Dual Views Crack Time-Series Reasoning: TSLM Accuracy Jumps Nearly 9 Points

Time-series data is everywhere—from stock movements to sensor readings to medical monitoring signals. Enabling language models to understand and reason about such data is an active area of AI research. But a fundamental question remains unresolved: what information should a time-series language model (TSLM) representation preserve, and which variations should it filter out?

A recent paper titled "Twice Upon a Time: Complementary Representations for Time-Series Reasoning" offers an illuminating answer—don't give the model just one view, but two complementary representations simultaneously.

The Dilemma of a Single View

The mainstream approach in existing TSLMs is to convert time series into a single representation before feeding it into a language model. Common choices include textualization of numerical sequences, screenshots of line plots, or some form of normalized embedding vector. This design implies a fixed inductive bias: the model is assumed to need only one "correct" perspective.

The problem is that different downstream tasks have vastly different information requirements. Forecasting tasks may care more about temporal structure such as trends and periodicity, while question-answering or reasoning tasks may need to know specific values and coordinate references. A single representation cannot simultaneously satisfy these conflicting needs. It is precisely from this point that the paper's authors propose the design principle of "complementary representations."

Two Views, Each with Its Own Role

The model combines two visual views. The first is the traditional line plot, which preserves absolute measurement values and the observational coordinate frame—you can directly read the specific value at a given moment from the chart. The second is the normalized delay embedding, which actively eliminates the effects of positive amplitude scaling and offset, instead exposing the essential patterns of temporal structure.

The complementarity of these two views lies in their opposite invariance characteristics. Line plots are sensitive to numerical changes but not sharp enough for structural changes, while delay embeddings are insensitive to amplitude and offset but excel at capturing periodicity and dynamical features. The paper validates this through interventional experiments during evaluation: the two views indeed exhibit different invariance profiles.

Integrating two views into the same language model requires a specially designed visual encoder. The model injects both representations into a shared language model backbone through visual tokens and multi-level feature fusion. This architectural choice ensures that information from the two views is not simply concatenated early on and lose their respective characteristics.

Delay embedding is a classic method in time-series analysis, with its core idea originating from Takens' embedding theorem: a scalar time series is constructed into a high-dimensional vector through time delays, for example mapping x(t) to (x(t), x(t-τ), x(t-2τ), …), thereby reconstructing the topological structure of the original dynamical system in phase space. This way, even if only one dimension is observed, dynamical features such as periodicity and chaos of the system can be captured. The normalization step (typically subtracting the mean and dividing by the standard deviation, i.e., z-score transformation) eliminates the absolute amplitude and offset of the sequence, making the embedding result insensitive to different scales and zero points. Therefore, the normalized delay embedding highlights the intrinsic shape and dynamic patterns of the time series rather than specific values. The paper chooses this representation precisely because of its complementarity with line plots in terms of invariance: line plots preserve absolute values and coordinates, while delay embeddings strip away amplitude and offset, focusing on structure.

Invariance here refers to the degree of sensitivity of a model representation to a certain type of data transformation. For example, line plot representations are highly sensitive to vertical translation (offset) or scaling (amplitude changes)—when values change, the graph changes accordingly, but the visual features of temporal structure (such as periodicity, trends) may not change much. In contrast, normalized delay embeddings are invariant to amplitude scaling and offset, yet highly sensitive to structural changes such as periodicity and trends. This complementary invariance profile means: line plots excel at answering "what is the specific value at a certain moment," while delay embeddings excel at answering "what pattern does the sequence exhibit." The paper observes changes in model performance through interventional experiments (such as artificially scaling or translating input sequences), validating that the two views indeed carry different invariance properties. This concept is key to understanding why dual views outperform single views.

Experimental Results: More Than Just Perceptual Improvement

The paper builds a dual-view model based on Qwen3-VL-8B and conducts systematic evaluation on TSRBench. Single-view ablation experiments reveal a key finding: the combination of two views is not only beneficial for basic perception tasks, but the gains are particularly significant on forecasting and reasoning tasks.

In terms of specific data, the model using only line plots achieves 44.5% accuracy on TSRBench, while the dual-view model raises this figure to 53.2%, an improvement of nearly 9 percentage points. For comparison, the strongest specialized TSLM baseline evaluated in the paper is only 36.7%, which the dual-view model substantially surpasses.

On other benchmarks, the model also performs well: 89.9% accuracy on TSExam and 86.8% numerical accuracy on the multivariate ChatTS alignment task. These results indicate that complementary representations can not only improve performance on specific tasks but also have the potential for cross-task generalization.

TSRBench is a benchmark suite used in the paper to evaluate time-series reasoning capabilities, typically containing multiple task types such as forecasting, anomaly detection, numerical question answering, and structural reasoning. It aims to comprehensively measure a model's understanding and reasoning level for time series, rather than merely testing perception (such as identifying trends). The single-view line plot model mentioned in the paper achieves 44.5% accuracy on this benchmark, the dual-view model improves it to 53.2%, while the specialized TSLM baseline is only 36.7%. These numbers indicate that the dual-view design has a clear advantage on tasks requiring integrated numerical precision and temporal pattern reasoning. Additionally, ChatTS is a multimodal task/dataset that aligns time series with text, requiring the model to generate or match corresponding time-series values based on natural language descriptions, examining the model's fine-grained numerical understanding of time-series data. The dual-view model achieves 86.8% numerical accuracy on this task, indicating that complementary representations not only benefit reasoning but also improve the generation and alignment precision of specific values.

Why This Approach Deserves Attention

The value of this work lies not only in numerical improvements but also in the generality of the design principle it proposes. Rather than having researchers carefully select a representation for each task, it is better to let the model learn to extract information from multiple perspectives on its own. This approach shares similarities with complementary information fusion in multimodal learning, but its application in the time-series domain is still novel.

The core conclusion of the paper is: complementary representations allow the model to leverage temporal structure without sacrificing access to the original measurement frame. For general-purpose TSLMs that need to handle both numerical precision and temporal pattern reasoning simultaneously, this may be a more pragmatic direction than pursuing a single "optimal representation."

One-line Summary

The paper proposes using two complementary views—line plots and normalized delay embeddings—to enhance the reasoning capability of time-series language models.

Paragraph Summary

This paper addresses the problem that a single representation in time-series language models (TSLMs) struggles to balance numerical precision and temporal structure, proposing the design principle of "complementary representations." The model simultaneously uses line plots (preserving absolute measurement values and coordinate frames) and normalized delay embeddings (eliminating amplitude and offset, highlighting temporal structure), injecting both views into a shared language model backbone through a specialized visual encoder. The dual-view model built on Qwen3-VL-8B improves accuracy on TSRBench from 44.5% to 53.2%, substantially surpassing the specialized baseline of 36.7%; it achieves 89.9% on TSExam and 86.8% numerical accuracy on the ChatTS alignment task. Experiments show that the two views have opposite invariance characteristics, and their combination yields particularly significant gains on forecasting and reasoning tasks. This design principle is general and provides a more pragmatic direction for general-purpose TSLMs than pursuing a single optimal representation.