Short papers
Up to 4 pages (excluding references and appendices); includes complete results, work in progress, position papers, preliminary findings, negative results, replication studies, critical analyses, and deployment experience reports.
Machine Learning/CORE Unranked
FMTS 2026
NeurIPS 2026 Workshop: Foundation Models for Temporal Systems From Forecasting to World Modeling
FMTS 2026 is a NeurIPS workshop focused on foundation models for temporal systems, bringing together research on forecasting, simulation, multimodal temporal data, and reliable decision-making in evolving real-world systems. The workshop emphasizes four interconnected research axes—tasks, data and environments, models, and evaluation—to advance temporal world modeling beyond traditional forecasting.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Datasets, benchmarks, and simulators
Treated as primary research artifacts; submissions should include contamination audits and deduplication logs where applicable.
Research areas in scope
01
Forecasting and simulation tasks
Long-horizon and multi-resolution forecastingMultimodal contextual forecasting conditioned on text, video, events, covariates, or actionsCalibrated probabilistic forecastingPrediction under sparse observations, regime changes, and distribution shiftScalable trajectory simulationAdaptive, retrieval-augmented, or agentic forecasting systemsModel selection, ensembling, tool use, planning, and inference-time orchestration
02
Temporal data and environments
Large-scale temporal pretraining corporaEnvironments combining time series with text, video, graphs, sensors, trajectories, actions, and eventsVideo-as-environment corporaPhysical evaluation suitesSynthetic temporal dataSimulation environmentsBenchmark realismData qualityScalable evaluation
03
Temporal models
Time-aware models for irregular samplingEvent-based models for marked point processes and transaction streamsHierarchical multi-timescale and state-space architecturesGenerative temporal modelsTime-series foundation modelsScalable pretrainingMultimodal fusionMemory and persistent stateEmbodied vision-language-action systemsAdaptation across domains, modalities, and horizons
04
Evaluation and reliability
Benchmark realism, data quality, and leakage-aware evaluationRobustness under distribution shift and test-time adaptationCalibration and conformal uncertainty quantificationLong-horizon and simulation consistencyNeural scaling lawsReproducibility standardsContamination audits for temporal datasets and pretraining corpora
Policies worth checking twice
- Submissions are limited to 4 pages (excluding references and appendices); over-length submissions will be desk-rejected.
- Review is double-blind; submissions must not include author names, affiliations, acknowledgements, or self-identifying links.
- Submissions must be anonymized; failure to comply results in desk rejection.
- All papers are non-archival and do not appear in proceedings; acceptance does not preclude future archival publication.
- Concurrent submission to other venues is permitted, but previously published work (e.g., at ML conferences or NeurIPS main conference) is ineligible.
- Preprints (e.g., arXiv) are allowed, but the submitted PDF must remain anonymized.
- LLMs and AI assistants cannot be authors; important or non-standard use must be documented; routine editing assistance does not require disclosure.
- Prompt injection and attempts to manipulate reviewing are prohibited.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026