arXiv Machine Learning

Latent Inference-Time Guidance of Time Series Foundation Models

arXiv Machine Learning
Oct 9

SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting

SteerCast is a retrieval‑based latent steering technique that enhances decoder‑only time‑series forecasters during inference without updating model parameters. It builds a database from the training set, storing each history window’s representation and a steering vector that captures the difference between the model’s prediction and the ground‑truth continuation in latent space. During testing, SteerCast retrieves nearest neighbors, aggregates their steering vectors, and injects the resulting signal into the forecaster’s hidden states at each autoregressive step, guiding predictions toward trajectories seen in similar training cases. Experiments on diverse multivariate benchmarks and multiple horizons demonstrate consistent accuracy improvements over fine‑tuned backbones and other retrieval‑based baselines, all while requiring no additional training beyond the original fine‑tuning and using only the training set as a retrieval corpus.

By Van Dai Do, Huu Hiep Nguyen, Minh Hoang Nguyen, Hung Le
arXiv Machine Learning
Aug 27

NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates

NVExplain is a model‑agnostic framework that explains time‑series forecasting by attributing each forecast horizon to temporally relevant historical lags. It models forecasting as a latent trajectory, introduces semantic flow to track information evolution, and aggregates this into a lag‑horizon attribution matrix. The method also generates structure‑preserving perturbations and fits sparse local surrogates to produce human‑readable, temporally coherent explanations, and demonstrates competitive faithfulness and stability across benchmark datasets.

By Muyan Anna Li, Manikandan Ravikiran, Aditi Gautam
arXiv Machine Learning
Sep 29

Aurora-X: Built for Extreme Time Series Forecasting

Aurora‑X is a billion‑parameter time‑series foundation model designed for extreme forecasting tasks. It employs a progressive curriculum that starts with channel‑independent pretraining, then adds cross‑variable dependencies, variable context and horizon lengths, and optional future covariates during mid‑training. A variable‑resolution post‑training stage allows adjustable temporal spans per token at inference, while a pattern‑guided mixture‑of‑experts expands capacity through sparse activation and expert specialization. An implicit quantile network head predicts arbitrary quantiles, enhancing probabilistic forecasting flexibility. Experiments on GIFT‑Eval, TIME, FEV‑Bench, TFB, and DAG‑Bench show state‑of‑the‑art performance against both pretrained TSFMs and task‑specific supervised models.

By Xingjian Wu, Chenjuan Guo, Xiangfei Qiu, Zhigang Hu, Hanyin Cheng, Peng Chen, Yang Shu, Jilin Hu, Bin Yang
Hugging Face Trending Papers
Sep 28

Instance-Adaptive Prompts as Context for Time-Series Foundation Models

The paper introduces PaCTS, a method that generates instance‑adaptive latent prompts—continuous embedding tokens—to provide compact contextual information for frozen time‑series foundation models (TSFMs). These prompts are constructed from instance‑specific global statistics and refined with segment‑level temporal data, enabling the model to capture both global characteristics and local temporal variations. Experiments show that PaCTS improves forecasting performance across various context lengths and model architectures, often outperforming the same backbone with double the context while reducing inference computation, and it also offers stronger improvements and better out‑of‑distribution generalization compared to weight‑space adaptation methods.

Hugging Face Trending Papers
Oct 27

Timestral

Timestral is a family of four time‑series foundation models (5M, 28M, 123M, and 209M parameters) that achieve state‑of‑the‑art results on GIFT‑Eval, TIME, and fev‑bench, with the 209M and 123M variants topping every metric. The models use Decoupled Horizon Forecasting to separate prediction tokens from context, a covariate framework for known‑future inputs, and a two‑stage training pipeline that first pre‑trains on synthetic data and then post‑trains on a mix of real and synthetic data with extended context. All four models and inference code are publicly available on Hugging Face and GitHub.