Audio reasoning stops spending equal tokens on every step: self-rated section difficulty lifts MMAR to 62.0% while cutting tokens by about 60%
Synopsis
Large audio-language models that reason explicitly have mostly used one overall difficulty signal to decide whether to think longer, which gives little guidance on how much effort each part of the reasoning trace deserves. SWEAR scores difficulty along five dimensions covering acoustic quality, acoustic complexity and reasoning demand, uses those offline scores to compress reference traces, teaches the model during supervised fine-tuning to predict each section's difficulty before writing it, and then lets reinforcement learning set section length from those self-ratings. On MMAU-test-mini and MMAR, accuracy rises by 2.6 and 6.3 percentage points over CoTA-SFT without difficulty supervision, while generated tokens fall by 60.1% and 62.3%.
Interpretation
Reasoning effort in audio models moves from one difficulty judgment for the whole trace to a per-section judgment inside the trace. Earlier difficulty-aware methods typically rely on a single overall measure such as rollout performance, attention statistics or trace-level rewards, which says whether more reasoning is needed but not which part of the trace is hard. On the 1,000-example test sets of MMAU-test-mini and MMAR, SWEAR improves accuracy by 2.6 and 6.3 percentage points over CoTA-SFT without difficulty supervision while reducing generated tokens by 60.1% and 62.3%.
The difficulty signal is predicted by the model itself before each section is written and acts as a causal condition for that section. Difficulty is no longer only a post-hoc annotation; it precedes the corresponding section and can therefore decide its length in real time. During supervised fine-tuning the model learns to predict each section's difficulty rating, and during reinforcement learning that rating sets the target length ratio, with higher ratings allowing longer generated sections.
Reasoning length grows with difficulty instead of using one budget for all samples. Uniform trace compression gives easy and hard samples similar reasoning lengths. As difficulty increases, average output length rises from 172.8 to 222.5 tokens on MMAU and from 174.4 to 223.2 on MMAR, while staying 58.3% to 63.3% shorter than CoTA-SFT across all splits.
Perspective
The reported scope is the two 1,000-example audio reasoning test sets MMAU-test-mini and MMAR, with Qwen2.5-Omni-7B as the backbone and training data drawn from 402,496 supervised fine-tuning examples and 113,410 reinforcement learning examples in DiCoTA. It is aimed at researchers and engineers working on audio-language model reasoning efficiency and adaptive compute allocation. The authors propose extending the difficulty-aware principle to more datasets and multimodal reasoning domains.
Difficulty ratings are self-predicted, and their calibration relies on soft-label constraints against unquantized offline difficulty; removing that calibration lowers accuracy by 3.2 and 2.6 points on MMAU and MMAR. Removing length shaping raises MMAU accuracy by 0.7 points but lowers MMAR accuracy by 1.6 points, so the link between length control and accuracy is not consistent across the two benchmarks. The curriculum-bucket ablation shows four buckets give the highest accuracy, with more buckets not improving further.