TSR-Suite: Time-Series Reasoning Benchmark

• TSR-Suite is a benchmark framework that defines four atomic tasks across perception, extrapolation, and decision-making to foster explicit time-series reasoning.
• It employs structured chain-of-thought annotations, human-guided supervision, and reinforcement learning to ensure context-rich, reliable performance evaluation.
• The suite has evolved into a broader ecosystem integrating multimodal inputs and varied benchmarks, driving research in context-aware temporal reasoning and decision-making.

Time Series Reasoning Suite (TSR-Suite) is a benchmark-and-training framework for time-series reasoning introduced in "TimeOmni-1: Incentivizing Complex Reasoning with Time Series in LLMs" (Guan et al., 29 Sep 2025). In that formulation, TSR-Suite defines four atomic tasks spanning three fundamental capabilities—perception, extrapolation, and decision-making—and provides more than 23K samples, including 2.3K examples curated through a human-guided hierarchical annotation process, to support both evaluation and training of time series reasoning models (TSRMs) (Guan et al., 29 Sep 2025). This suggests that TSR-Suite is both a specific dataset introduced with TimeOmni-1 and a broader organizing concept for later work on multimodal, cognitively structured, and context-aware time-series reasoning benchmarks (Qiu et al., 20 Jun 2026, Yu et al., 26 Jan 2026, Ding et al., 19 Apr 2026).

1. Definition and benchmark philosophy

In the original TimeOmni-1 formulation, TSR-Suite is motivated by the claim that many earlier multimodal time-series datasets do not require genuine time-series reasoning, because their questions are too shallow or their context is insufficient for uniquely determined answers (Guan et al., 29 Sep 2025). Two explicit design principles are proposed.

The first is that question answering must reward reasoning. Let $M_{\text{RM}}$ denote a reasoning model that produces rationales $R$ and an answer $y$, and let $M_{\text{NRM}}$ denote a non-reasoning model that directly outputs $y$. A task is reasoning-critical only if

$$\bar{S}(M_{\text{RM}}) \gg \bar{S}(M_{\text{NRM}}),$$

where $\bar{S}(\cdot)$ is the mean task score (Guan et al., 29 Sep 2025).

The second is that context must be sufficient. If the time series $X$ or auxiliary context $C$ is underspecified, then even infinite reasoning capacity should not materially outperform chance: $$\lim_{RC \to \infty} P(\text{correct} \mid X, C) \approx \frac{1}{K},$$ where $R$0 is the number of options. With sufficient context, the probability should be substantially above $R$1 (Guan et al., 29 Sep 2025).

Within this framework, a TSRM is defined as a model that explicitly samples rationales and answers: $R$2 with rationales emitted inside > tags and final answers inside <answer> tags (Guan et al., 29 Sep 2025). This formalization distinguishes TSR-Suite from conventional forecasting or classification benchmarks that compress all intermediate reasoning into latent states and evaluate only end-task accuracy.

2. Atomic tasks and capability structure

TSR-Suite formalizes four atomic tasks aligned with three capabilities: perception, extrapolation, and decision-making (Guan et al., 29 Sep 2025).

Capability	Atomic task	Typical output
Perception	Scenario Understanding	Discrete choice
Perception	Causality Discovery	Discrete choice
Extrapolation	Event-aware Forecasting	Numerical sequence
Decision-making	Decision Making	Discrete choice

Scenario Understanding presents a single time series together with several textual scenario options and asks which real-world scenario most likely generated the series. The prescribed reasoning template checks series length, magnitude, shape, temporal pattern, alignment with external events, and final elimination of alternatives (Guan et al., 29 Sep 2025).

Causality Discovery uses two river discharge time series sampled every 12 hours and asks whether station A causes B, B causes A, or the pair is non-causal. The reasoning template first checks trend consistency, then peak and dip alignment within $R$3–$R$4 time steps, and finally uses the mean discharge magnitude to determine upstream and downstream direction (Guan et al., 29 Sep 2025).

Event-aware Forecasting provides a historical time series, past events in the history window, and descriptions of upcoming events, then asks for a future numerical sequence. The intended reasoning decomposes the task into baseline pattern extraction, estimation of event impact, and composition of baseline and event effects. Evaluation uses mean absolute error: $R$5 For reinforcement learning, the sequence reward is

$R$6

with $R$7 (Guan et al., 29 Sep 2025).

Decision Making instantiates energy storage arbitrage under dynamic prices. Inputs include 48 hours of historical building load, a price schedule, battery specifications, and a small set of candidate strategies. The reasoning template first forecasts next-day load, then evaluates charge/discharge timing under valley and peak prices, and finally checks savings and feasibility. The paper gives the savings expression

$R$8

subject to state-of-charge constraints (Guan et al., 29 Sep 2025).

This task organization makes TSR-Suite a progressive pipeline from temporal perception to extrapolation and then to action selection. Later benchmark proposals echo this layered view. TSCognition defines a five-task ladder—Decoding, Grounding, Inferring, Extrapolating, and Acting—while TSRBench organizes 15 tasks under Perception, Reasoning, Prediction, and Decision-Making, and HiTSR introduces a four-level taxonomy from numerical read-out to predictive inference (Qiu et al., 20 Jun 2026, Yu et al., 26 Jan 2026, Ding et al., 19 Apr 2026). This suggests that the original TSR-Suite capability structure became a reference point for subsequent taxonomies.

3. Dataset composition, annotation, and evaluation protocol

TSR-Suite contains 23,605 question-answer pairs across four tasks and ten domains, with 2,339 high-quality chain-of-thought examples in the curated Stage-1 subset (Guan et al., 29 Sep 2025). The per-task counts are: Scenario Understanding $R$9 Stage-1 train, $y$0 Stage-2 train, $y$1 ID test, $y$2 OOD test; Causality Discovery $y$3, $y$4, $y$5, $y$6; Event-aware Forecasting $y$7, $y$8, $y$9, $M_{\text{NRM}}$0; and Decision Making $M_{\text{NRM}}$1, $M_{\text{NRM}}$2, $M_{\text{NRM}}$3, $M_{\text{NRM}}$4 (Guan et al., 29 Sep 2025).

The Stage-1 subset supplies rationale supervision. The Stage-2 subset is larger and supports reinforcement learning. The data pipeline uses three components: an LLM Analyzer, human reviewers, and an LLM Rewriter. GPT-4.1 first generates structured reasoning and answers using task-specific templates; if the answer is correct, the rationale is retained. If the answer is incorrect, human reviewers determine whether the problem is sufficiently specified and, if so, write a reasoning trace, which the Rewriter standardizes to the required format (Guan et al., 29 Sep 2025).

For Event-aware Forecasting, exact matching is not possible because real trajectories are noisy. The authors therefore retain examples where GPT-4.1 produces relatively low-MAE forecasts and report that ground-truth-guided chains—where the model is told the future values—hurt training compared with self-consistent but slightly imperfect model-generated forecasts (Guan et al., 29 Sep 2025). This is a distinctive feature of TSR-Suite: it treats imperfect but structured forecasting rationales as useful supervision.

TSR-Suite also distinguishes instruction following from reasoning correctness. It reports Success Rate (SR), defined as the fraction of outputs that can be parsed into valid answers using the required <think>...<answer>...</answer> schema; ACC for discrete tasks; and MAE for Event-aware Forecasting, all computed on valid outputs (Guan et al., 29 Sep 2025). For discrete tasks, correctness reward is

$M_{\text{NRM}}$5

and the combined reward is

$M_{\text{NRM}}$6

with $M_{\text{NRM}}$7 in the reported experiments (Guan et al., 29 Sep 2025).

4. Training role, TimeOmni-1, and empirical results

TSR-Suite was designed not only as a benchmark but also as a training curriculum for TSRMs. TimeOmni-1 uses Qwen2.5-Instruct-7B as its base model and adopts a two-stage training paradigm: supervised fine-tuning on the curated rationale subset, followed by reinforcement learning with Group Relative Policy Optimization (GRPO) on the larger Stage-2 data (Guan et al., 29 Sep 2025).

The supervised stage optimizes the negative log-likelihood of rationale-answer pairs over $M_{\text{NRM}}$8: $M_{\text{NRM}}$9 The reinforcement stage samples multiple rationale-answer trajectories for each input, computes task-grounded rewards, and uses GRPO with group-relative advantages and a KL penalty to the supervised reference model (Guan et al., 29 Sep 2025).

The paper reports that reinforcement learning without the Stage-1 rationale injection is ineffective or harmful, whereas CoT-SFT followed by GRPO yields robust gains, especially on causality discovery and decision-making (Guan et al., 29 Sep 2025). Multi-task training over all four TSR-Suite tasks also outperforms single-task training, and perception-plus-extrapolation supervision transfers to decision-making even when direct task-4 supervision is absent (Guan et al., 29 Sep 2025).

On the main benchmark, TimeOmni-1 achieves 69.3% ID accuracy and 64.0% OOD accuracy on Causality Discovery, compared with 28.7% and 35.9% for GPT-4.1; 90.7% ID and 87.7% OOD on Scenario Understanding; 47.9% ID and 58.9% OOD on Decision Making, compared with 25.5% and 27.8% for GPT-4.1; and 145.53 OOD MAE on Event-aware Forecasting, compared with 170.78 for GPT-4.1, while also improving valid response rate from 76.1% to 82.3% on that task (Guan et al., 29 Sep 2025). These results position TSR-Suite as both an evaluation protocol and a task mixture for cultivating explicit time-series reasoning rather than pure numerical regression.

5. Expansion into a broader benchmark and modeling ecosystem

Later work extends the underlying idea of TSR-Suite in several directions. TSCognition defines cognitive time-series reasoning as a shift from trajectory prediction to context-grounded reasoning over multivariate time series and text, and introduces a multimodal benchmark with 41,086 multiple-choice QA pairs from 15 public sources and five tasks: Decoding, Grounding, Inferring, Extrapolating, and Acting (Qiu et al., 20 Jun 2026). TSRBench broadens the scope to 4,125 problems from 14 domains across 15 tasks and four dimensions—Perception, Reasoning, Prediction, and Decision-Making—and reports that scaling laws hold for perception and reasoning but break down for prediction (Yu et al., 26 Jan 2026). HiTSR contributes a four-level taxonomy from numerical read-out to predictive inference and 83k samples, together with LLaTiSA, a dual-view VLM that combines plots and precision-calibrated numerical tables (Ding et al., 19 Apr 2026). TFRBench, in turn, targets forecasting systems and evaluates reasoning traces over cross-channel dependencies, trends, seasonality, and external events across ten datasets and five domains (Ahamed et al., 7 Apr 2026). TSVer adapts time-series reasoning to fact verification with 287 real-world claims, 400 curated time series, structured evidence ranges, and verdict justifications (Strong et al., 2 Nov 2025). This suggests that the phrase "TSR-Suite" increasingly denotes a family of benchmark designs rather than a single fixed corpus.

A parallel literature addresses the modeling side. TSAlign aligns patch-level time-series representations with semantic directions in an LLM embedding space through attention alignment, gated residual injection, and multivariate fusion, and reports strong gains on TSCognition and TimerBed (Qiu et al., 20 Jun 2026). TS-Reasoner instead freezes a pretrained TS foundation model, inserts TS-to-text adapter outputs at <ts></ts> placeholders inside a Qwen2.5-7B-Instruct backbone, and uses a two-stage alignment-then-instruction recipe built from 120K synthetic caption pairs and 30K instruction examples (Yu et al., 3 Oct 2025). RationaleTS treats retrieved rationales as guiding reasoning units rather than post-hoc explanations, using label-conditioned rationales, hybrid retrieval over temporal and semantic embeddings, and in-context inference on Finance, Traffic, and Power tasks (Liu et al., 6 Jan 2026). SenTSR-Bench separates TSLM perception from GRLM reasoning and injects RLVR-trained TSLM traces directly into a GRLM’s reasoning buffer for industrial diagnostics (He et al., 23 Feb 2026). ReasonTSC shows that, in time-series classification, tailored multi-turn reasoning plus plug-in classifier logits can outperform both vanilla CoT and standalone plug-in models (Zhou et al., 1 Jun 2025). Together, these systems indicate that later TSR-style research combines benchmark design with increasingly explicit interfaces between temporal encoders, rationales, retrieval, and general-purpose reasoning models.

6. Limitations, misconceptions, and future directions

Several recurrent limitations appear across the TSR-Suite literature. The original TSR-Suite covers ten domains and four tasks, but this is not exhaustive, and its chain-of-thought annotations partly inherit the biases of GPT-4.1 and a limited volume of human review (Guan et al., 29 Sep 2025). Event-aware Forecasting is evaluated by MAE on a single realized future, even though multiple plausible futures may exist (Guan et al., 29 Sep 2025). TSCognition identifies multiple-choice format as a limitation for explanation quality and interactive decision-making, and notes open problems involving noise, missing data, irregular sampling, and more severe domain shifts (Qiu et al., 20 Jun 2026). TSRBench shows that strong semantic reasoning does not guarantee accurate context-aware forecasting and that multimodal models still fail to fuse textual and visual representations effectively for reciprocal gains (Yu et al., 26 Jan 2026). HiTSR argues that high-level reasoning benchmarks can remain fragile if lower-level numerical grounding is not explicitly evaluated (Ding et al., 19 Apr 2026).

A common misconception addressed by several papers is that LLMs should primarily be used as regressors over tokenized numbers. "Toward Reasoning-Centric Time-Series Analysis" argues instead for reframing time-series analysis from “extracting trends and correlations from observed data” to “a reasoning process,” and warns that many current LLM-based methods exploit numerical regression ability while neglecting deeper reasoning potential (Wang et al., 14 Oct 2025). TSCognition similarly argues that textualization leads to extremely long token sequences and poor discretization of continuous signals, whereas chart-based visual encodings can distort temporal dependencies and fine-grained numeric patterns (Qiu et al., 20 Jun 2026). The BlueSky vision paper goes further, defining time-series reasoning systems as models that produce both outputs and reasoning paths,

$y$0

and frames interpretability and trustworthiness as first-class requirements for future systems (Ning et al., 19 Oct 2025).

Future work therefore converges on several directions. One is broader task coverage: open-ended QA, explanation generation, counterfactual reasoning, long-form reporting, root-cause analysis, and multi-turn decision dialogue are repeatedly proposed as extensions (Qiu et al., 20 Jun 2026, Wang et al., 14 Oct 2025, Strong et al., 2 Nov 2025). Another is richer multimodality and retrieval, including text, images, tables, logs, and retrieved historical episodes (Ning et al., 19 Oct 2025, Ahamed et al., 7 Apr 2026). A third is more faithful reasoning evaluation, including rationale faithfulness, evidence alignment, attribution consistency, and calibrated uncertainty rather than only end-task accuracy (Ning et al., 19 Oct 2025, Strong et al., 2 Nov 2025). This suggests that TSR-Suite, in its mature form, is best understood not as a closed benchmark but as a research program for moving from surface-level time-series recognition to explicit, context-aware, and decision-relevant temporal reasoning.