mlsanityData to 14 Oct 2026

Pythia: Toward Foundation World Models for Multimodal Time Series

arXiv 2610.05240 ↗First posted 4 Oct 2026cs.LG cs.AIcited by 0

On the map: World models → Latent predictive world models

Abstract (arXiv)

Time-series foundation models offer a unified approach to forecasting across heterogeneous domains. Textual context and auxiliary observations provide complementary information about temporal dynamics, yet reusable multimodal predictive representations remain underexplored. We introduce Pythia, a foundation world model that learns context-conditioned latent dynamics across datasets through a joint-embedding predictive architecture. A stop-gradient numerical reference guides contextual corrections to predicted future states. A separate probabilistic decoder then adapts to the frozen predictive representation and observed history, decoupling world-model pretraining from observation-space forecasting. On MUSE, Pythia-Tiny's normalized mean absolute scaled error (MASE) and weighted sum quantile loss (WSQL) are 0.6879 and 0.4269, reducing errors by 6.26% and 5.00% relative to the strongest model evaluated in the published MUSE leaderboard. Through a series of controlled experiments, we investigate how to design a time-series world model through shared pretraining and how joint-embedding predictive learning can incorporate multimodal information. The results support separating predictive representation learning from probabilistic readout and show complementary contributions from entity descriptions, events, and covariates.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Latent predictive world models, one of 11 directions in World models. Builds on 56 of its 59 references that are in the map. First posted on arXiv 4 Oct 2026.

Builds on

56 of 59 references.

Closest ancestors · 3

Same direction, newest first

  1. 2026Phys-JEPA: Physics-Informed Latent World Models for Multivariate Time-Series Forecasting15 Jun
  2. 2024Revisiting Feature Prediction for Learning Visual Representations from Video15 Feb
  3. 2023Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture19 Jan

Other references · 53

  1. 2022A Time Series is Worth 64 Words: Long-term Forecasting with Transformerscited by 2,039
  2. 2022Are Transformers Effective for Time Series Forecasting?cited by 1,424
  3. 2022TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysiscited by 1,058
  4. 2021VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learningcited by 1,057
  5. 2025V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planningcited by 809
  6. 2024Chronos: Learning the Language of Time Seriescited by 759
  7. 2023Time-LLM: Time Series Forecasting by Reprogramming Large Language Modelscited by 700
  8. 2023A decoder-only foundation model for time-series forecastingcited by 692
  9. 2023One Fits All:Power General Time Series Analysis by Pretrained LMcited by 632
  10. 2024Unified Training of Universal Time Series Forecasting Transformerscited by 614
  11. 2023Large Language Models Are Zero-Shot Time Series Forecasterscited by 500
  12. 2024TimeMixer: Decomposable Multiscale Mixing for Time Series Forecastingcited by 346
38 more
  1. 2024Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Expertscited by 287
  2. 2026Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Rankingcited by 230
  3. 2024Timer: Generative Pre-trained Transformers Are Large Time Series Modelscited by 208
  4. 2025Chronos-2: From Univariate to Universal Forecastingcited by 194
  5. 2025Sundial: A Family of Highly Capable Time Series Foundation Modelscited by 180
  6. 2024Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Seriescited by 127
  7. 2024CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuningcited by 108
  8. 2024TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignmentcited by 103
  9. 2024ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Datacited by 100
  10. 2024Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Expertscited by 95
  11. 2024Context is Key: A Benchmark for Forecasting with Essential Textual Informationcited by 87
  12. 2024Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysiscited by 74
  13. 2025Moirai 2.0: When Less Is More for Time Series Forecastingcited by 69
  14. 2025This Time is Different: An Observability Perspective on Time Series Foundation Modelscited by 68
  15. 2025VL-JEPA: Joint Embedding Predictive Architecture for Vision-languagecited by 68
  16. 2024AutoTimes: Autoregressive Time Series Forecasters via Large Language Modelscited by 65
  17. 2025Language in the Flow of Time: Time-Series-Paired Texts Weaved into a Unified Temporal Narrativecited by 40
  18. 2024GenRL: Multimodal-foundation world models for generalization in embodied agentscited by 31
  19. 2024How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networkscited by 31
  20. 2026Toto 2.0: Time Series Forecasting Enters the Scaling Eracited by 30
  21. 2025Aurora: Towards Universal Generative Multimodal Time Series Forecastingcited by 29
  22. 2026Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scalingcited by 28
  23. 2024Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learningcited by 20
  24. 2025Joint Embeddings Go Temporalcited by 14
  25. 2025From Samples to Scenarios: A New Paradigm for Probabilistic Forecastingcited by 10
  26. 2025Fidel-TS: A High-Fidelity Multimodal Benchmark for Time Series Forecastingcited by 10
  27. 2026Falcon-X: A Time Series Foundation Model for Heterogeneous Multivariate Modelingcited by 10
  28. 2026Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Modelscited by 9
  29. 2024LaT-PFN: A Joint Embedding Predictive Architecture for In-context Time-series Forecastingcited by 8
  30. 2026Overcoming the Modality Gap in Context-Aided Forecastingcited by 5
  31. 2026From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Spacecited by 5
  32. 2026Unlocking the Value of Text: Event-Driven Reasoning and Multi-Level Alignment for Time Series Forecastingcited by 5
  33. 2026Giving Sensors a Voice: Multimodal JEPA for Semantic Time-Series Embeddingscited by 5
  34. 2025Koopman Invariants as Drivers of Emergent Time-Series Clustering in Joint-Embedding Predictive Architecturescited by 4
  35. 2026SC-JEPA: Stabilizing Latent Predictive Learning for Time-Series Anomaly Predictioncited by 4
  36. 2026Position: Universal Time Series Foundation Models Rest on a Category Errorcited by 4
  37. 2026TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learningcited by 4
  38. 2026Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamicscited by 3

50 most cited of 53 All 53 (JSON)

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026WorldTS: World Modeling for Multimodal Covariate-aware Time Series Forecasting
  2. 2025UniDiff: A Unified Diffusion Framework for Multimodal Time Series Forecasting
  3. 2025The language of time: a language model perspective on time-series foundation models
  4. 2025Conversational Time Series Foundation Models: Towards Explainable and Effective Forecasting
  5. 2025LLM-Integrated Bayesian State Space Models for Multimodal Time-Series Forecasting
  6. 2025Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Representation Alignment
  7. 2025Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting
  8. 2026Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting