VITRO: Vocabulary Inversion for Time-series Representation Optimization

Filippos Bellos, Nam H. Nguyen, Jason J. Corso

Abstract

Large Language Models (LLMs) possess significant reasoning capabilities but struggle with the continuous, numerical nature of time-series data due to their discrete, text-based vocabularies. We present VITRO (Vocabulary Inversion for Time-series Representation Optimization), a framework that introduces "pseudo-word" vocabularies to bridge this gap. By learning dataset-specific embeddings while keeping the underlying LLM frozen, VITRO enables state-of-the-art long-term forecasting performance. Qualitative analysis reveals that VITRO learns structured temporal manifolds, outperforming traditional text-tethered reprogramming methods across multiple benchmarks.

Introduction

While LLMs excel at symbolic reasoning, they face a "language barrier" when applied to time-series forecasting. Standard vocabularies are designed for text, not the nuanced temporal dynamics or seasonal patterns of numerical data. Existing methods often attempt to "reprogram" time-series patches into text-like prototypes, but these remain constrained by the original LLM vocabulary. VITRO addresses this by asking whether a custom-built vocabulary, optimized specifically for time-series, can better unlock the latent power of frozen LLMs.

Methodology: Vocabulary Inversion

VITRO is inspired by "textual inversion" techniques from the image generation domain. The framework operates in two primary stages:

  1. Vocabulary Inversion: Instead of mapping data to existing tokens, VITRO learns a set of "pseudo-words"—new embedding vectors ($v_i$) representing specific instances and a shared "domain" embedding ($s$) for the entire dataset. These are optimized iteratively to minimize forecasting error.
  2. Forecasting Integration: The learned vocabulary is integrated into the LLM pipeline. Two variants are explored: VITRO-Sim, which uses similarity-based selection of core lexicon embeddings, and VITRO-TimeLLM, an attention-based approach.
The VITRO architecture showing the integration of numerical patches and learned pseudo-word embeddings into a frozen LLM.

The pipeline includes data normalization via RevIN, patching, and projecting these patches into the LLM's embedding space alongside the optimized pseudo-words.

Key Results

VITRO was evaluated on seven public benchmarks, including Weather, Electricity, Traffic, and the ETT datasets. The results demonstrate that replacing standard text vocabularies with VITRO’s learned embeddings consistently improves performance.

Comparison of VITRO-enhanced methods against state-of-the-art forecasting models across multiple datasets and metrics.

As shown in the table, VITRO-enhanced methods achieve the lowest MSE and MAE in most categories. For instance, on the ETTh1 dataset, VITRO-Sim achieved an MSE of 0.412, significantly outperforming the standard Time-LLM (0.437) and PatchTST (0.444). These gains were consistent across different LLM backbones like GPT-2 and Llama-7B.

Analysis and Insights

Qualitative analysis using PCA and t-SNE reveals that VITRO’s learned vocabulary forms a structured, U-shaped manifold, suggesting it captures the underlying geometric structure of temporal data far better than the diffuse, circular clouds formed by standard text embeddings. Furthermore, attention weight heatmaps show "horizontal striping," indicating that specific pseudo-words act as specialized temporal markers, consistently anchoring to certain parts of the input sequence. This confirms that the "knowledge" inside the LLM is already sufficient; it simply requires a better way to be prompted.

Conclusion

VITRO demonstrates that the inherent knowledge within frozen LLMs is sufficient for high-performance time-series forecasting, provided the input representation is optimized. By moving away from text-based tokens toward learned pseudo-words, VITRO achieves state-of-the-art results without the need for expensive model fine-tuning. Future work will explore the application of these learned vocabularies to other tasks such as anomaly detection and classification.

Summary updated

BibTeX
@inproceedings{bellos2025vitro,
  author = {Filippos Bellos and Nam H. Nguyen and Jason J. Corso},
  title = {{VITRO: Vocabulary Inversion for Time-series Representation Optimization}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10889449},
}
Figure