Large Language Models (LLMs) possess an extraordinary capacity for sequential reasoning, yet they were trained on discrete text tokens, not continuous numerical signals. To apply these models to time series forecasting and classification, researchers have moved beyond simple "textualization" (converting numbers to strings like "1.2, 3.4"). Instead, the field is converging on a strategy of cross-modal semantic alignment: mapping numerical patterns directly into the LLM’s latent embedding space.
By treating temporal fluctuations as "pseudo-tokens" that share a mathematical home with words, these frameworks allow frozen LLMs to apply their pre-trained world knowledge to everything from stock market trends to brain waves.
Bridging the Modality Gap
The fundamental challenge in this research is the "modality gap." As noted by Yushan Jiang and Xiyuan Zhang in their surveys, discrete text and continuous time series occupy different representational worlds. Traditional methods often fail because they lose the high-precision nature of numerical data or ignore the rich structural priors that LLMs have already learned about sequences.
The solution, popularized by frameworks like Time-LLM by Ming Jin and colleagues, is to reprogram the time series. This involves two steps:
- Patching: Breaking the continuous signal into discrete sub-series segments (patches).
- Embedding Alignment: Using a lightweight encoder to project these patches into the same high-dimensional space as the LLM’s word embeddings.
Semantic Anchors and Latent Mapping
The core intuition behind this alignment is that temporal patterns (like a "rapid spike" or a "seasonal dip") have semantic equivalents in language. Zijie Pan and Dongjin Song introduced the concept of Semantic Space Informed Prompt Learning (S2IP-LLM), which aligns numerical embeddings with "semantic anchors"—pre-trained word embeddings that represent high-level concepts.
In this setup, a model doesn't just see a vector of numbers; it sees a representation that has been mathematically "pushed" toward the linguistic concept of a trend. Zhe Zhao and Haibin Wen further refined this in STEM-LTS by decomposing signals into trend and seasonal components before aligning them with semantic prompts. This ensures the LLM receives a "structured vocabulary" of the data's dynamics rather than raw, noisy signals.
Why Alignment Works: Transferable Reasoning
Why should a model trained on the internet be good at predicting energy loads or traffic flow? The research suggests that the sequential logic of language—grammar, syntax, and long-range dependencies—is structurally similar to the Markovian properties of time series.
Siming Sun and Qinmin Yang argue in their work on Structure-Guided Cross-Modal Alignment that both modalities are governed by state-transition patterns. By aligning the latent state space of a time series with the transition patterns of natural language, the LLM can treat a numerical sequence as a "pseudo-language" with its own grammar. This allows the model to perform zero-shot forecasting, as seen in the TSLLM-Load Forecasting Mechanism, where the model predicts energy demand in new regions without prior training.
Grounding Signals in Domain Knowledge
Beyond general forecasting, this alignment strategy is proving transformative in specialized fields like medicine and engineering. In these domains, the "text" part of the alignment isn't just general word embeddings, but expert descriptions.
- Clinical Phenotyping: Sam Gijsen and Kerstin Ritter developed EEG-Language Models that align brain waves with physician reports. By using contrastive learning to map signal frequencies to medical terminology, the model learns to "read" an EEG like a neurologist.
- Cardiology: Similarly, the ETP framework by Sibo Cheng aligns ECG signals with clinical text. This grounding allows the model to distinguish between clinically relevant features and background noise, significantly improving diagnostic accuracy in data-scarce scenarios.
Efficiency through Frozen Backbones
A recurring theme across these papers—from AutoTimes by Xiangdong Huang to TimeCMA by Chenxi Liu—is the use of frozen LLMs. Instead of fine-tuning the billions of parameters in a model like LLaMA or GPT, researchers only train the alignment layers (often less than 1% of the total parameters).
This efficiency proves that the "intelligence" required for time series analysis is already present within the LLM's pre-trained weights; the challenge is simply a matter of translation. By aligning the numerical "language" of sensors and markets with the semantic "language" of the model, researchers are unlocking a universal engine for temporal reasoning.