Voice agents traditionally follow a rigid listen‑think‑speak loop, inserting seconds of silence before each reply. By adding a lightweight interrupt‑and‑resume orchestrator to an unmodified text model, we enable continuous‑time cognition—thinking while listening and while speaking—reducing overall latency by 19% and by about half in the targeted regime.
To measure the practical impact of continuous‑time thinking, we introduce ReactiveBench, which contains 120 interactive scenarios scored against pre‑registered binary requirements and a verifiable streaming track scored by exact correctness. The benchmark reveals a pitfall: LLM judges tend to reward visible reasoning, giving continuous thinking a large apparent advantage, but an independent judge flips this advantage and judge‑trained models actually satisfy fewer requirements when they think.
A five‑stage training study then pinpoints the right signal at three levels. Source: verifiable objectives turn thinking from harmful to helpful; Structure: whatever a uniform reward omits gets optimized away, and overly brief replies erode multi‑hop tool chaining; Optimizer: preference optimization can only trade conflicting sub‑goals, whereas on‑policy RL with a type‑shaped reward improves every correctness axis, raising streaming completion from 48% to 73%±5% across seeds and replicating at larger scale and on a second model. Orchestration makes continuous‑time interaction possible; a correctly sourced, shaped, and optimized verifiable signal makes it beneficial.
Review: Continuous‑time interaction achieved via soft interruption yields a more natural dialogue flow, but merely rewarding visible reasoning is insufficient for real‑world task success. Combining verifiable objectives with on‑policy RL is essential to turn thinking into measurable performance gains.