X$^3$-OPD: Distilling Reasoning into Large Audio-Language Models via On-Policy Alignment
The paper claims audio-language models can pick up stronger reasoning by training on their own audio-conditioned chains, with a text model guiding each step.
X$^3$-OPD uses matched text inputs and verified answers from a teacher model while the audio student produces its own reasoning trajectories. The authors also build a three-part corpus spanning speech-rendered text reasoning, complex audio-event reasoning, and spoken-dialogue reasoning with prosody and context. They report gains on MMSU, MMAU, BIG Bench Audio, and MMAR, including better audio-grounded reasoning and chain-of-thought quality while mostly preserving prior capabilities under domain shift. Source: ArXiv · AI/CL/LG's note.
X$^3$-OPD uses matched text inputs and verified answers from a teacher model while the audio student produces its own reasoning trajectories. The authors also build a three-part corpus spanning speech-rendered text reasoning, complex audio-event reasoning, and spoken-dialogue reasoning with prosody and context. They report gains on MMSU, MMAU, BIG Bench Audio, and MMAR, including better audio-grounded reasoning and chain-of-thought quality while mostly preserving prior capabilities under domain shift. Source: ArXiv · AI/CL/LG's note.
score 5