Tracking the field's attempt to make GPT care about timestamps

What it does

LLM4TS is a curated bibliography that tracks research attempting to apply large language models to time series analysis, alongside native time-series foundation models that skip the NLP lineage entirely. It organizes papers, surveys, datasets, and code repositories into two camps: those fine-tuning GPT-style models on temporal data and those training transformers from scratch on massive time-series corpora. It is a reading list, not a framework.

The interesting bit

The repo’s value is taxonomic. By explicitly separating “TS-LLMs” from standalone foundation models like Chronos and MOMENT, it surfaces the field’s central tension: whether language architectures are genuinely useful for forecasting or just a familiar hammer looking for a new nail.

Key highlights

• Maps the divergence between repurposed LLMs and purpose-built time-series foundation models
• Tracks emerging datasets like LOTSA and Timeseries-PILE aiming for NLP-scale pretraining
• Catalogs models from commercial offerings (TimeGPT) to open research (Lag-Llama, Tiny Time Mixers)
• Includes skeptical surveys explicitly questioning whether LLMs help time-series forecasting
• Links adjacent territories: tabular foundation models, time-series pretraining, and LLMOps

Verdict

Worth bookmarking if you are researching or investing in temporal foundation models and need a curated map of the arXiv flood. Skip it if you are looking for a drop-in forecasting library.