AI

HiPLEX: Hierarchical Policy Factorization for Full Duplex Speech Language Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 02:18:25 UTC.

Briefings

HiPLEX is an RL framework for full-duplex speech language models that factorizes text policy into timing and content controls. It reduces takeover rates during pauses and shortens response latency compared to GRPO while maintaining quality. The model uses event-causal masks for timing and an LLM-judge for semantic advantages.

Key points

• HiPLEX factorizes a pretrained full-duplex text policy into a control policy for timing and a conditional content policy for content selection.
• The framework routes timing advantages to the token-group factor via event-causal masks and semantic advantages to the conditional content factor using an LLM-judge.

Context

Full-duplex speech language models are critical for natural real-time dialogue, requiring coordination of turn-taking, backchanneling, and floor management. Existing RL methods struggle to jointly optimize timing and content, prompting the need for a hierarchical factorization approach.

Limitations

Results are based on three Moshi seeds on Full-Duplex-Bench v1. Performance on Moshi and PersonaPlex is reported but specific metrics for those platforms are not detailed in the evidence.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 24/10/2026, 20:41:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicltgixy6n2c45cxx26izad7s4kzjzztht5jizkhzkqvyhtuiepse