CGTime decouples perception to beat GPT-4o-mini on time-series analysis

Training multimodal models to understand time series with language has a fundamental flaw: the usual approach asks an LLM to describe a series, but label quality is capped by the very perception the model is supposed to learn. This self-supervision trap gets worse with multivariate data, where cross-channel correlations and lead-lag structures are most important yet hardest for LLMs to capture. Researchers from Xi'an Jiaotong University and collaborators address this with a new model called CGTime, which decouples perception from description. Deterministic code computes statistical features from real open-source multivariate series, while an LLM only verbalizes those precomputed facts. This lets the system handle perception with computation and expression with language, avoiding the labeling bottleneck entirely.

CGTime, at just 4B parameters, outperforms far larger general-purpose models on multivariate understanding tasks. It achieves a multivariate fact score of 0.283 on their held-out benchmark, compared to 0.173 for GPT-4o-mini and 0.203 for GPT-5.4-nano, with the gap surviving Holm-corrected paired significance tests. The model also generates captions with more accurate verifiable numerical facts and covers a broader range of statistical properties. The paper is submitted to AAAI 2027, and the approach could make specialized time-series AI far more reliable and cost-efficient than scaling general models.