Teaching a time-series LLM to describe a signal is not the same as teaching it to understand how that signal evolves. Labels, captions, and question-answer pairs push the model toward linguistic representations, but they do not force it to preserve temporal information. Lotus starts from that gap and introduces a different pretraining principle: before adapting the representations to downstream tasks, the model must predict the next segment of the series. The authors call this forecast grounding.

In practice, Lotus divides each time-series channel into non-overlapping patches and projects them into the LLM embedding space. It then asks the model to predict the next patch from the hidden states of the time-series tokens. The constraint is not cosmetic: it forces the representation to retain temporal structure, not only semantic content. To capture both temporal changes and periodic patterns, each raw patch is combined with its spectral representation before being mapped into the embedding space.

The numbers tell part of the story. Across three LLM backbones and six downstream evaluation datasets, Lotus reaches 59.91% average accuracy, 3.36 points above the strongest matched baseline, with minimal trainable parameters. Representation analysis shows the model uses the numerical input and keeps temporal structure in the learned patches. On TSRBench, compact models trained with this principle outperform counterparts 25 to 50 times larger.

The significance goes beyond a benchmark. A good temporal representation can reduce the model size needed for a task. Today, deploying a time-series LLM on-prem often hinges on video memory and inference costs: large models require expensive accelerators or cloud offload. If a pretraining principle lets compact models retain temporal dynamics, the decision weights shift. Training and serving costs fall, and the threshold for keeping data in-house lowers. For industrial telemetry, IoT sensors, or financial series, using smaller models without sacrificing relative accuracy changes the balance between local control and external providers.

There is a second effect on model supply. The race to parameters has rewarded centralized clusters. A pretraining approach that shows gains with minimal trainable parameters makes representation efficiency a competitive axis, not just scale. This does not eliminate large models, but it makes the assumption that progress must come from size less automatic.

Structurally, the signal is that temporal alignment is not solved by adding more labels or invoking larger models. It requires a pretraining objective compatible with the sequential nature of the data. Teams designing fine-tuning pipelines for time-series should think about that predictive constraint, not only about caption quality. For readers evaluating on-prem deployment, AI-RADAR offers analytical frameworks at /llm-onpremise to assess the trade-off between model footprint and data control.