Heartbeat Thinking: faster voice replies can still act too early

How preparation works

The system periodically reasons over the partial transcript, then hands available work to the responder. Authors report more useful tool preparation and lower final-response latency on their evaluation; those measurements were not reproduced here.

Heartbeat Thinking ↗

A correction can leave an earlier action behind

The appendix describes a cart update made before the speaker finished stating a budget. A later search found an affordable product, but it did not undo the earlier cart action. Some final answers also reused stale speculative results.

Heartbeat speculative-tool analysis ↗

An evaluation worth running

Use a mock store with fictional items. Say “add a lamp” and then finish with “under thirty dollars.” A useful test checks the final cart, every intermediate write and the spoken answer. Checking only whether the last search used the right budget would miss the failure.

Heartbeat speculative-tool analysis ↗

What you can do

1. Read the speculative-tool failure analysis before using partial speech to authorize an external action. Open reference ↗

Where the evidence stops

This original test idea was not executed. The paper’s aggregate latency gains do not establish safe behavior for payments, bookings or other durable actions.

Sources read. Checked October 23, 2026. The paper and speculative-action appendix were read. No voice agent or mock-store test was run.

Sources & further reading