The first step is the hardest: pitfalls of representing and tokenizing temporal data for large language models

Large language models (LLMs) have demonstrated remarkable generalization and across diverse tasks, leading individuals to increasingly use them as personal assistants due to their emerging reasoning capabilities. Nevertheless, a notable obstacle emerges when including numerical/temporal data into these prompts, such as data sourced from wearables or electronic health records. LLMs employ tokenizers in their input that break down text into smaller units. However, tokenizers are not designed to represent numerical values and might struggle to understand repetitive patterns and context, treating consecutive values as separate tokens and disregarding their temporal relationships. This article discusses the challenges of representing and tokenizing temporal data. It argues that naively passing timeseries to LLMs can be ineffective due to the modality gap between numbers and text.

We conduct a case study by tokenizing a sample mobile sensing dataset using the OpenAI tokenizer. We also review recent works that feed timeseries data into LLMs for human-centric tasks, outlining common experimental setups like zero-shot prompting and few-shot learning.

The case study shows that popular LLMs split timestamps and sensor values into multiple nonmeaningful tokens, indicating they struggle with temporal data. We find that preliminary works rely heavily on prompt engineering and timeseries aggregation to "ground" LLMs, hinting that the "modality gap" hampers progress. The literature was critically analyzed through the lens of models optimizing for expressiveness versus parameter efficiency. On one end of the spectrum, training large domain-specific models from scratch is expressive but not parameter-efficient. On the other end, zero-shot prompting of LLMs is parameter-efficient but lacks expressiveness for temporal data.

We argue tokenizers are not optimized for numerical data, while the scarcity of timeseries examples in training corpora exacerbates difficulties. We advocate balancing model expressiveness and computational efficiency when integrating temporal data. Prompt tuning, model grafting, and improved tokenizers are highlighted as promising directions.

We underscore that despite promising capabilities, LLMs cannot meaningfully process temporal data unless the input representation is addressed. We argue that this paradigm shift in how we leverage pretrained models will particularly affect the area of biomedical signals, given the lack of modality-specific foundation models.