Audio reasoning stops paying for equal-length traces: per-section self-rated difficulty holds MMAR at 62.0% while cutting tokens 62.3%

Reasoning effort in audio models moves from one overall difficulty for the whole trace to a per-section decision about how long each part should be. Earlier difficulty-aware methods typically use a single global measure such as rollout performance, attention statistics, or trace-level rewards, which says whether more reasoning is needed but gives limited guidance on where to spend it. On the 1,000-example test sets of MMAU-test-mini and MMAR, SWEAR reaches 73.5% and 62.0% accuracy with 197.13 and 195.85 generated tokens, 60.1% and 62.3% fewer than CoTA-SFT.

During supervised fine-tuning the model learns to predict each section's difficulty rating before generating that section, so the rating acts as a causal conditioning token rather than a post-hoc annotation. DiCoTA extends CoTA by assigning difficulty targets to the Planning, Caption, and Reasoning sections while Summary keeps a fixed budget, and it characterizes difficulty along five dimensions: production quality, production complexity, integration score, reasoning demand, and question quality. Production quality and production complexity come from Audiobox Aesthetics, while integration score, reasoning demand, and question quality come from Qwen3-Omni-30B-A3B; training uses 402,496 examples for SFT and 113,410 for RL on a Qwen2.5-Omni-7B backbone.

In reinforcement learning the predicted rating controls each section's length, and a feedback-driven curriculum introduces harder samples as competence grows. The length reward follows a rating-dependent target ratio so higher ratings permit longer generated sections, and the KL coefficient adapts with predicted difficulty so harder inputs may deviate further from the frozen SFT reference. Ablations show that removing rating calibration costs 3.2 and 2.6 points on MMAU and MMAR, removing length shaping adds 19.5 and 29.6 tokens, and static KL lowers accuracy on both benchmarks.

The report covers results on the 1,000-example test sets of MMAU-test-mini and MMAR, where answer accuracy measures effectiveness and average generated tokens measure efficiency, with difficulty groups partitioned independently using the same procedure as DiCoTA. The intended readers are researchers and engineers tracking the inference cost of large audio-language models. The authors list extending this difficulty-aware principle to more diverse datasets and multimodal reasoning domains as future work.

Difficulty scoring depends on offline assessors: production quality and production complexity come from Audiobox Aesthetics, while integration score, reasoning demand, and question quality come from Qwen3-Omni-30B-A3B. Ablations show that removing length shaping raises MMAU accuracy by 0.7 points but lowers MMAR accuracy by 1.6 points, so length control does not act uniformly across benchmarks. Four curriculum buckets give the highest accuracy, while ten buckets use fewer tokens but lose accuracy.