Time-IMM is a comprehensive benchmark and open-source library designed for irregular, multimodal time series. It introduces nine real-world datasets across diverse domains and provides tools for fusing asynchronous text with numeric signals, showing that multimodal integration can significantly enhance forecasting in complex, real-world settings.

LLM4TS is a framework that adapts pre-trained Large Language Models for multivariate time series forecasting through a two-stage fine-tuning process. It captures multi-scale temporal patterns and achieves state-of-the-art performance across full-shot and few-shot settings.

This survey systematizes the training recipes, infrastructure, and environments of Agentic Reinforcement Learning to provide a unified framework for building robust, generalist autonomous agents from scratch, addressing the shortcomings of standard alignment methods like RLHF.

Defines time-series reasoning through a two-level taxonomy (reasoning topology × primary objectives) and a compact attribute tag set (e.g., decomposition, verification, tool use, multimodality, alignment), while curating research and non-research works with guidance on evaluation and deployment.

Perseus is an interactive framework for multi-granularity time series segmentation under sparse supervision. Each user prompt is encoded with its local context into a persistent memory bank, so predictions condition on all prompts accumulated across iterations rather than only their local neighborhood. Across six wearable-sensing and industrial datasets, it improves segmentation accuracy by 23% (single-granularity) and 85% (multi-granularity) over strong baselines, with more than double the per-iteration refinement gains of its predecessor PromptTSS.

This position paper argues that physiological waveform analysis should move beyond black-box prediction toward verifiable reasoning, mapping how foundation models and agentic LLMs can ground, verify, and explain conclusions drawn from signals such as ECG and PPG, and laying out the field's open challenges and research agenda.

FD-Bench is a modular benchmark for data-driven fluid simulation that enables fair, reproducible comparisons across spatial, temporal, and loss-function design choices, evaluating 85 baseline models over diverse flow scenarios with standardized protocols and a public leaderboard.

PromptTSS is a framework that unifies coarse- and fine-grained time series segmentation using prompts for dynamic adaptation. It achieves substantial accuracy gains in segmentation and transfer learning, showing strong effectiveness for hierarchical, evolving time series.

TimeDRL is a self-supervised learning framework for multivariate time series data that learns disentangled timestamp- and instance-level embeddings without relying on augmentations. It introduces dual-level objectives for predictive and contrastive learning, and achieves strong performance across forecasting and classification tasks, even in low-label scenarios.

COKE discovers causal graphs in manufacturing data with extreme missingness (up to 90%) by exploiting expert knowledge and the chronological order among sensors instead of imputing missing values, improving F1-score by 39.9% on average over prior methods on semiconductor manufacturing datasets.

RUN performs root cause analysis in microservice systems via neural Granger causal discovery with contrastive learning, exploiting temporal precedence in time series to localize failure sources more accurately than existing techniques.

Text2Freq integrates textual event information with time series through the frequency domain, aligning text to the low-frequency components of the series to bridge the modality gap and improve forecasting performance.

Proposes an align-then-fine-tune recipe for adapting pre-trained LLMs to time-series forecasting — first aligning temporal representations with the language model's embedding space, then fine-tuning for the forecasting task; the workshop precursor to LLM4TS.

Learns disentangled timestamp-level and instance-level representations of multivariate time series with self-supervised objectives, reducing reliance on augmentation-induced inductive biases; the workshop precursor to TimeDRL.