Recent studies have explored adapting LLMs to time series tasks by leveraging language reasoning paradigms. However, it remains unclear whether reasoning strategies are effective across time series tasks with diverse temporal structures. In this work, we systematically investigate this practice using five general reasoning strategies across 12 LLMs and 20 public datasets, together with a self-constructed benchmark of approximately 18k QA pairs. Our results show that general reasoning strategies are not universally beneficial: their effectiveness depends on alignment with task-specific temporal decision structures and output spaces, rather than on model scale or complexity. We further observe that prior information and common time series pre-processing techniques exhibit strong task-dependent effects. Based on these findings, we propose three time-aware reasoning strategies that explicitly model temporal regimes, latent dynamics, and output uncertainty. Across four tasks, they consistently outperform generic reasoning strategies, and enable reasoning to better benefit from larger model scales and multi-sampling. Our study highlights the importance of temporal structure awareness in designing reliable reasoning strategies for LLM-based time series analysis.

Large Language Models (LLMs) have recently shown strong potential for time series (TS) analysis by integrating numerical signals with rich text prompts. In this process, although TS are encoded into token sequences and fed into LLMs in a manner similar to text prompts, TS tokens differ from prompt tokens in their underlying structure and semantics, a distinction that remains underexplored in existing LLM-based TS systems. In this work, we investigate both TS tokens and prompt tokens from a new perspective. First, we re-examine TS tokens in the frequency-domain and reveal that token-level information contribution is highly uneven: large groups of tokens share redundant spectral patterns, while a small subset acts as critical frequency anchors. Motivated by this observation, we introduce a frequency-based token merging strategy for TS tokens, which preserves essential information while eliminating redundancy. Second, we uncover a unexplored pyramidal decay effect in prompt tokens, showing that their influence rapidly diminishes with model depth. Building on these findings, we propose TokenDecouple, a unified framework that decouples the compression of TS tokens and prompt tokens. Extensive experiments across diverse TS tasks show that TokenDecouple achieves up to 7.68× inference speedup while delivering performance improvements in 78% of evaluated scenarios, establishing an efficient paradigm for LLM-based TS analysis.