Prompting Underestimates LLM Capability for Time Series Classification
Not on the map. Why?
Abstract (arXiv)
Prompt-based evaluations suggest that large language models (LLMs) perform poorly on time series classification, raising doubts about whether they encode meaningful temporal structure. We show that this conclusion reflects limitations of prompt-based generation rather than the model's representational capacity by directly comparing prompt outputs with linear probes over the same internal representations. While zero-shot prompting performs near chance, linear probes improve average F1 from 0.15-0.26 to 0.61-0.67, often matching or exceeding specialized time series models. Layer-wise analyses further show that class-discriminative time series information emerges in early transformer layers and is amplified by visual and multimodal inputs. Together, these results demonstrate a systematic mismatch between what LLMs internally represent and what prompt-based evaluation reveals, leading current evaluations to underestimate their time series understanding.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 52 of its 89 references that are in the map. First posted on arXiv 6 Jan 2026.
Builds on
52 of 89 references.
References · 52
- 2014Adam: A Method for Stochastic Optimizationcited by 36,960
- 2016SGDR: Stochastic Gradient Descent with Warm Restartscited by 4,993
- 2017Neural Discrete Representation Learningcited by 4,650
- 2020Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecastingcited by 2,141
- 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
- 2021Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecastingcited by 1,654
- 2018An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modelingcited by 1,447
- 2016Understanding intermediate layers using linear classifier probescited by 1,279
- 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
- 2025gpt-oss-120b & gpt-oss-20b Model Cardcited by 1,258
- 2023ImageBind: One Embedding Space To Bind Them Allcited by 1,162
- 2017Mixed Precision Trainingcited by 1,135
38 more
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 831
- 2019On the Variance of the Adaptive Learning Rate and Beyondcited by 763
- 2019Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecastingcited by 748
- 2023One Fits All:Power General Time Series Analysis by Pretrained LMcited by 615
- 2023Large Language Models Are Zero-Shot Time Series Forecasterscited by 491
- 2021TS2Vec: Towards Universal Representation of Time Seriescited by 463
- 2024MOMENT: A Family of Open Time-series Foundation Modelscited by 453
- 2018What you can cram into a single vector: Probing sentence embeddings for linguistic propertiescited by 452
- 2019What do you learn from context? Probing for sentence structure in contextualized word representationscited by 436
- 2019Linguistic Knowledge and Transferability of Contextual Representationscited by 382
- 2023PandaGPT: One Model To Instruction-Follow Them Allcited by 330
- 2018Analysis Methods in Neural Language Processing: A Surveycited by 276
- 2018Scalable and accurate deep learning for electronic health recordscited by 275
- 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 225
- 2015Doctor AI: Predicting Clinical Events via Recurrent Neural Networkscited by 205
- 2016Layer-wise Relevance Propagation for Neural Networks with Local Renormalization Layerscited by 197
- 2017What do Neural Machine Translation Models Learn about Morphology?cited by 165
- 2023TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Seriescited by 165
- 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 138
- 2023LLM4TS: Aligning Pre-Trained LLMs as Data-Efficient Time-Series Forecasterscited by 127
- 2023Large Language Models are Few-Shot Health Learnerscited by 115
- 2024Large Language Models for Time Series: A Surveycited by 113
- 2017Explaining Recurrent Neural Network Predictions in Sentiment Analysiscited by 87
- 2024Language Models Still Struggle to Zero-shot Reason about Time Seriescited by 81
- 2024Aligning Large Language Models with Representation Editing: A Control Perspectivecited by 62
- 2018Universal features of price formation in financial markets: perspectives from Deep Learningcited by 59
- 2023The Wall Street Neophyte: A Zero-Shot Analysis of ChatGPT Over MultiModal Stock Movement Prediction Challengescited by 59
- 2018Evaluating Layers of Representation in Neural Machine Translation on Part-of-Speech and Semantic Tagging Taskscited by 56
- 2024Evaluating Large Language Models on Time Series Feature Understanding: A Comprehensive Taxonomy and Benchmarkcited by 38
- 2023DeWave: Discrete EEG Waves Encoding for Brain Dynamics to Text Translationcited by 36
- 2024A Picture is Worth A Thousand Numbers: Enabling LLMs Reason about Time Series via Visualizationcited by 31
- 2023TENT: Connect Language Models with IoT Sensors for Zero-Shot Activity Recognitioncited by 25
- 2023Multimodal Pretraining of Medical Time Series and Notescited by 22
- 2023Transfer Knowledge from Natural Language to Electrocardiography: Can We Detect Cardiovascular Disease Through Language Models?cited by 21
- 2024Contrastive Instruction Tuningcited by 17
- 2020Analyzing analytical methods: The case of phonology in neural models of spoken languagecited by 16
- 2023Text-to-ECG: 12-Lead Electrocardiogram Synthesis conditioned on Clinical Text Reportscited by 13
- 2025Unlocking Pretrained Vision Transformers for Time Series Classificationcited by 6
50 most cited of 52 All 52 (JSON)
Built on by
Citing papers per year: 2026: 1.
1 paper.