Preprint reports lower training cost for audio-visual reasoning

A preprint reports that text-only post-training improved Qwen2.5-Omni-7B's geometric mean of nine reasoning scores by 25.83% over the base model, outperforming a complete audio-visual training route while using 56.6% fewer GPU-hours. The method combines supervised fine-tuning with reinforcement learning, which trains a model using feedback on its outputs. Text-only training degraded perception, so the researchers added a smaller audio-visual reinforcement-learning stage to refine that capability. That refinement used about 90% fewer input tokens than full-data audio-visual reinforcement learning, restored perception above the base level and retained 93.5% of the best text-only pipeline's reasoning gain.