Paper deep dive
SWEAR: Section-Wise Efficient Audio Reasoning via Self-Rated Difficulty
Yixuan Li, Minghao Lv, Xuenan Xu, Zihao Zheng, Mengyue Wu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/12/2026, 6:52:15 AM
Summary
The paper introduces SWEAR, a difficulty-aware reasoning framework for Large Audio-Language Models (LALMs) that improves accuracy and efficiency by adapting reasoning effort based on section-wise difficulty scores. It utilizes a dataset called DiCoTA with five-dimensional difficulty annotations and employs Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to learn self-rated difficulty and allocate reasoning tokens dynamically.
Entities (10)
Relation Signals (8)
SWEAR → backedby → Qwen2.5-Omni-7B
confidence 96% · We use Qwen2.5-Omni-7B as the backbone.
SWEAR → evaluatedon → MMAU-test-mini
confidence 95% · Experiments on MMAU-test-mini and MMAR show that SWEAR improves accuracy
SWEAR → evaluatedon → MMAR
confidence 95% · Experiments on MMAU-test-mini and MMAR show that SWEAR improves accuracy
SWEAR → uses → DiCoTA
confidence 95% · We construct DiCoTA... SWEAR learns the ratings during Supervised Fine-Tuning
SWEAR → employs → Reinforcement Learning
confidence 94% · uses reinforcement learning to adapt reasoning allocation
SWEAR → employs → Supervised Fine-Tuning
confidence 94% · SWEAR learns the ratings during Supervised Fine-Tuning (SFT)
DiCoTA → derivedfrom → CoTA
confidence 90% · We built DiCoTA from CoTA [16]
SWEAR → outperforms → Audio-Reasoner-7B
confidence 90% · SWEAR improves accuracy... compared with SFT without difficulty supervision... Audio-Reasoner-7B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2610.13081v1
- Canonical: https://arxiv.org/abs/2610.13081v1
Trouble viewing inline? Open PDF directly →
Full Text
24,608 characters extracted from source content.
Expand or collapse full text
SWEAR: Section-Wise Efficient Audio Reasoning via Self-Rated Difficulty Yixuan Li Minghao Lv Xuenan Xu Zihao Zheng Mengyue Wu Corresponding author Abstract 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. Index Terms: Large Audio-Language Models, Structured Difficulty Modeling, Adaptive Reasoning †address: 1 MoE Key Lab of Artificial Intelligence, X-LANCE Lab, Shanghai Jiao Tong University, China 2 Shanghai AI Lab, China 3 KlingAI Research, China 1 Introduction LALMs increasingly use explicit reasoning to connect acoustic observations with answers [5, 13, 10]. As reasoning incurs additional computational cost, characterizing reasoning difficulty is essential for adaptive reasoning allocation, allowing more budget for challenging cases while reducing unnecessary computation for easier ones. However, accurately estimating input difficulty and determining the corresponding reasoning effort remain challenging, motivating difficulty-aware reasoning strategies that dynamically adjust reasoning allocation. Addressing this question, a parallel line of work has explored input-dependent reasoning effort rather than fixed CoT behaviors. In text reasoning, AdaptThink [20], AutoThink [12], and related methods [8] adapt reasoning length or computational budgets based on instance difficulty. Similar ideas have been investigated in visual reasoning, where FAST [15] and CAR [4] adjust reasoning effort according to input cues. These studies demonstrate the potential of adaptive reasoning for balancing accuracy and efficiency across modalities. Recent studies have explored difficulty-aware reasoning adaptation in LALMs. Audio-Thinker [14] and Omni-AutoThink [19] use reinforcement learning to adapt reasoning activation, while Think Smart, Not Hard [9] and CESAR [1] adjust reasoning length based on difficulty-related signals. However, these approaches typically characterize difficulty with a single overall measure, such as rollout performance, attention statistics, or trace-level rewards. Such global estimates indicate whether more reasoning is needed but provide limited guidance on where to allocate reasoning effort. This limitation is important for audio reasoning, where different stages, including question interpretation, acoustic evidence extraction, and answer inference, may introduce distinct and uncorrelated difficulties. Therefore, fine-grained and multi-dimensional difficulty characterization across reasoning sections is needed for adaptive allocation of reasoning effort within structured reasoning traces. To address this gap, we propose SWEAR, a framework for Section-Wise Efficient Audio Reasoning. In addition, We built DiCoTA from CoTA [16], which supplies section-wise supervision and compressed traces from five offline assessments. SWEAR learns the ratings during Supervised Fine-Tuning (SFT) and uses them during Reinforcement Learning (RL) to control section-wise reasoning length and reference-policy regularization, together with a feedback-driven curriculum. This encourages concise reasoning on easy inputs while preserving sufficient flexibility for challenging ones. Figure 1: Overview of our multi-dimensional difficulty-aware audio reasoning framework. (a) Offline linguistic and acoustic annotations estimate Planning, Caption, and Reasoning difficulty, which determine section ratings and compression budgets for structured targets. (b) Structured SFT teaches the model to produce concise, rated reasoning traces. (c) Curriculum RL progresses from easy to hard while jointly optimizing task quality, format, rating-conditioned section-length rewards, and teacher-anchored rating calibration under reference-policy KL regularization. Offline annotators are used only during training. Overall, our contributions are summarized as follows: • We construct DiCoTA, a 575K-example audio reasoning dataset with five-dimensional difficulty scores conditioned on reasoning sections and compressed reasoning traces. • We develop SWEAR, a difficulty-aware curriculum learning framework for multi-dimensional reasoning adaptation. SWEAR learns difficulty self-ratings during SFT and leverages them in RL for adaptive reasoning length control. • Experiments on MMAU and MMAR show that SWEAR improves accuracy by 2.6 and 6.3 percentage points while reducing generated tokens by 60.1% and 62.3%, respectively, compared with SFT without difficulty supervision. 2 Method SWEAR learns to allocate reasoning effort across trace sections through a three-stage pipeline, as illustrated in Fig. 1. First, we construct DiCoTA, an audio reasoning dataset with multi-dimensional difficulty annotations, by assigning difficulty scores to individual sections and compressing their content under difficulty-dependent budgets. and compressing each section under an adaptive budget. Second, SFT teaches the policy to predict each section’s difficulty rating before generating its content. Finally, difficulty-conditioned curriculum RL progressively aligns section-wise reasoning effort with the policy’s own difficulty predictions. 2.1 DiCoTA: Difficulty-Aware Trace Construction We construct DiCoTA, a difficulty-annotated extension of CoTA [16]. Each trace comprises Planning (P), Caption (C), Reasoning (R), and Summary (S), followed by a Response (A). We assign difficulty targets to P, C, and R, while S uses a fixed consolidation budget. We characterize audio reasoning difficulty through five dimensions spanning acoustic quality, audio complexity, and reasoning demand: production quality (PQ), production complexity (PC), integration score (IS), reasoning demand (RD), and question quality (Q). PQ and PC are obtained from Audiobox Aesthetics [11], measuring technical quality and the complexity of audio components, respectively. IS, RD, and Q are provided by Qwen3-Omni-30B-A3B [18], capturing reliance on fine-grained acoustic evidence, inference complexity, and question formulation quality. Together, these dimensions characterize complementary sources of difficulty, ranging from acoustic perception and evidence extraction to question understanding and answer reasoning. With gm(x)=(x−1)/(m−1)g_m(x)=(x-1)/(m-1) for an m-point score, we construct difficulty proxies by reversing Q and PQ: lower question clarity and audio quality receive higher proxy scores. q=1−g5(Q),i=g5(IS),r=g5(RD),p=1−g10(PQ),c=g10(PC). gatheredq=1-g_5(Q), i=g_5(IS), r=g_5(RD),\\ p=1-g_10(PQ), c=g_10(PC). gathered (1) We define the integrated difficulty of each trace section by aggregating the relevant normalized proxies according to its functional role: DP=q,DC=13(i+p+c),DR=12(r+i).D_P=q, D_C= 13(i+p+c), D_R= 12(r+i). (2) Specifically, DPD_P captures question interpretation difficulty. DCD_C reflects acoustic description difficulty by considering task-relevant evidence, audio quality, and scene complexity. DRD_R characterizes reasoning difficulty through inference demand and acoustic evidence requirements. The continuous difficulties are quantized into five equal-width levels, where bsb_s denotes the corresponding discrete difficulty rating index for section s∈=P,C,Rs =\P,C,R\. bs=min5,⌊5Ds⌋+1b_s= \5, 5D_s +1\ (3) represented by rating tokens 1,…,5 R_1,…, R_5. The resulting target response is y=⟨P⟩⟨bP⟩yP;⟨C⟩⟨bC⟩yC;⟨R⟩⟨bR⟩yR;⟨S⟩yS;⟨A⟩yA.y= P R_b_P y_P; C R_b_C y_C; R R_b_R y_R; S y_S; A y_A. (4) Unlike uniform trace compression, we allocate a fixed trace budget according to the difficulty of different sections. Let B denote the total budget for P, C, R, and S (excluding the final answer), and let BSB_S be the fixed Summary budget. For s∈s , the target budget is Bs=(B−BS)Ds+ϵ∑j∈(Dj+ϵ),ϵ>0,B_s=(B-B_S) D_s+ε _j (D_j+ε), ε>0, (5) where the difficulty floor prevents degenerate zero allocations. We use continuous difficulty scores DsD_s rather than quantized ratings bsb_s for finer-grained budget allocation. Conditioned on (Ds,Bs)(D_s,B_s), the teacher removes redundant steps while preserving each section’s function and consistency with the answer. This allocation assigns more tokens to difficult sections while producing more concise traces for easier ones. Given an audio–question input x and its structured target y, we perform SFT to learn the target response generation. Since each rating is predicted before its corresponding section, it serves as a causal conditioning token rather than a post-hoc annotation. SFT thus establishes the section-wise difficulty interface later optimized by curriculum RL. 2.2 Difficulty-Adaptive Curriculum RL To encourage difficulty-aware reasoning adaptation and improve the efficiency of reasoning generation, we optimize the SFT policy with the clipped Group Relative Policy Optimization (GRPO) objective [7], using group-normalized advantages and KL regularization to the frozen SFT reference. ℒGRPO=ℒclip−βiDKL(πθ∥πSFT),L_GRPO=L_clip- _iD_KL( _θ\| _SFT), (6) where ℒclipL_clip denotes the clipped policy objective with group-normalized advantages. Reward Function. For a sampled response yiy_i, we define ri=rfmt(yi)+rqual(yi)+rlen(yi),r_i=r_fmt(y_i)+r_qual(y_i)+r_len(y_i), (7) where rfmtr_fmt checks the required section and rating-token format, and rqualr_qual uses normalized BLEU for speech translation or answer accuracy for closed-choice tasks. For each section s∈s , let Li,sL_i,s and Ci,sC_i,s denote the generated and compressed DiCoTA reference lengths, respectively. Given the predicted rating bi,s∈1,…,5b_i,s∈\1,…,5\, we define a rating-dependent target length ratio τi,s _i,s, which increases with bi,sb_i,s, and compute the length reward as rlen(yi)=−∑s∈(Li,sCi,s−τi,s)2.r_len(y_i)=- _s ( L_i,sC_i,s- _i,s )^2. (8) Here τi,s _i,s is the target ratio between the generated and compressed reference lengths for section s; higher ratings favor longer generated sections. Feedback-driven curriculum. We use task-quality feedback to gradually introduce harder samples. The samples are ranked by Drms=(DP2+DC2+DR2)/3D_rms= (D_P^2+D_C^2+D_R^2)/3 and divided into KcK_c equal-frequency buckets from easy to hard. The competence score c∈[0,1]c∈[0,1] is a scalar estimate of the model’s mastery of the current bucket, updated from recent reward feedback. Following AdaCuRL [3], training starts with ℬ1B_1 and a competence score c(0)=0c^(0)=0, updated as c(t+1)=clip[0,1](c(t)+(r¯qual(t)−0.5)max1−c(t),γ).c^(t+1)=clip_[0,1]\! (c^(t)+ ( r_qual^(t)-0.5 ) \1-c^(t),γ\ ). (9) Here, r¯qual(t) r_qual^(t) averages the latest M prompt-level quality scores from the hardest active bucket, each computed by averaging the quality rewards of that prompt’s rollouts, and γ sets a minimum update scale. When c≥(k−1)/Kcc≥(k-1)/K_c, we activate the next bucket ℬkB_k and reset c to that threshold. Earlier buckets remain active with exponentially downweighted sampling. Adaptive KL regularization. We adapt the KL coefficient as βi=β0(1−0.5RRMS) _i= _0(1-0.5R_RMS), where RRMSR_RMS aggregates the normalized predicted section ratings. Higher predicted difficulty allows greater deviation from the frozen SFT reference. Rating calibration. To improve the accuracy of predicted difficulty, we calibrate the control ratings against the unquantized offline difficulty. Let μs=5Ds∈[0,5] _s=5D_s∈[0,5] (without rounding) be the unquantized offline difficulty, and define qs(j)∝exp[−(j−μs)2/(2σ2)]q_s(j) [-(j- _s)^2/(2σ^2)] over the integer rating levels j∈1,…,5j∈\1,…,5\, normalized across these levels. We obtain pθ,sp_θ,s by applying softmax only to the logits of the rating tokens 11–55 immediately before the rating digit, conditioned on the audio, question, and preceding generated content. The calibration loss for section s is ℒcal,s=DKL(qs∥pθ,s)L_cal,s=D_KL(q_s\|p_θ,s). The final loss is ℒ=ℒGRPO+0.05∑s∈ℒcal,sL=L_GRPO+0.05 _s L_cal,s. At inference, the model generates ratings, sections, and the answer autoregressively, without external difficulty scoring or reference traces. MMAU-test-mini MMAR Model Easy Medium Hard Sound Music Speech Overall Token Easy Medium Hard Sound Music Speech Overall Token Qwen2.5-Omni-7B [17] 73.87 70.96 61.26 71.47 66.47 68.17 68.7 14.34 60.18 60.36 56.16 58.18 50.00 62.93 58.9 13.18 Public Baselines Audio-Reasoner-7B [16] 69.37 69.16 61.26 66.67 64.67 68.77 66.7 504.36 49.40 43.84 43.84 45.45 39.81 48.98 45.6 537.18 Ke-Omni-R-7B [22] 79.88 75.75 64.86 78.08 70.36 72.07 73.5 76.64 58.38 56.46 49.85 55.15 42.23 58.84 55.0 70.34 Audio-Flamingo 3 [2] 78.68 76.65 65.17 77.48 69.46 73.27 73.4 199.90 61.68 55.26 45.95 49.70 48.06 58.16 54.5 189.10 Omni-CLST [21] 77.78 74.85 63.66 77.48 71.56 67.27 72.1 204.03 60.78 60.06 56.46 63.49 52.45 56.17 59.1 368.01 Omni-R1 [6] 77.18 73.65 68.47 80.48 67.96 70.87 73.2 7.11 66.17 63.66 54.37 58.81 53.46 65.14 61.4 6.82 Our Models CoTA-SFT 75.68 74.25 62.46 71.17 67.37 74.17 70.9 494.58 59.28 57.06 50.76 56.36 45.63 59.86 55.7 519.82 DiCoTA-SFT 74.17 76.35 65.47 70.27 69.76 75.38 71.8 216.18 58.68 56.46 53.45 56.36 48.06 57.82 56.0 222.74 SWEAR 78.68 72.46 69.37 78.38 73.65 68.47 73.5 197.13 62.28 66.07 57.66 62.42 54.85 64.29 62.0 195.85 Table 1: Accuracy (%) across predicted difficulty levels and audio categories on MMAU-test-mini and MMAR. Overall denotes accuracy on the full test set, and Token denotes the average number of generated tokens. The best and second-best results are highlighted in bold and underlined, respectively. Token rankings include only models with explicit CoT. 3 Experiment and Results 3.1 Experimental Setup Training Data. From DiCoTA, we use 402,496 examples for SFT and a separate set of 113,410 reward-compatible examples for RL. The RL set covers speech-to-text transformation and multiple-choice audio reasoning tasks, split by DrmsD_rms into four equal-sized difficulty buckets. Evaluation. We evaluate on the 1,000-example test sets of MMAU-test-mini and MMAR, using answer accuracy as the effectiveness metric and average generated tokens as the efficiency metric. For difficulty, we score each benchmark example using the same procedure as DiCoTA, partition each benchmark independently into 3 equal-sized groups from easy to hard. Answers are extracted by jointly matching the predicted option label and option content. We compare with representative open-source audio reasoning baselines and SFT variants based on CoTA and DiCoTA, denoted as CoTA-SFT and DiCoTA-SFT, respectively. All models follow the same generation and scoring protocol. Training Settings. We use Qwen2.5-Omni-7B as the backbone. SFT is conducted for one epoch with LoRA rank 64, scaling factor 128, a global batch size of 64, and a learning rate of 1×10−51× 10^-5. RL continues LoRA optimization from the SFT checkpoint for 2,000 steps with a learning rate of 1×10−61× 10^-6. Each prompt samples G=6G=6 responses, and the curriculum uses Kc=4K_c=4 difficulty buckets. For the length reward in Eq. (8), we set τi,s=(bi,s+27)/30 _i,s=(b_i,s+27)/30, which increases linearly from 0.930.93 to 1.071.07 as bi,sb_i,s goes from 11 to 55. 3.2 Results and Analysis As shown in Table 1, our full model achieves 73.5% accuracy on MMAU-test-mini, tying the best result, and leads on MMAR with 62.0%. These gains are achieved with only 197.13 and 195.85 generated tokens, reducing token usage by 60.1% and 62.3% relative to CoTA-SFT. These results demonstrate that SWEAR achieves a stronger accuracy-efficiency trade-off than prior explicit-CoT audio reasoning models, including Audio-Reasoner, Audio-Flamingo 3, Omni-CLST, and CoTA-SFT, consistently reducing reasoning tokens while maintaining or improving task performance. EasyMediumHard40405050606070708080DifficultyAccuracy (%)EasyMediumHardDifficulty200200300300400400500500600600Mean output tokens SWEAR DiCoTA-SFT CoTA-SFT Bars: Accuracy (%) Lines: Mean output tokens(a) MMAU(b) MMAR Figure 2: Accuracy (bars) and mean output tokens (lines) across difficulty under DICoTA standard on MMAU and MMAR. Figure 2 exhibits the intended difficulty-aware allocation behavior. As difficulty increases, our full model raises its average output length from 172.8 to 222.5 tokens on MMAU and from 174.4 to 223.2 on MMAR, while remaining 58.3–63.3% shorter than CoTA-SFT and 5.4–16.6% shorter than DiCoTA-SFT across all splits. With this smaller budget, it maintains competitive accuracy on easy MMAU examples and improves hard accuracy by 5.6 points over CoTA-SFT. On MMAR, it consistently outperforms both SFT variants across all difficulty levels. This trend verifies that SWEAR does not apply a uniform reasoning budget, but dynamically adjusts generation length according to estimated difficulty. 3.3 Ablation Studies MMAU MMAR Setting K Acc.↑ Token↓ Acc.↑ Token↓ SWEAR 4 73.5 197.1 62.0 195.9 Component ablations w/o rating calibration 4 70.3 188.5 59.4 193.6 w/o length shaping 4 74.2 216.6 60.4 225.5 w/o adaptive KL 4 71.6 198.4 61.5 197.7 Number of curriculum buckets uniform training 1 69.3 199.0 58.2 188.2 3 buckets 3 70.1 198.7 60.5 191.4 5 buckets 5 71.2 196.0 60.8 195.8 10 buckets 10 70.4 180.2 59.3 185.7 Table 2: Curriculum RL component ablations and sensitivity to the number of curriculum buckets. We further invstigate the impact of different designs and curriculum buckets. As shown in Table 2, removing rating calibration reduces accuracy by 3.2 and 2.6 points on MMAU and MMAR. Removing length shaping increases output length by 19.5 and 29.6 tokens, improving MMAU accuracy by 0.7 points but reducing MMAR accuracy by 1.6 points. Static KL also lowers accuracy on both benchmarks. Among curriculum settings, four buckets achieve the highest accuracy. These ablations confirm that rating prediction, adaptive length control, and curriculum scheduling each contribute to the final performance, validating the design of the proposed training strategy. 4 Conclusion We proposed SWEAR, a structured difficulty modeling framework for adaptive audio reasoning. Building on CoTA, we augment each trace with multi-dimensional difficulty annotations and distill them into the LALM’s self-ratings during SFT. These ratings subsequently guide curriculum RL, adapting both reasoning length and sectional demands. Experiments on MMAU and MMAR show that SWEAR produces concise traces for easy problems while preserving sufficient deliberation for hard ones. Extending this difficulty-aware principle to more diverse datasets and multimodal reasoning domains is our future direction. 5 Compliance with Ethical Standards This is a numerical simulation study for which no ethical approval was required. References [1] J. Fan, R. Ren, J. Li, R. Pandey, P. G. Shivakumar, I. Bulyko, et al. (2026) Incentivizing consistent, effective and scalable reasoning capability in audio LLMs via reasoning process rewards. In Proc. ICLR, Cited by: §1. [2] S. Ghosh, A. Goel, J. Kim, S. Kumar, Z. Kong, S.-g. Lee, et al. (2025) Audio flamingo 3: advancing audio intelligence with fully open large audio language models. In Proc. NIPS, Vol. 38. Cited by: Table 1. [3] R. Li, H. Huang, F. Wei, F. Xiong, Y. Wang, and X. Chu (2025) AdaCuRL: adaptive curriculum reinforcement learning with invalid sample mitigation and historical revisiting. arXiv:2511.09478. Cited by: §2.2. [4] J. Lu, H. Yu, S. Xu, S. Ran, G. Tang, S. Wang, et al. (2025) Prolonged reasoning is not all you need: certainty-based adaptive routing for efficient LLM/MLLM reasoning. arXiv:2505.15154. Cited by: §1. [5] Z. Ma, Z. Chen, Y. Wang, E.-S. Chng, and X. Chen (2025) Audio-CoT: exploring chain-of-thought reasoning in large audio language model. In Proc. ASRU, p. 1–6. Cited by: §1. [6] A. Rouditchenko, S. Bhati, E. Araujo, S. Thomas, H. Kuehne, R. Feris, et al. (2025) Omni-R1: do you really need audio to fine-tune your audio LLM?. In Proc. ASRU, p. 1–7. Cited by: Table 1. [7] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, et al. (2024) DeepSeekMath: pushing the limits of mathematical reasoning in open language models. arXiv:2402.03300. Cited by: §2.2. [8] Y. Shen, J. Zhang, J. Huang, S. Shi, W. Zhang, J. Yan, et al. (2025) DAST: difficulty-adaptive slow-thinking for large reasoning models. In Proc. EMNLP(Industry Track), p. 2322–2331. Cited by: §1. [9] Z. Sheng, S. Zhou, C. Gong, and Z. Li (2026) Think smart, not hard: difficulty adaptive reasoning for large audio language models. In Findings of ACL, p. 32771–32790. Cited by: §1. [10] F. Tian, X. T. Zhang, Y. Zhang, H. Zhang, Y. Li, D. Liu, et al. (2025) Step-audio-R1 technical report. arXiv:2511.15848. Cited by: §1. [11] A. Tjandra, Y.-C. Wu, B. Guo, J. Hoffman, B. Ellis, A. Vyas, et al. (2025) Meta audiobox aesthetics: unified automatic assessment for speech, music and sound. In Proc. ASRU, p. 1–8. Cited by: §2.1. [12] S. Tu, J. Lin, Q. Zhang, X. Tian, L. Li, X. Lan, et al. (2025) Learning when to think: shaping adaptive reasoning in R1-style models via multi-stage RL. In Proc. NIPS, Vol. 38, p. 16989–17015. Cited by: §1. [13] C. Wen, T. Guo, S. Zhao, W. Zou, and X. Li (2025) SARI: structured audio reasoning via curriculum-guided reinforcement learning. arXiv:2504.15900. Cited by: §1. [14] S. Wu, C. Li, W. Wang, H. Zhang, H. Wang, M. Yu, et al. (2026) Audio-thinker: guiding large audio language model when and how to think via reinforcement learning. In Proc. AAAI, Vol. 40, p. 33962–33970. Cited by: §1. [15] W. Xiao and L. Gan (2025) Fast-slow thinking GRPO for large vision-language model reasoning. In Proc. NIPS, Vol. 38, p. 190068–190098. Cited by: §1. [16] Z. Xie, M. Lin, Z. Liu, P. Wu, S. Yan, and C. Miao (2025) Audio-reasoner: improving reasoning capability in large audio language models. In Proc. EMNLP, p. 23829–23851. Cited by: §1, §2.1, Table 1. [17] J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, et al. (2025) Qwen2.5-Omni technical report. arXiv:2503.20215. Cited by: Table 1. [18] J. Xu, Z. Guo, H. Hu, Y. Chu, X. Wang, J. He, et al. (2025) Qwen3-Omni technical report. arXiv:2509.17765. Cited by: §2.1. [19] D. Yang, S. Liu, D. Wang, Y. Wang, G. Wan, and H. Meng (2025) Omni-AutoThink: adaptive multimodal reasoning via reinforcement learning. arXiv:2512.03783. Cited by: §1. [20] J. Zhang, N. Lin, L. Hou, L. Feng, and J. Li (2025) AdaptThink: reasoning models can learn when to think. In Proc. EMNLP, p. 3716–3730. Cited by: §1. [21] J. Zhao, H. Su, L. Fan, Z. Luo, H. Wang, H. Sun, et al. (2025) Omni-CLST: error-aware curriculum learning with guided selective chain-of-thought for audio question answering. arXiv:2509.12275. Cited by: Table 1. [22] S. Zhao, T. Guo, C. Wen, B. Xiang, W. Zou, and X. Li (2025) Ke-Omni-R: achieving advanced audio reasoning with a concise 50-words think process. Note: GitHub repository: https://github.com/shuaijiang/Ke-Omni-RAccessed: 2026-09-04 Cited by: Table 1.