iGEN
Visit IGEN World Explore IGEN Expo
EXPLORE UPGRADE PLANS
BREAKING
Lightweight Distillation of SAM 3 and DINOv3 for Edge-Deployable Livestock Monitoring Varanasi to Host 2-Day Wheat Products Promotion Society CEO's Conclave from July 9 Uncertainty Quality of VGGT: Analysis on DTU Benchmark Dataset Reveals Effective Confidence Threshold for 3D Reconstruction New Diffusion Model Learns Permutation Distributions with Softer, More Tractable Trajectories RidgeCut: Reinforcement Learning Framework Optimizes Logistics Network Partitioning with Rings and Wedges SDS-LoRA: New Low-Rank Adaptation Method Fixes Gradient Distortion in Large Model Fine-Tuning NeuronFabric Architecture Cuts Memory for On-Chip Transformer Training, Promises Efficient Edge AI Kharif Pulses Sowing Off to a Weak Start: Acreage Down 43% as of June 12 Self-Gated Clarification Method Boosts AI Accuracy in Complex Tariff Classification Tyler Framework Boosts LLM Reasoning by Up to 14 Points with Smarter Compute Allocation Lightweight Distillation of SAM 3 and DINOv3 for Edge-Deployable Livestock Monitoring Varanasi to Host 2-Day Wheat Products Promotion Society CEO's Conclave from July 9 Uncertainty Quality of VGGT: Analysis on DTU Benchmark Dataset Reveals Effective Confidence Threshold for 3D Reconstruction New Diffusion Model Learns Permutation Distributions with Softer, More Tractable Trajectories RidgeCut: Reinforcement Learning Framework Optimizes Logistics Network Partitioning with Rings and Wedges SDS-LoRA: New Low-Rank Adaptation Method Fixes Gradient Distortion in Large Model Fine-Tuning NeuronFabric Architecture Cuts Memory for On-Chip Transformer Training, Promises Efficient Edge AI Kharif Pulses Sowing Off to a Weak Start: Acreage Down 43% as of June 12 Self-Gated Clarification Method Boosts AI Accuracy in Complex Tariff Classification Tyler Framework Boosts LLM Reasoning by Up to 14 Points with Smarter Compute Allocation
Home ›› Technology ›› Ai ›› Llms ›› FlowState: New Time-Series Model Handles Any Sampling Rate Without Retraining

FlowState: New Time-Series Model Handles Any Sampling Rate Without Retraining

IBM Research has developed FlowState, a novel time-series foundation model (TSFM) that is sampling-rate-equivariant, meaning it can handle data sampled at different rates without retraining. The model uses a state space encoder and a functional basis decoder to achieve continuous-time modeling, and it outperforms larger models on the GIFT-Eval benchmark while being one of the smallest TSFMs.

iG
iGEN Editorial
June 16, 2026
FlowState: New Time-Series Model Handles Any Sampling Rate Without Retraining

Enterprise time-series forecasting—whether for supply chain demand, logistics throughput, or trade finance cash flow—has long been hampered by a fundamental mismatch: most models are trained on fixed sampling rates and forecast horizons, so when data arrives at a different frequency (e.g., daily vs. hourly) or a different prediction length is needed, the model must be retrained or replaced. This rigidity costs time and compute, and limits the agility that modern global trade demands.

Researchers at IBM Research have proposed a solution called FlowState, described in a preprint on arXiv (2508.05287). According to the paper, authored by Graf, Lars, Ortner, Thomas, Woźniak, Stanisław, and Pantazi, Angeliki, FlowState is a novel time-series foundation model (TSFM) architecture that achieves sampling-rate-equivariant forecasting through a unified design. This means the same trained model can accept input data at any temporal resolution and generate forecasts for any desired horizon without any retraining or fine-tuning.

How FlowState Works

FlowState pairs a state space model (SSM) encoder with a functional basis decoder (FBD). The SSM encoder processes the input sequence in continuous time, avoiding the discretization errors that plague traditional recurrent or transformer models when sampling rates change. The FBD then represents the forecast as a linear combination of basis functions, allowing the model to dynamically adjust the output length. This design enables continuous-time modeling and dynamic time-scale adjustment, so the model inherently generalizes across all possible temporal resolutions.

The authors also introduce an efficient pretraining strategy that improves robustness and accelerates training. Despite being one of the smallest TSFMs in terms of parameter count, FlowState achieves state-of-the-art results on the widely used GIFT-Eval benchmark, a standard evaluation suite for time-series forecasting. Detailed analyses in the paper confirm the effectiveness of each component, and the authors demonstrate FlowState's unique ability to adapt to varying input sampling rates without performance degradation.

Implications for Trade and Supply Chain Forecasting

For enterprises that manage global trade and logistics—where sensor data, customs filings, and trade documentation arrive at irregular intervals—a model that can handle multiple sampling rates natively is a significant step forward. A single FlowState model could replace a portfolio of models each tuned to a specific data frequency, reducing maintenance overhead and enabling more responsive demand sensing and inventory optimization.

Feature Benefit for Enterprise
Sampling-rate equivariance One model works across data frequencies (e.g., daily, weekly, hourly) without retraining
Dynamic forecasting horizon Predict any number of steps ahead without modifying the network
Continuous-time state space encoder Handles missing data and irregular timestamps common in trade flows
Efficient pretraining Faster deployment and lower compute cost vs. existing TSFMs
State-of-the-art on GIFT-Eval Competitive accuracy vs. much larger models

While the paper is a research preprint and not yet a commercial product, the architecture points toward a new class of forecasting engines that could be embedded in trade finance platforms, logistics visibility systems, and supply chain control towers. The ability to adapt to unseen sampling rates is particularly relevant for cross-border trade, where data sources—port scanners, customs databases, IoT trackers—often publish at different rhythms.

Technical Context

FlowState stands in contrast to existing TSFMs such as Lag-Llama, Chronos, and TimesFM, which are typically based on transformer variants. According to the authors, these models "lack adaptability to different sampling rates, struggle with generalization across varying context and target lengths, and are computationally inefficient." FlowState's use of a state space model (SSM) encoder with a functional basis decoder addresses all three shortcomings. The model is one of the smallest TSFMs, yet it achieves top performance on GIFT-Eval, a benchmark that includes datasets with diverse sampling rates and patterns.

The research was conducted at IBM Research (the author affiliations are not explicitly listed but the domain suggests IBM). The paper is available on arXiv under the identifier 2508.05287. No code or dataset has been released yet, but the detailed methodological description allows replication by other teams.

For CTOs and digital transformation leaders in trade and supply chain, the key takeaway is that time-series forecasting is moving toward models that are more flexible by design, reducing the engineering overhead of maintaining multiple bespoke forecasters and enabling real-time adaptation to changing data landscapes.


Sources:

Keep Reading

Recommended Stories

Why Low-Precision Transformer Training Fails: Research Explains Flash Attention Instability Technology

Why Low-Precision Transformer Training Fails: Research Explains Flash Attention Instability

A new paper from researchers Qiu and Yao provides the first mechanistic explanation of why low-precision training with flash attention fails catastrophically. The authors identify two intertwined phenomena—emergent low-rank representations and biased rounding errors—and introduce a minimal modification that stabilizes training.

June 16, 2026
Phys-JEPA Model Promises More Accurate Multivariate Time-Series Forecasting with Physics-Informed Latent States Technology

Phys-JEPA Model Promises More Accurate Multivariate Time-Series Forecasting with Physics-Informed Latent States

Phys-JEPA is a new architecture that imposes physical consistency on latent states rather than only on outputs, improving multivariate time-series forecasting. On standard benchmarks, it reduces mean squared error across multiple horizons, suggesting a promising direction for interpretable temporal world models.

June 16, 2026
RAID: Semantic Graph Diffusion Enables True Cold-Start and Cross-Lingual Forecasting Technology

RAID: Semantic Graph Diffusion Enables True Cold-Start and Cross-Lingual Forecasting

A new framework called RAID (Retrieval-Augmented Iterative Diffusion) addresses the true cold-start forecasting problem where no prior observations exist. By leveraging textual metadata and semantic graph diffusion, RAID outperforms strong foundation models on accuracy and prediction interval coverage while reducing inference latency by an order of magnitude. It also enables zero-shot cross-lingual transfer, allowing models trained in one language to generalize to others.

June 16, 2026
Proximal Policy Optimization Achieves Faster Convergence in Discrete Sampling Research Technology

Proximal Policy Optimization Achieves Faster Convergence in Discrete Sampling Research

A new paper on arXiv explores policy gradient algorithms for training stochastic policies under the Generative Flow Network (GFlowNet) framework. The authors derive equivalents of standard policy gradient algorithms and, for the first time, successfully apply proximal policy optimization (PPO) to GFlowNets, demonstrating improved convergence speed and data efficiency on benchmarks including synthetic energies and molecular graph generation.

June 16, 2026