Step-Audio-R1.5 Technical Report

Recent advancements in large audio language models have extended Chain-of-Thought (CoT) reasoning into the auditory domain, enabling models to tackle complex acoustic and spoken tasks. To elicit these reasoning chains, the prevailing paradigm relies on Reinforcement Learning with Verified Rewards (RLVR), which optimizes models to distill continuous auditory contexts into isolated, verifiable text labels. However, the authors identify this as a 'verifiable reward trap': RLVR yields high scores on standardized benchmarks but systematically degrades real-world conversational feel, reducing dynamic interactions to mechanical 'answering machines' that compromise prosodic naturalness, emotional continuity, and user immersion, especially in long-turn dialogues.

To bridge the gap between objective verification and genuine sensory empathy, the team introduces Step-Audio-R1.5, marking a paradigm shift toward Reinforcement Learning from Human Feedback (RLHF) in audio reasoning. Comprehensive evaluations show that Step-Audio-R1.5 not only maintains robust analytical reasoning but profoundly transforms the interactive experience, redefining the boundaries of deeply immersive long-turn spoken dialogue. This approach prioritizes acoustic nuance and emotional continuity over isolated correctness, offering a more natural and engaging user experience.