Live intelligence

Latest AI updates

3 updates
Research

FlexCast: Flexible Context Learning for Time Series Foundation Models

FlexCast introduces a flexible context-target learning framework for time series foundation models that reformulates forecasting as variable-length masked prediction over time series patches. During pretraining, it dynamically samples visible context and masked target patches to expose the model to diverse context-target configurations without sequential decoding, while inference supports one-step forecasting, confidence-based iterative refinement, and next-token-style generation through a unified mask-and-predict interface. The authors report accuracy competitive with or better than strong TSFM baselines on BLAST and multiple zero-shot benchmarks while substantially reducing inference latency, and ablations indicate that flexible training context matters more than left-to-right decoding.

Time Series Foundation ModelsForecastingMasked Prediction
Research

Theoretical Analysis of Axial Attention in Time Series Foundation Models

The authors extend the Single Location Regression framework from Marion et al. [2025] to 2D structured data in order to study sparse-information settings where a target depends on only a small number of tokens. They formulate an analytically tractable axial attention-type predictor and analyze its statistical properties and training dynamics. Given oracle parameters in a specific asymptotic regime, the predictor is proven to achieve Bayes optimality, whereas a class of semi-linear simplifications provably fails, highlighting the role of the inner non-linearity. They also prove that Projected Gradient Descent converges globally to the optimal oracle parameters under a specific temperature scaling, and numerical experiments emphasize the importance of inverse temperature scheduling for convergence in practice.

time series foundation modelsaxial attentionstatistical learning theory
Models Editor selected

Mistral Releases Timestral, a CPU-Friendly Time Series Foundation Model Family

Mistral AI has released Timestral, a family of four zero-shot time series forecasting models ranging from 5M to 209M parameters, with open weights and inference code. The 209M checkpoint is reported by Mistral to rank first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, though several margins are narrow. The models support long context, 21 quantiles per step and covariate-aware attention, and run inference on standard server CPUs without a GPU. Weights are released under a M

Why it matters: It gives developers a concrete, openly weighted forecasting model family with CPU-only inference and benchmark comparisons against TimesFM-3, Chronos-2 and Toto 2.0.

Mistral AItime series foundation modelsGIFT-Eval