Multi-axis reinforcement learning for full-duplex interactivity

What changed

Extracted short conversation segments for pause handling, turn-taking, backchannels and user interruptions and optimized axis-specific rewards with GRPO plus a semantic-quality judge reward.

Scientific scope

Extracted short conversation segments for pause handling, turn-taking, backchannels and user interruptions and optimized axis-specific rewards with GRPO plus a semantic-quality judge reward. Inherits GRPO, existing speech models and prior interactivity alignment. Tested on Moshi and PersonaPlex; judge rewards and finite multi-turn tests do not prove preservation of meaning or universal interaction quality.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits GRPO, existing speech models and prior interactivity alignment. Tested on Moshi and PersonaPlex; judge rewards and finite multi-turn tests do not prove preservation of meaning or universal interaction quality.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models
• Kyutai scientific publications index

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction