Beyond Uniform Tokens: Adaptive Compression for Time Series Language Models

The paper presents a novel adaptive token budgeting framework for time series language models (TS LLMs) that addresses the inefficiencies of uniform token processing by recognizing the distinct information structures of TS and prompt tokens. It introduces a method that compresses TS tokens based on their frequency-domain contributions while progressively reducing prompt tokens across model layers. Experimental results indicate up to 7.68× inference acceleration and performance improvements in 78% of tested scenarios, highlighting the framework's potential for enhancing scalability in TS foundation models.