Researchers at Griffith University and CSIRO have tested an AI model that keeps processing a stream of measurements while it writes an answer. In their benchmark, TimeInteract almost eliminated the pause in incoming data that occurred when comparison systems stopped reading to generate a response.
A conventional model can answer a question after receiving a complete set of measurements. Streaming systems read new values as they arrive, but some still alternate between reading and speaking: if a response takes time, later observations wait. That matters when an instruction is to report a power spike or keep watch for a changing pattern while the conversation continues.
TimeInteract separates those jobs. One processing branch keeps the incoming values and user instructions in view and decides when a reply is needed. When it decides to respond, it forks the model's working memory into another branch that generates text. New observations can then reach the first branch without waiting for the sentence to finish.
The team built StreamTSI-34K to test answering questions, obeying persistent instructions, issuing warnings and adapting to user guidance. It contains 34,588 interaction episodes drawn from synthetic signals and recorded data in six domains, but most interaction scenarios were constructed synthetically. TimeInteract scored above the evaluated language, vision-language and time-series models on all four task groups. On multi-turn user-guided adaptation, it succeeded in 36.96% of cases, against 13.04% for the strongest comparator on that task, Qwen3-14B. The comparison models differed in size and input format, and none natively handled streaming time-series input in the same way.
The speed test compared three inference approaches at several answer lengths. For answers longer than 70 tokens, TimeInteract's average pause in reading new observations was approximately zero, against about 2.04 seconds for the alternating approach. Completing a response took about 2.40 seconds, versus 3.83 seconds for that approach and 5.18 seconds for an offline one. The paper does not report latencies from a live monitor handling naturally occurring requests.
Sources
- TimeInteract research paper, 22 September 2026
- Full paper, including benchmark tables and limitations
- Illustrative oscilloscope photograph, not equipment used in the study: Jeff Keyzer, perspective correction by Pittigrilli, via Wikimedia Commons, CC BY-SA 2.0.
