Found first: a primary source the press has not covered yet.
A seven-author paper submitted October 19, 2026 proposes "Heartbeat Thinking," a method that fires periodic reasoning invocations while a user is still speaking, letting a voice agent begin executing tools before speech ends. On FDBench v3, it reduced mean final-response latency by 36.4% and raised tool coverage by 16 percentage points over a windowed baseline. The paper is at arxiv.org/abs/2610.26233.
What the source says
The paper, by Zhenyu Zhang, Kewei Fu, Yongqi Tong, Chenzhi Zheng, Wei Wu, Jiang-Ming Yang, and Jianshe Li (affiliations not listed in the abstract), introduces a "Thinker" component invoked at regular intervals during speech. Each invocation reads the latest partial transcript alongside a shared record of completed reasoning and prior tool results. Invocations can overlap and tools run asynchronously, so early work is not discarded when the transcript grows. The latency and coverage gains come from FDBench v3; the authors also confirm improvements on SpeechFCEval and AudioCRAG, but note the effects on final task quality vary by benchmark.
Why it matters
In current voice agent architectures, reasoning starts only after the utterance ends, and that post-speech wait is the main driver of perceived latency. Heartbeat Thinking reclaims that time by treating the listening phase as usable compute. The benchmark numbers are concrete and the method generalises across three evaluation sets, which makes this more than a narrow benchmark trick. The caveat the authors themselves raise, that final task quality gains are benchmark-dependent, is the honest limitation to watch: if early tool calls frequently misfire on production queries, the latency win shrinks.