← News

ResearchSources: arXiv · 20 Nov 2026

Duopra pairs full-duplex speech with tool execution, lifting benchmark scores

The arXiv paper introduces a voice model that learns to call tools during conversation, reporting large Pass@1 gains on Full-Duplex-Bench v3.

In plain words

Voice AI models that can listen and speak at the same time are good at natural conversation but often fail when a task requires actually running a tool. Duopra is a model trained to do both: manage the conversation and issue correct, complete requests to a separate execution system. Its authors report much higher success rates on a benchmark of such tasks.

What was announced

A new arXiv paper argues that full-duplex interaction models, which handle natural, low-latency turn-taking, backchannels and interruptions, still lack task agency: the ability to invoke tools on time and revise arguments after user corrections. The authors note that in their streaming pipeline, most failures of a Realtime-Venus-Audio setup with a harness on Full-Duplex-Bench v3 happened before any domain tool executed, since unsent requests cannot run and incomplete ones can misdirect the backend.

Duopra treats the backend as a tool that accepts a self-contained natural-language request, with direct calls and delegation sharing one action interface. Its three-stage training combines protocol SFT, full-duplex reinforcement learning that rewards timely and complete execution, and projector-only repair of talker conditioning after thinker updates. The authors report Pass@1 rising from 4.7% to 44.7% with the same backend, 47.7% when Duopra calls tools directly, and 49.3% with a DeepSeek backend; the model also improves three FDB v1.0 takeover rates over the base model. The paper concludes that task agency can coexist with key conversational agency.

What it does not show

  • Evidence is limited to the paper's abstract; the full article, method details and independent verification are not available.

For security and IT teams

Scope
The reported results come from the authors' own streaming pipeline and benchmarks (FDB v3 and FDB v1.0), not from independent evaluation.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.