TimeWorld: Towards Action-Conditioned Multimodal World Models for Time Series Forecasting

Hang Ni; Weijia Zhang; Hao Liu

cs.LGOct 26, 2026

TimeWorld: Towards Action-Conditioned Multimodal World Models for Time Series Forecasting

Authors: Hang Ni, Weijia Zhang, Hao Liu

Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Abstract

Time series forecasting supports decision-making in web applications such as online retail and network management, where planned actions can reshape future dynamics. World models provide a natural paradigm for predicting such action-dependent changes beyond extrapolating historical patterns. However, existing time series world models (TSWMs) typically rely on fixed, domain-specific covariate schemas that limit flexible intervention specification, while multimodal forecasters treat text primarily as passive predictive context. To investigate these gaps, we introduce TSWM-Bench, a cross-domain benchmark for multimodal TSWMs spanning seven real-world datasets, with textual interventions grounded in recorded action--outcome data. Our evaluation reveals that conventional multimodal forecasters underperform history-only time series foundation models (TSFMs) despite dataset-specific training, whereas simple text-prefix adaptations improve forecasting across all evaluated TSFM backbones. These findings suggest that pretrained TSFMs encode implicit world knowledge of temporal dynamics, offering a strong foundation for action-conditioned modeling beyond conventional multimodal fusion. Building on this insight, we propose TimeWorld, a plug-in framework that adapts pretrained TSFMs through explicit action-response modeling. Specifically, the Prototype-Guided Latent Modulation (PLM) module grounds textual actions in temporal dynamics by composing representative response prototypes to modulate future representations based on the action and historical patterns. The Action Contrastive Learning (ACL) component strengthens action-specific response modeling by contrasting these updates under alternative interventions sharing the same history, distinguishing action types while preserving relative intensity relationships.

Figures & tables

Study	Action- Conditioned	Action Format	Real-world Action	Multi- domain
Multimodal TSF
Time-MMD ( Liu et al., 2024a )	✗	–	–	✔
CiK ( Williams et al., 2025 )	✗	–	–	✔
Structured TSWM
MobiWM ( Qi et al., 2026 )	✔	Covariates	✗	✗
CEDAR ( Meng et al., 2026 )	✔	Covariates	✔	✗

Table 1. Comparison of benchmarks.

Model	nMAE $\downarrow$	nRMSE $\downarrow$	MASE $\downarrow$	Rank $\downarrow$
TimesFM 3.0	1.0540	1.2087	3.5023	8.33
Chronos-2	1.0237	1.1695	3.4948	5.00
Toto 2.0	1.0257	1.1730	3.4923	5.33
Moirai 2.0	1.0329	1.1759	3.4996	7.00
Time-LLM	1.4213	1.5774	4.6371	12.67
MM-TSFlib	1.1301	1.2530	3.8602	11.00

Table 2. Averaged forecasting performance across seven datasets. Rank is the mean rank over the three metrics. The best and second-best results in each column are shown in red and blue , respectively.

Figure 1. Overview of TimeWorld.

Dataset	Metric	TimesFM 3.0			Chronos-2			Toto 2.0			Moirai 2.0
Dataset	Metric	Vanilla	TSFM++	TimeWorld	Vanilla	TSFM++	TimeWorld	Vanilla	TSFM++	TimeWorld	Vanilla	TSFM++	TimeWorld
M5	nMAE	1.1612	0.9658	0.9803	0.8784	1.0204	0.9829	0.8705	0.9255	0.9917	0.8602	0.9788	1.0219
	nRMSE	1.7597	1.3084	1.2997	1.4231	1.2914	1.2991	1.4190	1.3742	1.3102	1.4071	1.3476	1.3002
	MASE	2.4538	2.0766	2.1151	1.9042	2.2255	2.1294	1.8886	1.9649	2.0875	1.8606	2.1473	2.1837
MOF	nMAE	0.7793	0.7787	0.8153	0.8019	0.7569	0.7347	0.7673	0.7908	0.7638	0.7235	0.7414	0.7527
	nRMSE	0.8066	0.8012	0.8402	0.8286	0.7845	0.7604	0.7958	0.8178	0.7923	0.7488	0.7677	0.7772

Table 3. Comparison results of four TSFM backbones. Vanilla performs zero-shot forecasting by TSFMs, and Overall averages the 13 settings across seven datasets. The best and second-best results of each metric are shown in red and blue , respectively. Underlining marks the best result for each TSFM backbone.

Figure 2. Ablation results of TimeWorld.

Panel of Figure 3 — Figure 3. Parameter sensitivity of TimeWorld across four TSFM backbones, measured by nMAE.

Model	Action Type		Action Intensity
Model	Same	Different	Adjacent	Distant
TimeWorld	0.1077	0.0533	0.9815	0.9380
w/o ACL	0.0293	0.0271	0.9311	0.9284

Table 4. Action-conditioned representation similarity.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$(L,H)$	Train	Val.	Test
M5	$(112,7)$	438	147	147
M5	$(112,28)$	437	142	147
MOF	$(60,1)$	57	15	21
MOF	$(60,5)$	57	15	21
FOMC	$(60,1)$	117	39	40
FOMC	$(60,5)$	116	39	40

Appendix

Table 5. History–horizon configurations and split sizes.

Method	nMAE $\downarrow$	nRMSE $\downarrow$	MASE $\downarrow$
Vanilla	1.0540	1.2087	3.5023
Covariates (zero-shot)	1.1097	1.2519	3.9601
Covariates (full tuning)	1.0207	1.1340	3.4507
Covariates (head + LoRA)	1.1035	1.2438	3.9267
TSFM++	0.9983	1.1116	3.3634

Appendix

Table 6. Comparison with covariate action conditioning.

Method	nMAE $\downarrow$	nRMSE $\downarrow$	MASE $\downarrow$
TimeWorld (frozen TSFM)	0.8997	1.0185	3.0651
+ Full tuning	0.9082	1.0224	3.0916
+ Head and LoRA	0.9003	1.0188	3.0682

Appendix

Table 7. Performance against variants that fine-tune TSFMs.

Method	TimesFM	Chronos	Toto	Moirai
TSFM++	$[-0.227,$ $-0.010]$	$[-0.099,$ $-0.001]$	$[-0.034,$ $0.008]$	$[-0.089,$ $-0.014]$
TimeWorld	$[-0.427,$ $-0.028]$	$[-0.124,$ $-0.044]$	$[-0.246,$ $-0.027]$	$[-0.243,$ $-0.031]$

Appendix

Table 8. 95% confidence intervals (CIs) for

\Delta\mathrm{nRMSE}

(method minus vanilla). Negative values indicate improvement.

Method	nMAE $\downarrow$	nRMSE $\downarrow$	MASE $\downarrow$
TimeWorld	0.8997	1.0185	3.0651
Random reassignment	0.9215	1.0438	3.1586

Appendix

Table 9. Performance against prototype composition variant.

Figure 4. Forecasting example on HUPA-UCM with Toto.

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 861 papers

May 28, 2026

Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting

Time-Series Foundation Models (TSFMs) excel at zero-shot unimodal forecasting using numerical data, but unlike LLMs they cannot consume multimodal, non-numerical context that often shape real-world trajectories. In this work, we bridge this gap and argue for a multimodal time-series forecasting approach that post-trains LLMs to act as context-guided revisors over strong numerical TSFM priors. We introduce PostTime, a post-training recipe combining Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR), along with a methodology to generate automated reasoning traces for forecast revisions. PostTime teaches an LLM to generate context-conditioned forecast interventions -- decisions to revise, preserve, or ignore the TSFM prior based on the multimodal context. We evaluate this approach on the TimesX multimodal forecasting benchmark using a Gemma-3-4B LLM and TimesFM-2.5 TSFM, and show that it significantly outperforms standalone TSFMs, LLM-only baselines, and existing multimodal forecasting approaches.

Haoxin Liu, Yichen Zhou, Rajat Sen +2

Georgia Institute of Technology · Google Research

Multivariate Time Series Forecasting LLM Post-Training Time-Series Foundation Models

Ask PDF

Oct 19, 2026

ActTime: Action-Conditioned World Modeling for Time Series

Time-series models increasingly condition on known future inputs. In controlled systems, some of these inputs are actions that can be changed, compared, and optimized before execution. A model must therefore do more than predict the realized future: it must predict how outcomes change under different actions from the same observed history and support reliable comparison of candidate plans. We introduce ActTime, an action-conditioned world model for time series. Over the observed history, ActTime updates variable-specific latent states through an action-conditioned transition and corrects them using observations. It then rolls the same transition forward under candidate actions to generate future trajectories. Training uses only factual trajectories, without paired alternative-action outcomes or planning supervision. We construct a simulator-grounded evaluation suite across building energy, traffic, greenhouse climate, and inventory management, covering factual prediction, intervention-effect prediction, offline decision making, and closed-loop planning. ActTime ranks first in 12 of 16 domain-task settings. Across eight methods, the mean pairwise model-ranking reversal rate is 17.8% between factual prediction and intervention fidelity, and rises to 28.4% between intervention fidelity and decision ranking. These results show that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable.

Xinxing Zhou, Xiaojie Yuan, Yanlong Wen +1

Nankai University · Griffith University

Action-Conditioned World Models Model-Based Planning Multivariate Time Series Forecasting Sequential Decision Making

Ask PDF

Sep 25, 2026

WorldTS: World Modeling for Multimodal Covariate-aware Time Series Forecasting

Time series forecasting is typically framed as learning a direct mapping from historical to future observations in the observation space. However, sequences of observations generally provide only a partial view of the dynamics of the underlying system, with future observations being shaped by latent dynamics. Recent latent-space forecasting methods thus achieve improved performance by predicting future observations from latent-space representations of historical observations rather than directly forecasting future observations in the observation space. Next, while future observations are also shaped by external factors, how to incorporate external, often multimodal, information into forecasting, so that it can shape latent-state formation and evolution directly, remains underexplored. We propose WorldTS, a world-modeling based forecasting framework that integrates multimodal covariates directly into the forecasting to further improve forecasting performance. Specifically, WorldTS employs a two-stage training strategy. First, it learns forecasting-relevant latent state dynamics conditioned on multimodal covariates, yielding encoded future states. Next, the learned state dynamics are frozen, and an observation decoder is trained to map the predicted future states back to future observations. Extensive experiments on 21 real-world datasets offer insight into WorldTS and its effectiveness.

Yuhan Zhu, Xiangfei Qiu, Hanyin Cheng +5

East China Normal University · Aalborg University

Time Series Forecasting with Covariates Time Series Forecasting Multivariate Time Series Forecasting World Model Learning

Ask PDF

Abstract

Time Series Forecasting

Explore similar work

Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting

ActTime: Action-Conditioned World Modeling for Time Series

WorldTS: World Modeling for Multimodal Covariate-aware Time Series Forecasting