Latent Dynamics · 2026-10-20

ActTime: Action-Conditioned World Modeling for Time Series

Xinxing Zhou, Xiaojie Yuan, Yanlong Wen, Ming Jin

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

ActTime is an action-conditioned world model for time series that learns action-conditioned dynamics from factual trajectories only. It updates variable-specific latent states via an action-conditioned transition plus observation correction, then rolls the same transition forward under candidate future actions to generate alternative-action trajectories. In simulator-grounded evaluations across four domains, it ranks first in 12 of 16 task settings and shows that factual accuracy, intervention fidelity, and decision quality are related but not interchangeable.

Why it matters

The authors argue that evaluating only factual prediction is insufficient for controlled settings: accurate realized-trajectory prediction does not ensure correct alternative-action effects, and effect prediction does not necessarily yield correct candidate-action rankings. ActTime targets this gap by learning dynamics that can be queried under alternative actions without paired counterfactual supervision, and the evaluation suite measures how performance transfers across factual prediction, intervention fidelity, and decision making.

Method

Learns variable-specific latent states over history using a shared action-conditioned transition and an observation-based state correction; transition effects propagate in latent dynamics.
Uses a frozen pretrained temporal backbone (EIDOS) to form a temporal prior at the history boundary, then rolls forward under candidate future actions with no future observation correction.
Trains only on factual trajectories with a quantile prediction objective; no same-history alternative-action outcomes or planning supervision are used.

Limitation

The evaluation is simulator-grounded; performance in these simulators does not establish transfer to real-world systems. Also, ActTime is trained exclusively on factual trajectories without paired outcomes for alternative actions, and the experiments do not address extrapolation to substantially different action regimes unsupported by the training data.

Abstract (from arXiv)

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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-19 · Indexed by Pub-AI: 2026-10-20 UTC

preprint