InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement

Abstract

Most existing time series classification methods follow a discriminative modeling paradigm that directly maps input sequences to one-hot encoded class labels. Although effective, this paradigm suffers from two inherent limitations: it is difficult to effectively exploit contextual features, and one-hot label distribution fails to capture the semantic relationships among classes. In this paper, we propose a novel framework, termed InstructTime, which reformulates time series classification as a multimodal generative task. The core idea is to treat continuous numerical sequences, discrete textual features, and task-specific instructions as multimodal inputs, while representing class labels as textual outputs, with classification realized through the generative capability of tuned language models (LMs). To bridge the modality gap, we introduce a time series discretization module that converts continuous numerical sequences into discrete temporal tokens, thereby alleviating the inconsistency between numerical and textual modalities. In addition, we employ an alignment projection layer together with a generative self-supervised pre-training strategy to enhance cross-modal representation alignment. Lastly, we further fine-tune the aligned language model with task-specific instructions to strengthen its multimodal reasoning capability. While InstructTime effectively models explicit contextual features, language models are generally not well-suited to capturing implicit features that are not directly observable from raw time series or contextual features, such as latent temporal and structural patterns. Based on the above analysis, we extend InstructTime to InstructTime++ by introducing an improved multimodal generative framework that additionally captures implicit features to compensate for the limited inductive bias of current LMs-based time series classifiers. In particular, InstructTime++ leverages a collection of specialized toolkits to automatically mine informative implicit patterns from multiple views of raw time series and contextual inputs. In this work, we instantiate this framework from two complementary perspectives by designing statistical feature extraction and vision–language–based image captioning for textual grounding. The discovered implicit features are subsequently translated into textual descriptions, enabling seamless integration with the original InstructTime framework. Finally, extensive experiments on multiple benchmark datasets demonstrate the superior performance of InstructTime++. This work further paves the way for agentic time series classification 1 .