$ cat wiki/papers/2026/2610.26233-heartbeat-thinking.md
Heartbeat Thinking
TL;DR
Heartbeat Thinking lets voice agents continue reasoning and tool execution while updated speech arrives. On FDBench v3, the authors report 16 percentage points more tool coverage and 36.4% lower mean final-response latency than a SHANKS-style windowed baseline. (source)
Authors & Org
Zhenyu Zhang, Kewei Fu, Yongqi Tong, Chenzhi Zheng, Wei Wu, Jiang-Ming Yang and Jianshe Li. First submitted 2026-10-19. Affiliations are not established by the abstract page. (source)
Method
Chunk-bounded or interrupt-on-update approaches can stop a speculative tool call before its arguments are complete. Heartbeat instead starts new Thinker invocations for updated input while allowing existing invocations and dispatched tools to finish. Periodic invocations read the latest partial transcript and a shared record of completed reasoning and tool results, using these to resolve missing arguments and initiate dependent calls. (source)
Results
Across SpeechFCEval, FDBench v3 and AudioCRAG, the abstract reports better aggregate speculative tool coverage or retrieval reuse than the windowed baseline. Effects on final task quality depend on the benchmark. The FDBench v3 latency reduction is a relative change in the mean; the tool-coverage increase is in percentage points. Neither is a general task-success rate. (source)
Significance
For Agents (LLM Agents), the result makes scheduling during listening part of the agent design. Interpretation: preserving useful computation across speech updates can matter alongside the choice of model; Eval Harness Configuration should retain the baseline and outcome definitions. (source)
Open Questions
The abstract does not specify absolute latency, additional compute cost, or how speculative tool execution handles irreversible side effects and later corrections to user intent. Those remain questions for a full-paper and implementation review; neither has been performed here. (source)
Cite
arXiv:2610.26233 · Immutable abstract and metadata capture. (source)