Large language models (LLMs) can struggle with time-series question answering (TS-QA), especially when numerical signals are serialized as text and require explicit computation. Tool-augmented approaches improve performance, but existing systems often intertwine language reasoning, computation, and perception, making it difficult to determine which components drive the gains. We present T-SMART, a neurosymbolic framework that separates these roles: a frozen LLM interprets questions and selects operations, deterministic tools perform numerical computation, and structured perception is invoked only when needed. Controlled paired ablations show that deterministic computation provides the dominant benefit, improving accuracy by 31.7 percentage points over direct LLM reasoning on serialized time series, while language understanding and perception offer smaller complementary gains. These results indicate that tool-augmented TS-QA benefits primarily from reliable numerical execution rather than additional language-model reasoning and provide a controlled framework for analyzing component contributions in neurosymbolic time-series systems.
Figures & tables
Fig. 1: The single-pass T-SMART pipeline. One frozen model serves as router, vision sensor, and answer reader; deterministic tools perform every numerical computation, and the visual sensor is invoked only when the trust gate fires.
TABLE I: Deterministic branches and the core tools each invokes; the per-branch subtype registry is released with the code.
Name
Setting and role
baseline
trust gate, no loop; A0 numeric head on free-response. Canonical T-SMART; the control condition
vision_override
suppress vision on noise, force anomaly line plots. TimeSeriesExam-selected routing intervention, frozen before OOD testing; a mechanism-transfer ablation
learned_gate
per-branch suppression filter over the trust gate; a lower-look-rate policy fitted on MMTS discordance, evaluated on held-out TimeSeriesExam
vision_off
suppress vision on all branches; the math-only counterfactual
head_grammar , head_planner
A1 / A2 parser over the A0 default; free-response numeric-head variants
llm_only
no router, branch, or tools; one direct answer call. Raw-model control for the fixed-backbone analysis
TABLE II: Configurations used as experimental conditions. A configuration is a named set of pipeline arguments; prompts are identical across conditions.
Model
Setting
OA
PR
NU
AD
SA
CA
Phi-3.5
text
31.8
44
24
25
41
25
Gemini-2.5
text
30.4
29
32
26
35
30
Mistral-7B
text
34.6
29
33
38
45
28
GPT-4o
text
38.8
35
32
40
51
36
ChatTS
TS model
39.0
37
39
36
53
30
Gemini-2.5
vision
48.8
51
54
46
69
24
TABLE III: TimeSeriesExam, per-category accuracy (%). OA is the category-mean overall accuracy.
Model
Micro OA
Base
InWild
Match
Align
ChatTS
49
39
50
37
80
TS-Agent
60
21
71
77
97
T-SMART baseline
55.6
72.7
48.3
59.0
32.9
TABLE IV: MMTS-Bench, accuracy per subset (%). Micro OA is the row-weighted micro-average. Base uses the numeric head on free-response.
Comparison
n
n10/n01
Δ
95% CI
p
all numerical rows
393
128/12
29.5
[+24.4,+34.6]
< 0.0001
closed-form
328
104/0
31.7
[+26.8,+36.9]
< 0.0001
structured vs. math-only
568
113/69
7.7
[+3.2,+12.3]
0.0014
raw pixels vs. math-only
568
119/62
10.0
[+5.5,+14.6]
< 0.0001
SVI cost (struct. − raw)
568
−2.3
TABLE V: Paired mechanism effects on identical rows (points). Top: MMTS Base, numeric head vs. llm_compute , scored by Acc@10. Bottom: TimeSeriesExam gated subset, each visual condition vs. math-only.
Layer
A0
A1
A2
A2 fire %
clean
100
100
100
0
compositional
15
100
100
0
typo
0
83
100
17
paraphrase
0
0
75
100
overall ( n=240 )
29
71
94
29
TABLE VI: Three-parser study on the synthetic numeric stress set (Acc@10, %; gold = the deterministic tool’s own answer; n=60 per layer).
Fig. 2: Gated vision helps only where shape and periodicity are decisive: on gate-fired rows (gemini-3.1-flash-lite), structured vision improves periodicity ( +24.8 ) and trend ( +14.3 ), is near zero on anomaly, and negative on noise ( −4.2 ).
Time series data widely exist in real-world cyber-physical systems. Though analyzing and interpreting them contributes to significant values, e.g, disaster prediction and financial risk control, current workflows mainly rely on human data scientists, which requires significant labor costs and lacks automation. To tackle this, we introduce TimeART, a framework fusing the analytical capability of strong out-of-the-box tools and the reasoning capability of Large Language Models (LLMs), which serves as a fully agentic data scientist for Time Series Question Answering (TSQA). To teach the LLM-based Time Series Reasoning Models (TSRMs) strategic tool-use, we also collect a 100k expert trajectory corpus called TimeToolBench. To enhance TSRMs' generalization capability, we then devise a four-stage training strategy, which boosts TSRMs through learning from their own early experiences and self-reflections. Experimentally, we train an 8B TSRM on TimeToolBench and equip it with the TimeART framework, and it achieves consistent state-of-the-art performance on multiple TSQA tasks, which pioneers a novel approach towards agentic time series reasoning.
Xingjian Wu, Junkai Lu, Zhengyu Li +5
East China Normal University, Shanghai, China · Aalborg University, Aalborg, Denmark
Large language models (LLMs) and time-series language models (TSLMs) are increasingly applied to time-series question answering (TSQA). Unlike text-only QA, TSQA requires models to ground answers in temporal signals whose patterns may occur at different scales, specific time locations, or across separated intervals. However, existing benchmarks are typically organized by task types or high-level reasoning categories, making it difficult to diagnose the underlying signal-level capabilities driving model performance. We introduce TS-Skill, a controlled benchmark for evaluating three composable analytical skills in TSQA: temporal scale selection (SK1), temporal localization (SK2), and cross-interval integration (SK3). TS-Skill provides timestamp-aware questions, broad domain coverage, and human-validated QA quality. To construct the benchmark at scale, we develop SKEvol, a skill-guided agentic framework that combines domain-aware time-series seed generation, skill-controlled question generation, metadata- and code-assisted answer construction, multi-phase signal-grounded verification, and human-in-the-loop curation. Experiments on ten state-of-the-art LLMs and TSLMs reveal substantial and uneven capability gaps across SK1-SK3. In particular, SK3 remains consistently challenging for non-agent models, whereas tool-augmented agents show a selective advantage on standalone SK3. These findings demonstrate that skill-level evaluation can uncover temporal reasoning failures that are obscured by aggregate TSQA scores.
Liying Han, Kang Yang, Oliver Wang +9
University of California, Los Angeles · Samsung Research America · Carnegie Mellon University +2
Recent advances in large language models (LLMs) have given rise to time-series question answering (TSQA), which formulates time-series analysis as natural-language question answering. However, directly feeding raw numerical series into LLMs suffers from a tokenization bottleneck: Byte Pair Encoding fragments continuous values into unstable tokens whose embeddings lack meaningful metric structure, resulting in the loss of magnitude, scale, and trend information. Prior methods use patch-based encoders that split the series into fixed windows, locking in one granularity that breaks patterns and hides exact timesteps, through a separate module that rarely transfers across datasets with different lengths or sampling rates. To address this challenge, we propose CADE (Contrastive Alignment with Direct Embedding), a novel framework for TSQA built upon two key components: direct timestep embedding and semantic alignment. The proposed framework maps each timestep directly into the LLM embedding space through a point-wise linear encoder and MLP projector, preserving exact index-level access while eliminating the need for patching and padding. To further bridge the semantic gap between time-series and language representations, we introduce a novel one-directional supervised contrastive loss that aligns time-series embeddings with frozen class-name text anchors. Experimental results on the public Time-MQA benchmark demonstrate that our framework consistently improves performance across six TSQA tasks, outperforming both open-source and proprietary LLM baselines.