Thinking While Listening: Simple Test Time Scaling For Audio Classification

We propose a framework that enables neural models to “think while listening” to everyday sounds, thereby enhancing audio classification performance. Motivated by recent advances in the reasoning capabilities of large language models, we ad- dress two central questions: (i) how can thinking be incorpo- rated into existing audio classification pipelines to enable rea- soning in the category space and improve performance, and (ii) can a new architecture be designed from the ground up to support both thinking and test-time scaling? We demon- strate that in both settings, our models exhibit improved clas- sification accuracy. Leveraging test-time scaling, we observe consistent gains as the number of sampled traces increases. Furthermore, we evaluate two open-source reasoning mod- els, GPT-OSS-20B and Qwen3-14B, showing that while such models are capable of zero-shot reasoning, a lightweight ap- proach—retraining only the embedding matrix of a frozen, smaller model like GPT-2 can surpass the performance of billion-parameter text-based reasoning models.