Source-linked AI summary
SWEAR: Section-Wise Efficient Audio Reasoning via Self-Rated Difficulty
Yixuan Li, Minghao Lv, Xuenan Xu, Zihao Zheng, Mengyue Wu
TL;DR
Large audio-language models use explicit reasoning at added computational cost, but coarse global difficulty estimates provide limited guidance on where to allocate effort. SWEAR builds on five-dimensional section-wise difficulty scores and compressed traces, then uses SFT and RL to adapt reasoning allocation; experiments report accuracy gains with substantially fewer generated tokens.
Problem
Global difficulty estimates in large audio-language models indicate whether more reasoning is needed but offer limited guidance on allocating effort across distinct reasoning stages.
Source: §1 · p. 1Method
SWEAR builds DICOTA, a 575K-example dataset with five-dimensional section-conditioned difficulty scores and compressed traces, then learns self-ratings in SFT and uses them in RL to control reasoning length.
Source: §1 · p. 1Results
Accuracy improves by 2.6 and 6.3 percentage points while generated tokens fall by 60.1% and 62.3% on MMAU and MMAR, respectively, versus SFT without difficulty supervision.
Source: §1 · p. 1Takeaways & Limitations
SWEAR produces concise traces for easy problems while preserving sufficient deliberation for hard ones on MMAU and MMAR.
Source: §4 · p. 4The reported experiments cover MMAU and MMAR; extension to more diverse datasets and multimodal reasoning domains remains future work.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Recent work on large audio-language models (LALMs) explores adaptive reasoning to balance accuracy and efficiency. However, existing approaches often rely on coarse-grained difficulty estimation, overlooking the diverse factors that determine reasoning effort. We address this limitation by deriving five-dimensional difficulty scores and compressed reasoning traces from complementary offline assessments. Based on this characterization, we propose SWEAR, a difficulty-aware reasoning framework that learns difficulty signals during supervised fine-tuning and uses reinforcement learning to adapt reasoning allocation across reasoning stages. Experiments on MMAU-test-mini and MMAR show that SWEAR improves accuracy while substantially reducing generated tokens, shortening reasoning for easy instances without sacrificing performance on difficult ones.
1. INTRODUCTION
Explicit reasoning in large audio-language models adds computational cost, while global difficulty estimates offer limited guidance on where to allocate effort. SWEAR combines section-wise difficulty supervision and compressed traces with SFT and RL to adapt reasoning effort.
1. INTRODUCTION
- Adaptive reasoning methods in text and vision adjust reasoning length or computational budgets using instance difficulty, with potential to balance accuracy and efficiency across modalities.Source: §1 · p. 1
- Existing LALM approaches often use a single global difficulty measure, which indicates whether more reasoning is needed but gives limited guidance on allocating effort across distinct reasoning stages.Source: §1 · p. 1
- Accuracy improves by 2.6 and 6.3 percentage points while generated tokens fall by 60.1% and 62.3% on MMAU and MMAR, respectively, versus SFT without difficulty supervision.Source: §1 · p. 1
- DICOTA contains 575K audio-reasoning examples with five-dimensional, section-conditioned difficulty scores and compressed traces; SWEAR learns self-ratings during SFT and uses them in RL for adaptive reasoning-length control.Source: §1 · p. 1
2. METHOD
SWEAR builds section-wise difficulty signals and compressed reasoning targets, then teaches a model to predict difficulty ratings and uses them to adapt reasoning during reinforcement learning. Its method combines difficulty-weighted section budgets with quality feedback, length rewards, and rating calibration.
2.1. DiCoTA: Difficulty-Aware Trace Construction
- DICOTA assigns difficulty targets to Planning, Caption, and Reasoning sections using five dimensions spanning acoustic quality, audio complexity, and reasoning demand.Production quality and complexity come from Audiobox Aesthetics; integration score, reasoning demand, and question quality come from Qwen3-Omni-30B-A3B.Source: §2.1 · p. 2
- Section difficulties aggregate normalized proxies by functional role, then map to five equal-width levels represented by rating tokens R1–R5.Planning captures question interpretation, Caption acoustic description, and Reasoning inference demand and acoustic-evidence requirements.Source: §2.1 · p. 2
- DICOTA allocates a fixed trace budget by continuous section difficulty, preserving a fixed Summary budget and assigning more tokens to harder sections.The teacher compresses redundant steps while preserving each section’s function and consistency with the answer; a difficulty floor prevents zero allocations.Source: §2.1 · pp. 2–3
- SFT trains the policy to predict each section’s rating before generating its content, making the rating a causal conditioning token for later curriculum RL.The structured trace includes Planning, Caption, Reasoning, Summary, and Response, with difficulty targets for the first three sections.Source: §2.1 · pp. 2–3
2.2. Difficulty-Adaptive Curriculum RL
- Curriculum RL optimizes a clipped GRPO objective with group-normalized advantages and KL regularization, while rewards cover output format, task quality, and rating-conditioned section length.The target generated-to-reference length ratio increases with the predicted section rating.Source: §2.2 · p. 3
- A feedback-driven curriculum ranks samples into easy-to-hard buckets and uses recent quality rewards to advance the model through them.Earlier buckets remain active with exponentially downweighted sampling after a new bucket is activated.Source: §2.2 · p. 3
- RL adapts KL regularization and calibrates predicted ratings against offline difficulty; at inference, ratings, sections, and answers are generated without external difficulty scoring or reference traces.Higher predicted difficulty permits greater deviation from the frozen SFT reference.Source: §2.2 · p. 3
3. EXPERIMENT AND RESULTS
Across MMAU-test-mini and MMAR, SWEAR ties the best accuracy on the former and leads on the latter while using fewer tokens than CoTA-SFT. Difficulty-stratified results and component ablations characterize its reasoning allocation and training design.
3.1. Experimental Setup
- Evaluation uses 1,000-example test sets for MMAU-test-mini and MMAR, with accuracy and average generated tokens as effectiveness and efficiency metrics.The study uses 402,496 examples for SFT and a separate 113,410 reward-compatible examples for RL.Source: §3.1 · p. 3
3.2. Results and Analysis
- 73.5% accuracy on MMAU-test-mini tied the best result and 62.0% on MMAR led, with 197.13 and 195.85 tokens, down 60.1% and 62.3% versus CoTA-SFT.Table 1 reports accuracy and average generated tokens across both benchmarks.Source: §3.2 · pp. 3–4, Table 1 · p. 4
- On MMAU, output length rises from 172.8 to 222.5 tokens across difficulty and hard accuracy beats CoTA-SFT by 5.6 points; on MMAR, SWEAR outperforms both SFT variants at every level.Across all difficulty splits, SWEAR uses 58.3–63.3% fewer tokens than CoTA-SFT and 5.4–16.6% fewer than DiCoTA-SFT.Source: Figure 2 · p. 4, §3.2 · p. 4
3.3. Ablation Studies
- An ablation lowers accuracy by 3.2 points on MMAU and 2.6 on MMAR; removing length shaping adds 19.5 and 29.6 tokens, changing accuracy by +0.7 and −1.6 points, respectively.Static KL lowers accuracy on both benchmarks; four curriculum buckets achieve the highest accuracy among curriculum settings.Source: §3.3 · p. 4
4. CONCLUSION
SWEAR models audio reasoning difficulty across trace sections and uses learned self-ratings to adapt reasoning during curriculum reinforcement learning. Experiments on MMAU and MMAR show concise traces for easy problems while preserving deliberation for hard ones.
4. CONCLUSION
- SWEAR augments CoTA traces with multi-dimensional difficulty annotations and distills them into the LALM’s self-ratings during supervised fine-tuning.The paper identifies extension to more diverse datasets and multimodal reasoning domains as future work.Source: §4 · p. 4
- The learned ratings guide curriculum reinforcement learning that adapts reasoning length and sectional demands.Source: §4 · p. 4
- On MMAU and MMAR, SWEAR produces concise traces for easy problems while preserving sufficient deliberation for hard ones.Source: §4 · p. 4
5. COMPLIANCE WITH ETHICAL STANDARDS
The passage identifies the work as a numerical simulation study and states that ethical approval was not required.