From Consistency to Complementarity: Aligned and Disentangled Multi-modal Learning for Time Series Understanding and Reasoning

TL;DR

This work tackles time series understanding and reasoning (TSUR) by addressing cross-modal misalignment and entanglement between numerical data, visual plots, and textual descriptions. It introduces MADI, a multi-modal LLM framework built on a pre-trained backbone and equipped with Patch-level Alignment (PA), Discrete Disentangled Interaction (DDI), and Critical-token Highlighting (CTH). PA enforces fine-grained numerical–visual–text alignment at the patch level, while DDI separates modality-common and modality-unique semantics via hierarchical vector quantization and cross-attention to enable synergistic fusion; CTH further emphasizes informative signals for robust reasoning. Empirical results on synthetic and real-world TSUR benchmarks show that MADI outperforms general-purpose LLMs and time-series-specific MLLMs on both understanding and reasoning tasks, with ablations confirming the necessity of each module for effective cross-modal integration and generalization.

Abstract

Advances in multi-modal large language models (MLLMs) have inspired time series understanding and reasoning tasks, that enable natural language querying over time series, producing textual analyses of complex temporal dynamics. Recent attempts hybridize numerical time series with their visualized plots, facilitating precise value reasoning and visual structure comprehension for comprehensive time series understanding of MLLMs. However, effective cross-modal integration remains challenging due to fine-grained temporal misalignment across modalities and severe entanglement between shared and modality-specific semantics, which hinder localized interpretation and complementary reasoning. To address these issues, we propose MADI, a multi-modal LLM enhanced with fine-grained alignment and disentangled interaction, featuring (1) Patch-level Alignment, which enforces physically grounded fine-grained correspondence across heterogeneous modalities, (2) Discrete Disentangled Interaction, which separates modality-common semantics into compact discrete latents and adaptively synergizes the purified modality-unique information, and (3) Critical-token Highlighting, which emphasizes informative, query-relevant signals for robust reasoning. Experiments on synthetic and real-world benchmarks show that MADI consistently outperforms general-purpose LLMs and time-series-specialized MLLMs.

Figures (13)

• Figure 1: Comparison of multi-modal architectures.
• Figure 2: The model architecture of MADI.
• Figure 3: Ablation results of MADI.
• Figure 4: Heatmaps of cross-modal patch embedding similarities.
• Figure 5: Distributions of cross-modal embedding similarities.