Change the schedule of the work
AsyncReasoning lets a language model generate reasoning and response text concurrently, instead of finishing one before starting the other. Its preprint was first submitted in December 2025 and revised in May 2026. The method changes inference—the execution of an already trained model—without additional training. [2]
- New input enters shared state
- Reasoning and writing can overlap
- Writing can pause for more reasoning
Reuse the computation, change the view
A key-value, or KV, cache stores computations associated with earlier tokens, the pieces of text a model processes. Yandex’s September 5 explanation describes treating this temporary state as reusable blocks. The paper gives reasoning and writing different logical orderings of shared state, rather than encoding every token twice. [1] [2]
Writing also needs a pause button
The paper alternates between concurrent reasoning/writing and a mode where writing waits. Its authors report that models sometimes make incorrect switching decisions. The Yandex overview also warns about stale information and duplicated work. Earlier output is therefore a scheduling capability, not proof that an answer is correct. [2] [1]
Go a little deeper
Optional reading · about 1 more minute
What would make the response useful?
Hypothetical example: While comparing two travel plans, an assistant acknowledges a newly supplied departure time before finishing the comparison. A useful test would check whether its final plan incorporates that correction, not merely whether some text appeared quickly.
Keep the claim at the tested scale
Our interpretation: Treat responsiveness, final accuracy and handling of interruptions as separate questions. The work suggests a way to overlap computation; it does not establish that every existing assistant uses this design or that concurrent output is always the right choice.
Original sources
Attributed synthesis, not original reporting. Examples labeled hypothetical or illustrative are explanatory. Reviewing a source does not independently validate its findings.
- Yandex Research: the KV cache as an agent runtime ↗
September 5, 2026 developer explanation; cache reuse, AsyncReasoning and limitations read September 26. No benchmark multiplier adopted.
- Yakushev and colleagues: Asynchronous Reasoning ↗
December 11, 2025 preprint, revised May 13, 2026; abstract and sections 3.1–3.3 read September 26. Work in progress, no independent replication.
