PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering

Junkai Lu; Peng Chen; Xingjian Wu; Yang Shu; Chenjuan Guo; Christian S. Jensen; Bin Yang

cs.AIFeb 26, 2026

PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering

Authors: Junkai Lu, Peng Chen, Xingjian Wu, Yang Shu, Chenjuan Guo, Christian S. Jensen, Bin Yang

Organizations: East China Normal University, Shanghai, China · Aalborg University, Aalborg, Denmark

Abstract

Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitations: they often treat time series merely as text or images, failing to capture the patterns like trends and seasonalities needed to answer specific questions; and when trained on a mix of simple and complex tasks, simpler objectives often dominate the learning process, hindering the development of deep reasoning capabilities. To address these limitations, we propose the Pattern-Aware Alignment and Balanced Reasoning model (PATRA), introducing a pattern-aware mechanism that extracts trend and seasonality patterns from time series to achieve deep alignment. Furthermore, we design a task-aware balanced reward to harmonize learning across tasks of varying difficulty, incentivizing the generation of coherent Chains of Thought. Extensive experiments show that PATRA outperforms strong baselines across diverse Time Series Question Answering (TSQA) tasks, demonstrating superior cross-modal understanding and reasoning capability.

Figures & tables

Figure 1: Comparison of alignment paradigms in Time Series Question Answering. (a) Single-Modality Reasoning treats time series as text sequences, converting continuous numerical points into text tokens. (b) Multimodal Shallow Alignment introduces time series as a separate modality but relies on simple concatenation of embeddings. (c) Multimodal Deep Alignment addresses these limitations by explicitly decomposing time series into distinct patterns and fusing them with text embeddings via a dedicated module, ensuring that semantic reasoning is firmly grounded in physical data behaviors.

Figure 2: The framework of PATRA, which contains a Text Encoder for obtaining text embedding from pre-trained representations, a TS Encoder for embedding the time series, a Pattern-Aware Alignment to align the gap between text embedding and time series embedding, and an LLM Backbone to reason the questions.

Figure 3: Overview of PATRA’s two training stages. Alignment Stage performs supervised fine-tuning (SFT) with cross-entropy (CE) loss, while Reasoning-Enhanced Stage applies reinforcement learning, including format and task reward under the GRPO strategy to optimize the policy model.

Modalities	Models	Comp.		Recog.		Reason.		Presc.
Modalities	Models	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$
Textual	GPT-4o	50.86	11.99	69.65	4.75	50.00	7.75	66.66	6.78
	TimeOmni-1 7B	37.93	14.16	41.93	9.10	42.56	12.78	42.59	10.40
	Deepseek-R1 7B	40.52	13.65	12.41	8.40	18.24	10.34	14.81	8.59
	Llama3 8B	33.62	5.15	36.27	4.11	27.02	5.26	25.92	3.22
	Qwen2.5 7B	42.24	18.77	45.51	10.32	36.48	18.72	26.85	10.67

Table 1: Evaluation results for the four tasks, with the best and second-best performances highlighted respectively in open-source model in red with bold and orange with underline. We regard GPT-4o as the upper bound, which is not in the comparison set.

Type	Models	Weather		Finance (7 days)			Finance (30 days)
Type	Models	MCQ. $\uparrow$	Pre. $\uparrow$	3-way $\uparrow$	5-way $\uparrow$	MSE $\downarrow$	3-way $\uparrow$	5-way $\uparrow$	MSE $\downarrow$
Proprietary Models	GPT-5.2	59.16	43.76	52.35	49.80	0.196	48.95	36.05	0.985
	Gemini-3-Pro	58.74	43.68	53.11	49.79	0.209	42.20	30.27	0.999
	GPT-4o	41.70	43.54	42.81	36.45	0.365	47.35	30.58	0.897
	Gemini	56.96	51.76	47.30	43.40	0.384	44.90	29.70	0.975
	Claude	59.78	56.87	44.90	33.40	0.373	52.05	31.70	1.171

Table 2: Out of domain results about weather news-driven multiple-choice QA (MCQA), weather trend prediction accuracy, finance 3 and 5 stock trend prediction accuracy and finance MACD prediction MSE. The best and second performances are highlighted respectively in red with bold and orange with underline. And "-" represents that the model can not generate correct answers.

Models	Reason.		Presc.
Models	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$
W/ Single-Pattern Alignment	35.81	26.19	37.03	24.03
W/O Pattern-Aware Alignment	28.37	16.81	30.55	16.94
W/O Reasoning-Enhanced Stage	13.51	2.92	16.66	13.06
PATRA	44.59	27.36	52.78	27.06

Table 3: Ablation Studies on PATRA. Due to space limitations, we only report results on the two tasks folowing, while other tasks show similar results in Appendix C .

Model	Setting	Comp. $\uparrow$	Recog. $\uparrow$	Reason. $\uparrow$	Presc. $\uparrow$
	Zero-shot	12.93	12.41	18.24	14.81
DeepSeek-R1-7B	Tuned	40.52	29.65	30.40	25.00
	Zero-shot	6.03	19.72	10.13	27.77
TimeMQA-Qwen-7B	Tuned	10.34	24.41	13.51	26.85
	Zero-shot	6.90	12.41	10.81	9.26
TimeMQA-Mistral-7B	Tuned	18.96	26.20	12.83	28.70

Table 4: Performance comparison under Zero-shot and Tuned settings across four tasks.

Figure 4: Case Study of TS Reasoning Task, comparing the answers of a single-modality LLM (Qwen2.5), a multimodal LLM (ChatTS), and our proposed model (PATRA).

Models	Comp.		Recog.		Reason.		Presc.
Models	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$
Original Reward	39.65	21.92	48.83	19.00	37.84	21.64	35.18	16.88
Balanced Reward	56.03	25.67	64.69	25.46	44.59	27.36	52.78	27.06

Table 5: Ablation study on the Reward Balancing Strategy. "Original Reward" refers to training with raw, unscaled rewards, while "Balanced Reward" maps all task rewards to

[0,2]

Figure 5: Attention bias in Pattern-Aware Alignment.

Figure 6: The analysis study of Latent Decomposition.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Models	Comp.		Recog.		Reason.		Presc.
Models	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$	Acc. $\uparrow$	Rou. $\uparrow$
W/ Single-Pattern Alignment	44.82	24.59	58.48	23.50	35.81	26.19	37.03	24.03
W/O Pattern-Aware Alignment	40.51	15.36	47.03	15.18	28.37	16.81	30.55	16.94
W/O Reasoning-Enhanced Stage	40.51	3.21	16.68	5.51	13.51	2.92	16.66	13.06
PATRA	56.03	25.67	64.69	25.46	44.59	27.36	52.78	27.06

Appendix

Table 6: Ablation Studies on PATRA. Due to space limitations, we only report results on the two tasks folowing, while other tasks show similar results in Appendix C .

Figure 7: The Loss of Alignment stage and the Total Reward Curve in Reasoning-Enhanced Stage.

Type	Models	Forecasting	Anomaly Detection	Classification
Type	Models	MSE $\downarrow$	Acc. $\uparrow$	Acc. $\uparrow$
Proprietary Models	Doubao	-	0.52	0.44
Proprietary Models	GPT-4o	1.79	0.64	0.32
Open-source Models	backbones	TimeMQA
	Llama-3-8B	2.01	0.54	0.24
	Qwen-2.5-7B	1.82	0.68	0.52

Appendix

Table 7: Three Common Time Series Tasks’ Evaluations. We use the Accuracy metric for Anomaly Detection and Classification tasks and the MSE metric for the Forecasting task.

Time Series Reasoning

Time series reasoning studies how models interpret temporal patterns, answer questions about sequences, and use evidence to make forecasts.

See all 44 papers

Lapras: Latent Reasoning for Time Series Language Models
Oct 8, 2026Yuliang Chen, Yu Yvonne Wu, Patrick Langer +11Time Series Language Modeling Temporal Reasoning in Language Models
Dartmouth College · Aionic Labs · Agentic Systems Lab, ETH Zurich +2
TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models
Sep 11, 2026Sudarshan Regmi, Arvind Pillai, Yu Yvonne Wu +7Time Series Reasoning Time Series QA
Dartmouth College
From Recognition to Understanding: Unlocking Cognitive Time Series Reasoning with LLMs
Jun 20, 2026Xin Qiu, Junlong Tong, Yao Zhang +3Time Series Reasoning Time Series QA
Institute of Digital Twin, Eastern Institute of Technology, Ningbo · Zhejiang University · Munich Center for Machine Learning, LMU

Oct 8, 2026J/K move · Enter open · S save

Lapras: Latent Reasoning for Time Series Language Models

Yuliang Chen, Yu Yvonne Wu, Patrick Langer, Arvind Pillai+10

Dartmouth College · Aionic Labs · Agentic Systems Lab, ETH Zurich+2

Time Series Language Modeling Temporal Reasoning in Language Models Time Series QA Continuous Latent Reasoning

Open paper AskPDF arXiv

Time Series Language Models (TSLMs) offer a promising path toward time series understanding by reasoning over temporal signals and producing natural language answers and explanations. A common approach is Chain-of-Thought (CoT), which generates step-by-step rationales linking relevant signal patterns to final answers. Although these models learn from reference CoT traces during post-training, generating faithful descriptions of input time series at inference remains challenging. Expressing high-dimensional, continuous temporal representations in discrete language tokens may cause the model to neglect task-relevant patterns or describe them inaccurately. Because later reasoning steps build on these descriptions, early errors propagate, leading to incorrect answers with plausible explanations that are inconsistent with the input signal. We propose Lapras (Latent Post-trained Reasoning Across Series), a post-training framework that equips TSLMs with latent reasoning. A model trained with Lapras reasons through a sequence of continuous thoughts in the joint time series-language space, producing text only for the final answer. It learns this through teacher-student self-distillation, where a teacher trained on CoT reference traces reasons explicitly through text. The student aligns its hidden states with the teacher's at the answer stage, transferring the teacher's reasoning ability into its latent computation. We evaluate Lapras across four TSLM backbones on five time series question answering benchmarks. Lapras improves average F1 by up to 10.79% over explicit CoT while generating 23.9x fewer tokens. Lapras's continuous thoughts can also be decoded into readable reasoning traces via standard language decoding, preserving textual explanations. Together, these results highlight Lapras as a promising post-training paradigm for efficient, effective, and interpretable TSLM reasoning.

Abstract

Time Series Reasoning

Explore similar work

Lapras: Latent Reasoning for Time Series Language Models

TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models

From Recognition to Understanding: Unlocking Cognitive Time Series Reasoning with LLMs