than half the training data of ChatTS-7B, and instruction tuning with $X$8K samples suffices to outperform ChatTS-7B (Yu et al., 3 Oct 2025).

The ablation studies isolate the main design choices. Removing LLM-generated captions lowers TimeSeriesExam overall accuracy to $X$9, a drop of $K$0. Removing attributes while keeping visual captions lowers it to $K$1, with an approximately $K$2 effect on MTBench. Removing Stage 1 yields overall accuracy $K$3 on TimeSeriesExam, while removing Stage 2 yields $K$4, which is described as a severe degradation in instruction following. Replacing TimesFM-200M with MOMENT-200M reduces overall accuracy from $K$5 to $K$6. Projecting patches directly to the LLM without the TSFM reduces performance to $K$7, corresponding to a $K$8 drop on TimeSeriesExam and a $K$9 drop on MTBench. The scaling study over Qwen2.5 backbones at $S = \{T_0, \ldots, T_K\}$0B, $S = \{T_0, \ldots, T_K\}$1B, and $S = \{T_0, \ldots, T_K\}$2B reports a positive scaling law for both the baseline and TS-Reasoner, with consistent TS-Reasoner leads of $S = \{T_0, \ldots, T_K\}$3, $S = \{T_0, \ldots, T_K\}$4, and $S = \{T_0, \ldots, T_K\}$5 overall accuracy, respectively (Yu et al., 3 Oct 2025).

6. Positioning, limitations, and broader uses of the label

TS-Reasoner is positioned against prior TS-LLM approaches that either train LLMs to ingest raw numeric strings or train time-series encoders from scratch for narrow domains. Its stated novelty consists of freezing a strong pretrained TSFM to provide robust temporal features and data efficiency, using alignment-first training through attribute-aware high-fidelity synthetic caption pairs, separating grounding from instruction following through a two-stage recipe, and allowing multiple series to be inserted flexibly into the LLM context via special tokens and an adapter (Yu et al., 3 Oct 2025).

The reported limitations are equally explicit. Although the TSFM grounds temporal features, the LLM may still be brittle on fine-grained numerical queries or extremely long-range dependencies. Performance can degrade under unseen regimes, irregular sampling, or heavy missingness, and explicit imputation or irregular-time handling is not part of the current recipe. As with LLMs more generally, TS-Reasoner can hallucinate causal claims. Domain bias may arise from the domains represented in captions and instructions. The experiments focus on up to 7B LLMs, and larger models were not tested because of compute limitations. The future directions proposed in the literature include integrating richer modalities such as events, knowledge graphs, and exogenous variables; adding tool use, statistical tests, retrieval augmentation, or uncertainty estimation for improved numerical reasoning; extending to causal inference over time series; studying scaling with larger backbones; and improving online or irregular-time adaptation (Yu et al., 3 Oct 2025).

The label “TS-Reasoner” also has a broader and sometimes ambiguous use in recent literature. In test-time scaling work, “TS-Reasoner” is described as a broad class of methods that adapt test-time compute and reasoning by scaling think steps, sampling more chains, invoking tools, or stopping early; within that usage, Route-To-Reason is presented as an instance that jointly routes across models and reasoning strategies under a cost-aware objective (2505.19435). In agent architectures, the term is also used for the slow, deliberative System-2 component of a Talker-Reasoner design, where a Reasoner performs multi-step reason–act–extract loops, tool use, and belief updates while a Talker handles low-latency user interaction (Christakopoulou et al., 2024). The same label has been applied in formal reasoning contexts as well, including a configurable trace-semantics reasoner for software traceability and a C++ KE-tableau reasoner for the description logic $S = \{T_0, \ldots, T_K\}$6 (Erata et al., 2024, Cantone et al., 2017). This suggests that, in contemporary research usage, “TS-Reasoner” functions both as the proper name of a specific time-series reasoning model and as a broader descriptor for systems that mediate between structured signals and higher-level reasoning.