CVC: Further aligning LLMs via cross-view correction for time series forecasting

Open Webpage

Feifei Li, Haote Xu, Haodi Xu, Yinhao Liu, Xinghao Ding

Published: 01 Sept 2025, Last Modified: 05 Nov 2025Knowledge-Based SystemsEveryoneRevisionsCC BY-SA 4.0
Abstract: Highlights•We introduce Cross-View Correction, a novel framework for aligning time-series and textual embeddings in LLM-based time series forecasting.•Cross-attention match module integrates a wavelet-decomposed signals with LLM embeddings to inject rich semantic context into temporal representations.•Graph-based prompt learning module dynamically selects window-specific textual prompts via a sparse similarity graph and GCN refinement.•Contrastive correction applies layer-wise and output-level losses to remove redundancy and further improve cross-modal alignment.•Extensive benchmarks demonstrate that CVC achieves improvements in MAE/MSE over state-of-the-art baselines.