X³-OPD: On-Policy Distillation Closes the Reasoning Gap Between Audio-Language Models and Text LLMs

arXiv 2607.21550·medium signal

Dongjie Fu, Di Cao and Xize Cheng target the gap where large audio-language models have strong auditory perception but lag text LLMs on deep logical reasoning, using on-policy alignment to distill reasoning capability into the audio model. On-policy distillation matters here because the audio model's own error distribution differs from the text teacher's, so off-policy transfer misses the failure cases that actually occur. Relevant to anyone building voice agents where the model hears correctly but reasons poorly about what it heard.