Lychee-FD mitigates modality interference in full-duplex SLMs via hierarchical parameter separation

Researchers introduce Lychee-FD, a native end-to-end full-duplex framework designed to address severe modality interference that degrades the performance of Spoken Language Models (SLMs). The work identifies that this interference stems from gradient conflicts between acoustic and semantic modeling when they share a deep parameter space.

• Proposes a hierarchical parameter separation strategy that decouples conflicting modalities in deep layers while preserving cross-modality coherence via a dedicated semantic alignment channel.
• Achieves substantial improvements in speech intelligence, yielding a +7.4% gain on Spoken QA benchmarks.
• Improves full-duplex interaction fluidity by +28.5% on FullDuplexBench 1.5 without compromising inference efficiency.

This approach represents the first method to elucidate the root cause of modality interference and provides a practical solution for seamless, high-performance native intelligent full-duplex SLMs.