Jarvis keeps background tasks aligned with live conversations
In plain words
Some assistants talk to a user while other parts of the system work on requests in the background. If the user changes a request, an old plan or answer can become out of date.
Jarvis is a coordination harness that records tasks and checks that decisions still match the current task before code acts on them. On one test of voice customer-service calls, Jarvis solved 192 of 278 tasks, while Qwen Audio Agent solved 29 under the same front model, admission path and back end. A task counted only when its final database state and required actions were correct.
The research
When an assistant works on a request while the user keeps talking, how can it avoid acting on an old version of the task? Recognition and reasoning can be accurate, yet a plan may no longer match what the user wants.
Jarvis addresses this coordination problem by linking decisions to current task state and checking them before they take effect.
Jarvis ties each model decision to a live task version and uses code to reject stale choices before background work can act or reach the user.
The question is whether coordination can keep an assistant’s actions aligned with a changing conversation while background work continues. Jarvis records task versions, offers decisions tied to current tasks and checks those choices in code before they take effect.
On τ-Voice, a test of voice customer-service calls, tasks count as solved only when the final database state and required actions are correct. With the front model, admission path and back end fixed, Jarvis solved 192 of 278 tasks on Venus-Audio, against 29 for Qwen Audio Agent. The authors interpret this comparison as evidence that coordination design matters in this setting.
The evidence comes from specific task sets, scripted timing cases and recorded game sessions. It does not establish how Jarvis performs across other interactive systems or live settings.
Analysis
Deep
- Original paper authors
- Yuxiang Wang, Shengbo Cai, Zhixian Zhao, Yingda Shen, Yuancheng Wang
- Submitted to arXiv
- License
- arXiv non-exclusive