Continuous Audio Thinking for LALMs: Qwen2-Audio, Flamingo3

Researchers Gyojin Han, Dong-Jae Lee, Changho Choi, Jongsuk Kim and Junmo Kim introduced Continuous Audio Thinking (CoAT) on 5 Jun 2026 in an arXiv preprint (arXiv:2606.18273), a framework that adds a continuous latent workspace to large audio language models. The paper evaluates CoAT on three LALMs — Qwen2-Audio, Qwen2.5-Omni-7B, and Audio Flamingo 3 — and reports performance gains across a benchmark suite spanning audio reasoning, audio understanding, music classification, speech emotion and speech transcription.

How does Continuous Audio Thinking work?

CoAT creates a continuous latent workspace, called a continuous thinking block, that organizes acoustic information before the model generates text. The authors ground the thinking space by distillation from audio experts so the model can “utilize the rich acoustic information provided by expert distillation when generating its response.” The continuous thinking block is processed in a single prefill, and the paper states that "CoAT does not require additional autoregressive decoding cost over the baseline."

The paper frames the problem as follows: typical LALMs shape hidden states for text generation, which erodes diverse acoustic content such as phonetic detail, prosody, sound events, affect and pitch. CoAT inserts an intermediate latent workspace to preserve and expose that acoustic detail to the downstream text decoder, with auxiliary supervision distilled from specialist audio models.

Which models and tasks did the authors test?

The evaluation runs CoAT across three named large audio language models: Qwen2-Audio, Qwen2.5-Omni-7B, and Audio Flamingo 3. The reported improvements cover a broad benchmark suite that the paper describes as spanning audio reasoning, audio understanding, music classification, speech emotion and speech transcription. The authors also analyze signal flow inside the model and find that the auxiliary supervision given to the thinking positions propagates into the model's textual responses.

The submission to arXiv is dated 5 Jun 2026 and appears as arXiv:2606.18273. The paper includes an explicit claim that the thinking block can be processed in a single prefill, which keeps inference autoregressive cost at parity with the baseline model.

Why it matters

CoAT addresses a concrete architectural gap in large audio language models: hidden states optimized for text generation tend to lose fine-grained acoustic cues. By providing a continuous latent workspace and distilling audio experts into that space, the approach preserves acoustic detail and makes it available to text generation. For researchers and engineers, that means improved performance on tasks that require retaining pitch, prosody, affect or non-speech sound events while still producing text-aligned outputs.

Integrating a thinking block that incurs no extra autoregressive decoding cost lowers the engineering barrier to trying the idea on existing LALMs. If the claimed propagation of auxiliary supervision into textual outputs holds across use cases, CoAT could become a practical extension to multi-modal stacks that must balance acoustic fidelity with natural text responses.

What to watch

Check the paper's Code, Data and Media section and the associated links (Hugging Face, Replicate, DagsHub toggles shown on the arXiv page) for released implementations, weights or benchmarks. The next concrete signals will be released code, reproduced benchmark numbers on the named LALMs, or community replications that confirm the paper’s reported gains across the audio reasoning and transcription tasks.

References

Han G., Lee D.-J., Choi C., Kim J., Kim J., "Continuous Audio Thinking for Large Audio Language Models," arXiv:2606.18273, submitted 5 Jun 2026.