Teaching an AI to "think longer" about audio input can make it worse at understanding that audio.
Researchers studying Large Audio-Language Models - AI systems that process both sound and text - found that training these models to reason step-by-step, a trick that reliably helps text-based AI, produced only marginal gains or outright regressions compared to just answering directly. To figure out why, the team built an evaluation framework called CAFE to isolate exactly where audio reasoning breaks down. The diagnosis: as a model's reasoning chain gets longer, it progressively loses track of what it actually heard, a failure mode the researchers call audio perception decay.
That is a specific and unglamorous failure of chain-of-thought reasoning, the same technique credited with big jumps in math and coding AI over the past two years. Extra reasoning steps only help if the model keeps checking its work against the real input, and audio models apparently do not.
The proposed fix, MPAR2, forces models to break questions into smaller audio-focused sub-steps and uses reinforcement learning to keep them anchored to the sound - lifting perception accuracy on CAFE from 31.74% to 63.51% and reaching 74.59% on the MMAU benchmark, though both numbers come from the same paper introducing the method.