Decoupling Perception from Description: Computation-Grounded Representation Alignment between Multivariate Time Series and Language
Summary
This paper introduces CGTime, a 4B-parameter computation-grounded time-series-language model that decouples perception from description using deterministic statistics and LLM verbalization, outperforming larger general-purpose models on multivariate understanding tasks.
View Cached Full Text
Cached at: 08/07/26, 07:48 AM
# Decoupling Perception from Description: Computation-Grounded Representation Alignment between Multivariate Time Series and Language Source: [https://arxiv.org/abs/2608.05238](https://arxiv.org/abs/2608.05238) [View PDF](https://arxiv.org/pdf/2608.05238) > Abstract:Training multimodal models to align time series with language runs into a self\-supervision trap\. The usual recipe asks an LLM to read a series and write a description, so label quality is capped by the perceptual skill the model is supposed to learn\. The data can never teach more than the labeler already knows\. A second gap makes this worse: most datasets use a single variable, but the patterns that matter \(cross\-channel correlation, lead\-lag structure, co\-occurring anomalies\) appear only with several variables, right where the labeling LLM's limits are most exposed\. These two problems create a trilemma: existing methods are reliable, realistic, or scalable, but none achieves all three\. We resolve this by decoupling perception from description\. Deterministic code computes a set of statistics from real, open\-source multivariate series; the LLM verbalizes those precomputed facts\. Perception, which LLMs do poorly, is handled by computation, while the LLM handles expression\. This produces CGTime, our 4B\-parameter computation\-grounded time\-series\-language model\. CGTime outperforms far larger general\-purpose models on multivariate understanding tasks: it attains the best multivariate fact score on our held\-out benchmark \(0\.283 vs\. 0\.173 for GPT\-4o\-mini and 0\.203 for GPT\-5\.4\-nano\), a gap that survives Holm\-corrected paired significance tests against every baseline\. It also states verifiable numerical facts in generated captions more accurately and covers a broader range of statistical properties\. ## Submission history From: Xinran Feng \[[view email](https://arxiv.org/show-email/7e231c08/2608.05238)\] **\[v1\]**Wed, 5 Aug 2026 13:57:08 UTC \(6,665 KB\)
Similar Articles
Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding
Chronicle is a 324M-parameter decoder-only transformer pretrained from scratch on both natural language and time series, achieving competitive performance on NLU and time series classification tasks, and setting new state-of-the-art for frozen-embedding time series classification on UCR/UEA datasets.
TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models
TimeThink introduces a synthetic framework to enhance compositional reasoning in timeseries large language models via reinforcement learning with verifiable rewards, showing significant improvements over baselines on synthetic and real-world tasks.
Time-Series Retrieval for Grounding Multimodal Language Models in Remaining Useful Life
This paper proposes a framework for remaining useful life prediction using multimodal large language models grounded by time-series retrieval, demonstrating that retrieval-based methods improve accuracy and stability over non-retrieval baselines.
ConceptTS: LLM-Guided Concept Bottlenecks for Interpretable Multivariate Time-Series Forecasting
ConceptTS introduces an interpretable forecasting framework that uses large language models to propose human-readable concepts for multivariate time-series prediction, achieving competitive accuracy with transparency through concept bottlenecks.
GigaChat Audio: Time-aware Large Audio Language Model
This paper introduces GigaChat Audio, a time-aware large audio language model that answers questions with explicit timestamps for up to 120 minutes of audio, using interleaved periodic time markers and synthetic supervision. The model achieves strong temporal grounding accuracy on benchmarks and the authors release model weights and datasets.