TimeInteract introduces a new mode for streaming time series where a model continuously ingests incoming observations and user intent, decides when to stay silent or respond, and keeps processing new data even while generating a response. It combines a dual-view streaming encoder, a learned response trigger, and decoupled inference so input handling is not blocked by output generation.
The authors built StreamTSI-34K, a dataset of 34,588 episodes and 77,505 responses spanning synthetic and real-world time series in single- and multi-turn interactions. Across a four-level hierarchy of interaction capabilities, TimeInteract outperforms existing LLMs, VLMs, and TSLMs by up to 23.92 points on challenging tasks, while also improving response triggering, achieving near-zero stream stall, and delivering up to 2.15x inference speedup.
This matters because real-time monitoring in healthcare, industry, finance, and energy needs models that can answer and act while the underlying stream keeps evolving. Existing static or interleaved designs introduce latency exactly when timely response is critical.