Time-series question answering (TSQA) is an emerging application of large language models that turns numerical analysis into natural-language queries. However, feeding raw numerical values directly into an LLM is not straightforward. A new arXiv paper (2606.18986v2) addresses this by proposing a method that goes beyond conventional tokenization.
The authors suggest using direct timestep embeddings to represent time-series data, rather than converting numbers into text tokens. They also employ contrastive alignment to better match the time-series representations with the language space of the LLM. The abstract indicates that this approach is intended to overcome the limitations of raw numerical input, though specific experimental results are not included in the available summary.
As the paper is a preprint and the abstract is truncated, the full details of the architecture and evaluation are not yet accessible. The proposal is notable for its focus on embedding design and alignment, rather than relying on prompt engineering or external tools, as a way to make LLMs more capable of handling time-series questions.