TSFM Meets LLM: Context as Covariates

查看原文 / View Original 编辑 / Edit

Computer Science > Machine Learning arXiv:2611.05515 (cs) [Submitted on 3 Nov 2026] Title:TSFM Meets LLM: Context as Covariates Authors:Hemanth Ram, Sai Shankar Narasimhan, Shubhankar Agarwal, Sandeep Chinchali View a PDF of the paper titled TSFM Meets LLM: Context as Covariates, by Hemanth Ram and 3 other authors View PDF HTML (experimental)

计算机科学 > 机器学习 arXiv:2611.05515 (cs) [提交于 2026 年 11 月 3 日] 标题：TSFM 与 LLM 的结合：作为协变量的上下文作者：Hemanth Ram, Sai Shankar Narasimhan, Shubhankar Agarwal, Sandeep Chinchali 查看论文 PDF：TSFM 与 LLM 的结合：作为协变量的上下文，作者 Hemanth Ram 及其他 3 位作者查看 PDF HTML（实验性）

Abstract:Textual information is ubiquitous in real-world time series datasets, appearing as news, reports, and annotations that hold key predictive signals not deducible from numerical history alone. Large Language Models (LLMs) interpret text but lack numerical precision, while Time Series Foundation Models (TSFMs) are quantitatively accurate but cannot reason about the effects of text. Moreover, the lack of large-scale paired multimodal datasets limits progress in bridging this gap.

摘要：文本信息在现实世界的时间序列数据集中无处不在，表现为新闻、报告和注释，其中包含无法仅从数值历史中推断出的关键预测信号。大型语言模型（LLM）能够解读文本但缺乏数值精度，而时间序列基础模型（TSFM）在定量上准确，却无法推理文本的影响。此外，缺乏大规模配对的多模态数据集限制了弥合这一差距的进展。

In this work, we address this challenge by providing a scalable approach for generating realistic multimodal datasets by annotating real-world time series with LLMs. Building on this, we propose Context as Covariates (CoCo), a forecasting framework in which an LLM distills its textual reasoning into numerically grounded forecast and confidence covariates that guide a TSFM backbone, thereby exploiting the strengths of both LLMs and TSFMs.

在这项工作中，我们通过提供一种可扩展的方法来解决这一挑战，即利用 LLM 对现实世界的时间序列进行注释，从而生成逼真的多模态数据集。在此基础上，我们提出了“作为协变量的上下文”（CoCo），这是一个预测框架，其中 LLM 将其文本推理提炼为具有数值基础的预测和置信度协变量，用以指导 TSFM 主干模型，从而发挥 LLM 和 TSFM 的各自优势。

We align the two models in stages, first fine-tuning the LLM via Group Relative Policy Optimization (GRPO) with a reward tied to covariate-induced improvement in forecasting performance, and then fine-tuning the TSFM on covariates generated by the aligned LLM. Our experiments show that models trained on our synthetic multimodal corpus generalize to real benchmarks spanning finance, economics, weather, traffic, and security.

我们分阶段对这两个模型进行对齐，首先通过组相对策略优化（GRPO）微调 LLM，其奖励与协变量带来的预测性能提升挂钩，然后利用对齐后的 LLM 生成的协变量对 TSFM 进行微调。我们的实验表明，在我们的合成多模态语料库上训练的模型可以泛化到涵盖金融、经济、天气、交通和安全等领域的真实基准测试中。

Notably, CoCo with Qwen3-4B-Instruct-2507 as the LLM and Chronos2 as the TSFM achieves up to 12% MAE improvement over state-of-the-art TSFMs and LLM forecasters, such as GPT-4o. Code, checkpoints, and the evaluation datasets are publicly available.

值得注意的是，使用 Qwen3-4B-Instruct-2507 作为 LLM 和 Chronos2 作为 TSFM 的 CoCo 模型，在平均绝对误差（MAE）上比 GPT-4o 等最先进的 TSFM 和 LLM 预测器提升了高达 12%。代码、检查点和评估数据集均已公开。