GIFT-Eval
PulseAugur coverage of GIFT-Eval — every cluster mentioning GIFT-Eval across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
New HART model combines attention and RWKV for improved time series forecasting
Researchers have developed HART, a novel time series foundation model that combines attention and RWKV blocks to improve long-horizon forecasting. The model addresses the degradation observed in pure RWKV models when pr…
-
New time series foundation models tackle zero-shot forecasting and scale invariance
Researchers have introduced Timer-M1, a novel multivariate time series foundation model designed for zero-shot forecasting by learning temporal and relational primitives. This model utilizes a primitive-based data synth…
-
Aurora-X: New foundation model advances time series forecasting
Researchers have introduced Aurora-X, a large-scale time series foundation model designed to overcome limitations in training potential and architectural versatility. The model employs a progressive curriculum, starting…
-
FlowTSFM introduces novel recurrent transport for time series models
Researchers have introduced FlowTSFM, a novel encoder architecture for time series foundation models that utilizes depth as a recurrent transport process. Instead of multiple independent Transformer layers, FlowTSFM emp…
-
Tabby: Open-Source Time Series Foundation Model Unveiled
Researchers have introduced Tabby, an open-source probabilistic time series foundation model designed for long contexts. Tabby utilizes an encoder-only patch Transformer architecture and was trained on a diverse corpus …
-
IBM releases top-performing Granite Time Series PatchTST-FM-r2 model with open license
IBM has released its new Granite Time Series PatchTST-FM-r2 model, a foundation model designed for general-purpose zero-shot time-series forecasting. This approximately 385 million parameter model offers probabilistic f…
-
Google AI releases TimesFM-3 for multivariate time series forecasting
Google AI has unveiled TimesFM-3, a 330 million parameter foundation model designed for multivariate time series forecasting. Unlike its predecessors, which were limited to univariate predictions, TimesFM-3 can jointly …
-
New SATS method improves time series foundation model pretraining efficiency
Researchers have introduced SATS, a novel pretraining method for time series foundation models designed to handle heterogeneous datasets with varying sampling frequencies. SATS employs a scale-aware token alignment mech…
-
TinyCast: Compact Zero-Shot Forecaster Achieves High Accuracy with Minimal Parameters
A new time series forecasting model named TinyCast has been introduced, featuring a compact design with only 146,505 parameters. This model utilizes a zero-parameter spectral detector to identify periodicity and dilated…
-
TinyCast model achieves state-of-the-art zero-shot forecasting with minimal parameters
Researchers have introduced TinyCast, a novel zero-shot forecasting model that utilizes computed periodicity rather than learning it, making it highly efficient with only 146,505 parameters. This model outperforms exist…
-
New research tackles training and evaluation challenges for time series foundation models
Two new research papers explore challenges in training and evaluating time series foundation models (TSFMs). The first paper introduces ORBIT, a novel training paradigm designed to control distribution properties like d…
-
Google Research releases TimesFM 2.5 for zero-shot time-series forecasting
Google Research has released TimesFM 2.5, an open-source foundation model for time-series forecasting. This model, with 200 million parameters and a context window of up to 16,384 points, can predict future trends witho…
-
TiRex-2 model advances multivariate time series forecasting with recurrent xLSTM design
Researchers have introduced TiRex-2, a novel recurrent foundation model based on xLSTM architecture designed for multivariate time series forecasting. This model addresses limitations of existing Transformer-based appro…
-
FlowState Model Achieves Sampling-Rate-Equivariant Time-Series Forecasting
Researchers have introduced FlowState, a new time-series foundation model designed for enhanced adaptability and efficiency. Unlike previous transformer-based models, FlowState utilizes a state space model encoder paire…
-
New MoE frameworks enhance time series forecasting efficiency and accuracy
Researchers have developed new Mixture-of-Experts (MoE) frameworks for time series forecasting that aim to improve efficiency and accuracy. AME-TS uses structure-guided routing to align expert specialization with tempor…
-
Toto 2.0 models show scaling benefits in time series forecasting
Researchers have introduced Toto 2.0, a suite of five open-weight time series forecasting models that demonstrate the effectiveness of scaling foundation models. The models, trained using a single recipe, show improved …