Audio-Mind: new auditable framework boosts audio QA to 82.8% accuracy

Audio-Mind, introduced by Yucheng Wang and eight co-authors, addresses a key challenge in audio understanding: deciding when agentic evidence acquisition actually improves performance. Existing audio agents blindly decompose questions into tool calls, but as large audio-language models (LALMs) grow stronger, the ability to keep their holistic judgment intact becomes critical. Audio-Mind acts as an auditable, pluggable layer that dynamically combines a strong LALM frontend with a planner that invokes external tools only when initial evidence is insufficient. This conditional acquisition avoids the orchestration bottleneck seen in prior agentic systems, where forced decomposition degrades the frontend's native reasoning.

On two benchmarks – MMAR and MSU-Bench – Audio-Mind achieves 80.4% and 82.8% accuracy respectively, surpassing all previous audio-agent baselines. The authors also performed a matched-backbone comparison that reveals why the design works: when the audio frontend is already powerful (e.g., a large model pre-trained on extensive audio data), an unconstrained agentic workflow can actually hurt performance. Audio-Mind preserves the frontend's end-to-end judgment when it is already confident, only pulling in bounded external evidence (e.g., specialized tool outputs) for questions with clear evidence gaps.

Beyond raw accuracy, Audio-Mind generates higher-quality reasoning traces that are fully auditable. Each answer comes with a record of uncertainty, tool evidence, and the chain of rationale used to derive the final result. This transparency makes the framework particularly valuable for error analysis and for generating reliable audio-QA annotations at scale, where understanding why a model arrived at a certain answer is as important as the answer itself. The work is available on arXiv under reference 2605.28480.