New CADE framework enhances time-series question answering by bypassing tokenization

Researchers have developed CADE (Contrastive Alignment with Direct Embedding), a new framework designed to improve time-series question answering (TSQA) by addressing the limitations of traditional tokenization methods. CADE directly embeds each timestep into the LLM's embedding space using a point-wise linear encoder and MLP projector, preserving crucial magnitude, scale, and trend information lost in standard tokenization. The framework also incorporates a supervised contrastive loss to align time-series embeddings with language representations. Experiments on the Time-MQA benchmark show CADE consistently enhances performance across six TSQA tasks, surpassing existing LLM baselines. AI

IMPACT This research offers a novel approach to processing time-series data for LLMs, potentially improving performance on tasks requiring analysis of sequential numerical data.

RANK_REASON The cluster contains a research paper detailing a novel framework for time-series question answering.

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• Cade
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• multilayer perceptron
• ScienceCast
• Time-MQA

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →