Continuous Audio Thinking for Large Audio Language Models

A new framework called Continuous Audio Thinking (CoAT) equips large audio language models with a continuous latent workspace to preserve acoustic detail before generating text responses, according to a paper submitted to arXiv in June 2026 [1][2]. Large audio language models (LALMs) can handle tasks from speech transcription to music analysis, but their hidden states are shaped for text generation rather than for retaining acoustic information such as phonetic detail, prosody, sound events, affect, and pitch [1][2]. The CoAT framework addresses this by introducing a continuous thinking block that organizes acoustic information prior to response generation, grounded by distillation from audio experts [1][2]. The block can be processed in a single prefill, so CoAT does not require additional autoregressive decoding cost over the baseline [1][2]. The authors evaluated CoAT across three LALMs: Qwen2-Audio, Qwen2.5-Omni-7B, and Audio Flamingo~3 [1][2]. Performance gains were observed on a benchmark suite spanning audio reasoning, audio understanding, music classification, speech emotion, and speech transcription [1][2]. Further analysis confirmed that auxiliary supervision propagates from the thinking positions to the model’s textual responses [1][2]. The paper was posted on arXiv, an open-access repository of electronic preprints that has been operating since August 1991 and now receives roughly 24,000 submissions per month as of late 2024 [10]. The work appears under the Computation and Language category and is accessible through arXiv’s abstract page, which also features community-developed tools via the arXivLabs framework [7][8]. arXivLabs, launched in 2020, allows collaborators to build experimental features such as bibliographic explorers and code finders that sit directly on article record pages, under guidelines that enforce openness, community values, and user data privacy [8][9]. LALMs belong to the broader class of generative AI models that learn patterns from training data and produce new content in response to prompts, a field that expanded rapidly during the AI boom of the 2020s [3]. The CoAT approach, which separates acoustic reasoning from text generation inside a neural network, echoes principles explored in neuro-symbolic AI, where researchers combine statistical learning with structured reasoning to improve reliability and generalization [5]. Deep learning architectures, including transformers, provide the substrate for such models, using multilayered neural networks trained on large datasets [6].