List your company

Paralinguistic dialogue behavior on VoxSafeBench

60.97Child PresenceGRPO
-2.438814.023130.48546.9469Jun 9, 2026

Evaluation results

MethodLinks
GRPOInference Setting=Scaffolded2026.0660.97—21.8648.1513.9935.5
ParaBridgeInference Setting=Scaffold-free2026.0658.56—18.4355.5614.9340.33
ParaBridgeInference Setting=Scaffolded2026.0657.01—16.835515.7639.85
RFTInference Setting=Scaffold-free2026.0650.68—14.911.2916.231.64
RFTInference Setting=Scaffolded2026.0650.45—17.8820.3714.3432.35
BaselineInference Setting=Scaffolded2026.0645.83—17.1329.4113.1929.02
BaselineInference Setting=Scaffold-free2026.060—6.1107.9914.64
GRPOInference Setting=Scaffold-free2026.060—6.5804.8415.67