Beyond Tokenization: Direct Timestep Embedding and Contrastive Alignment for Time-Series Question Answering

The paper introduces CADE (Contrastive Alignment with Direct Embedding), a novel framework for time-series question answering (TSQA) that addresses the limitations of traditional tokenization methods in large language models (LLMs). CADE utilizes direct timestep embedding through a point-wise linear encoder and MLP projector to maintain index-level access without patching, while a one-directional supervised contrastive loss aligns time-series embeddings with textual class-name anchors. Experimental results on the Time-MQA benchmark show significant performance improvements over existing LLM baselines, making CADE a promising approach for practitioners in TSQA applications.