Rationale-Grounded In-Context Learning for Time Series Reasoning with Multimodal Large Language Models

TL;DR

This work tackles the poor time-series reasoning performance of existing multimodal LLMs by introducing rationale-grounded in-context learning. It proposes RationaleTS, a framework that (i) generates label-conditioned abductive rationales as reasoning priors, (ii) retrieves temporally and semantically similar rationales via a hybrid data-centric and semantic-centric approach, and (iii) conducts rationale-grounded in-context inference to produce outcomes with explicit reasoning. Across finance, transportation, and energy datasets, RationaleTS achieves superior F1 and AUC scores and demonstrates the importance of combining data priors with semantic context in rationale retrieval. The method highlights the value of explicit, transferable reasoning paths to improve reliability, interpretability, and efficiency in multimodal time-series inference, and makes code available for reproduction.

Abstract

The underperformance of existing multimodal large language models for time series reasoning lies in the absence of rationale priors that connect temporal observations to their downstream outcomes, which leads models to rely on superficial pattern matching rather than principled reasoning. We therefore propose the rationale-grounded in-context learning for time series reasoning, where rationales work as guiding reasoning units rather than post-hoc explanations, and develop the RationaleTS method. Specifically, we firstly induce label-conditioned rationales, composed of reasoning paths from observable evidence to the potential outcomes. Then, we design the hybrid retrieval by balancing temporal patterns and semantic contexts to retrieve correlated rationale priors for the final in-context inference on new samples. We conduct extensive experiments to demonstrate the effectiveness and efficiency of our proposed RationaleTS on three-domain time series reasoning tasks. We will release our code for reproduction.

Figures (9)

• Figure 1: Comparison in time series reasoning paradigms with MLLMs (red part) and rationale-grounded in-context learning in RationaleTS (blue part). In MLLMs the prediction outcome is generated by pattern extrapolation, while in RationaleTS, rationales provide reasoning priors connecting observations and implications, for the in-context learning on new samples.
• Figure 2: The workflow of RationaleTS, which includes ➊ Abductive Rationale Generation (§ \ref{['sec_gen']}), ➋ Hybrid Retrieval (§ \ref{['sec_hybrid']}), and ➌ Rationale-Grounded In-Context Inference (§ \ref{['sec_infer']}).
• Figure 3: Left: Time series chart of a sample from Traffic dataset. Right: The generated rationales include 5 reasoning paths. Blue: Observations. Red: Implications. Each reasoning path provides the evidence-grounded analysis on implications to the final outcome.
• Figure 4: Hyperparamter analysis of $K$ and $\lambda$ on Finance ((a)-(d)) and Power ((e)-(h)) datasets.
• Figure 5: Efficiency analysis in terms of AUC, F1 score, and # Input Tokens on Finance and Power datasets.