Paper deep dive

When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning

Ruixiang Mao, Xiangnan Ma, Dan Chen, Ziming Zhu, Yuan Ge, Aokai Hao, Haishu Zhao, Yifu Huo, Qing Yang, Kaiyan Chang, Xiaoqian Liu, Chenglong Wang, Qiaozhi He, Tong Xiao, Jingbo Zhu

Year: 2026Venue: arXiv preprintArea: cs.SDType: PreprintEmbeddings: 96

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 7/20/2026, 4:51:53 AM

Summary

This paper investigates the phenomenon of 'audio perception decay' in Large Audio-Language Models (LALMs), where extended reasoning chains lead to degraded audio perception and marginal or negative performance gains. The authors introduce CAFE, an evaluation framework to quantify audio reasoning errors, and propose MPAR^2, a multi-step perception-aware reasoning strategy using reinforcement learning to improve perception accuracy and mitigate decay, achieving state-of-the-art results on benchmarks like MMAU.

Entities (8)

MPAR^2 · method · 98%CAFE · framework · 97%Audio Perception Decay · phenomenon · 96%MMAU · benchmark · 95%MMAR · benchmark · 94%Large Audio Language Models · model-class · 93%GRPO · algorithm · 90%Gemini-3-pro · model · 85%

Relation Signals (5)

MPAR^2 → achievesaccuracyon → MMAU

confidence 95% · achieve a significant 74.59% accuracy on the MMAU benchmark

MPAR^2 → mitigates → Audio Perception Decay

confidence 95% · MPAR^2 ... effectively mitigates perception decay

MPAR^2 → improves → Perception Accuracy

confidence 93% · MPAR^2 improves perception performance on CAFE from 31.74% to 63.51%

CAFE → evaluates → Audio Perception Decay

confidence 92% · CAFE, an evaluation framework designed to precisely quantify audio reasoning errors... reveal LALMs struggle with perception... audio perception decay

MPAR^2 → uses → GRPO

confidence 90% · Leveraging reinforcement learning... leverage Group Relative Policy Optimization (GRPO)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Test-Time Scaling has shown notable efficacy in addressing complex problems through scaling inference compute. However, within Large Audio-Language Models (LALMs), an unintuitive phenomenon exists: post-training models for structured reasoning trajectories results in marginal or even negative gains compared to post-training for direct answering. To investigate it, we introduce CAFE, an evaluation framework designed to precisely quantify audio reasoning errors. Evaluation results reveal LALMs struggle with perception during reasoning and encounter a critical bottleneck: reasoning performance suffers from audio perception decay as reasoning length extends. To address it, we propose MPAR$^2$, a paradigm that encourages dynamic perceptual reasoning and decomposes complex questions into perception-rich sub-problems. Leveraging reinforcement learning, MPAR$^2$ improves perception performance on CAFE from 31.74% to 63.51% and effectively mitigates perception decay, concurrently enhancing reasoning capabilities to achieve a significant 74.59% accuracy on the MMAU benchmark. Further analysis demonstrates that MPAR$^2$ reinforces LALMs to attend to audio input and dynamically adapts reasoning budget to match task complexity.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

95,069 characters extracted from source content.

Expand or collapse full text

When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Ruixiang Mao 1 Xiangnan Ma 1 Dan Chen 1 Ziming Zhu 1 Yuan Ge 1 Aokai Hao 1 Haishu Zhao 1 Yifu Huo 1 Qing Yang 1 Kaiyan Chang 1 Xiaoqian Liu 1 Chenglong Wang 1 Qiaozhi He 1 Tong Xiao 1 2 Jingbo Zhu 1 2 Abstract Test-Time Scaling has shown notable efficacy in addressing complex problems through scaling in- ference compute. However, within Large Audio- Language Models (LALMs), an unintuitive phe- nomenon exists: post-training models for struc- tured reasoning trajectories results in marginal or even negative gains compared to post-training for direct answering. To investigate it, we introduce CAFE, an evaluation framework designed to pre- cisely quantify audio reasoning errors. Evaluation results reveal LALMs struggle with perception during reasoning and encounter a critical bottle- neck: reasoning performance suffers from audio perception decay as reasoning length extends. To address it, we propose MPAR 2 , a paradigm that encourages dynamic perceptual reasoning and de- composes complex questions into perception-rich sub-problems. Leveraging reinforcement learn- ing, MPAR 2 improves perception performance on CAFE from 31.74% to 63.51% and effectively mitigates perception decay, concurrently enhanc- ing reasoning capabilities to achieve a significant 74.59% accuracy on the MMAU benchmark. Fur- ther analysis demonstrates that MPAR 2 reinforces LALMs to attend to audio input and dynamically adapts reasoning budget to match task complexity. 1. Introduction Test-Time Scaling (Snell et al., 2024) has emerged as a dominant paradigm that enables models to solve complex problems by scaling inference compute, typically realized through extended chains of thought (CoT) (Guo et al., 2025; Wei et al., 2022; Brown et al., 2024). Inspired by this suc- cess, recent works have extended similar strategies to Large Audio-Language Models (LALMs) through reinforcement learning (RL), yielding notable gains on complex audio 1 Northeastern University,China 2 NiuTrans Research. Corre- spondence to: Tong Xiao<xiaotong@mail.neu.edu.cn>. Preprint. March 4, 2026. R1-AQA (A)SARI (A)Audio-Thinker (A)Audio-Thinker (O) 50 55 60 65 Negative and Marginal Gains 64.5 61.1 59.9 59.5 63.8 65 69.7 69.8 RLRL † Figure 1. Comparison of direct-answer RL baselines (denoted by RL) and RL with explicit reasoning prompts (denoted by RL † ). Three RL variants (R1-AQA, Audio-Thinker, and SARI) are tested on the MMAU benchmark. Here, the suffix (A) and (O) indicate Qwen2-Audio and Qwen2.5-Omni as base model. More details of the prompts used here are presented in Appendix B.3. tasks that demand deep audio understanding and reason- ing (Tian et al., 2025; Wu et al., 2025). However, a marked exception persists: optimizing models for structured reason- ing trajectories via RL results in marginal or even negative gains compared to direct-answering tuning. As illustrated in Figure 1 on the MMAU benchmark (Sakshi et al., 2024), direct-answer RL baselines, which are trained via implicit prompts, outperform CoT-based variants like R1-AQA (Li et al., 2025) and SARI (Wen et al., 2025), where explicit rea- soning prompts are utilized to generate<thinking>con- tent. Meanwhile, Audio-Thinker (Wu et al., 2025) achieves comparable results. This phenomenon differs from the infer- ence scaling anticipation that increasing inference compute yields predictable performance improvement. Consequently, this observation substantially challenges increasing infer- ence time compute in LALMs reasoning. In light of this, we raise a question: what causes the inefficacy of extended reasoning in complex audio tasks? Recent research has attempted to tackle similar challenges. Step-Audio-R1 (Tian et al., 2025) attributes this issue to an over-reliance on textual reasoning, exploring modality- grounded frameworks to mitigate the gap. Meanwhile, an analysis in MMAR (Ma et al., 2025b) reveals audio event perception and utilization errors are the dominant factor. 1 arXiv:2603.02266v1 [cs.SD] 28 Feb 2026 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning However, despite these efforts, a systematic empirical inves- tigation of this problem is still rare. To fill this void, we introduce CAFE, a Comprehensive Audio Fidelity Evaluation framework designed to accurately quantify audio reasoning errors. Leveraging the “LLM-as-a- Judge” paradigm (Zheng et al., 2023), CAFE extracts and counts audio events from reasoning processes to quantify perception and utilization errors. Experimental results re- veal that audio perception errors pose substantial challenges, and the audio perception decays as reasoning length in- creases, with perceptual and reasoning performance declin- ing simultaneously. Motivated by these results, we propose a hypothesis: LALMs’ perceptual capabilities do not scale alongside reasoning length. This misalignment explains why RL training for extended CoT harms, rather than helps audio model reasoning performance. To address this limitation, we propose MPAR 2 , a Multi-step Perception-Aware Reasoning and Review strategy designed to improve audio reasoning performance. Extending the reasoning power of CoT, MPAR 2 establishes a “think-while- listening” paradigm. It implements through a structured pipeline: initially enforcing a fine-grained scan for explicit event perception, followed by step-wise perception-aware decomposition to tackle complex questions, and conclud- ing with a post-reasoning review to validate event usage and ensure accuracy. To realize this strategy, we adopt a two-stage training scheme. In the first stage, supervised cold-start training teaches the model structured reason- ing. In the second stage, RL further optimizes the model for audio perception and reasoning. Experiments shows MPAR 2 achieves 74.59% on MMAU (origin) test-mini and 60.32% on MMAR. Under the CAFE evaluation, MPAR 2 achieves state-of-the-art (SOTA) perception and utilization accuracy. Further analysis demonstrates that MPAR 2 re- inforce LALMs to attend to audio input during reasoning and dynamically adjusts reasoning budget based on task complexity. In summary, our contributions are 1 : • We present the first systematic analysis of audio percep- tion decay in LALMs, where longer reasoning chains lead to degraded audio perception, resulting in marginal or even negative gains from complex CoT reasoning. •We introduce CAFE, an evaluation framework that pre- cisely quantifies audio reasoning error. Experimental results validate the audio perception decay phenomenon and pose a challenge for existing audio models. •We propose MPAR 2 , an RL-based two-stage training strategy designed to enhance both audio events perception and utilization. Experiments show that MPAR 2 achieves SOTA perception accuracy, and exhibits reinforced audio attention during reasoning and adaptive reasoning budget. 1 The code for the framework and training can be accessed via https://github.com/Moriiikdt/MPAR2 2. Related Work Large Audio Language ModelsAs large language mod- els (LLMs) continue to advance, models that integrate multi- ple modalities have gradually gained growing focus. These multimodal LLMs are capable of jointly modeling and rea- soning over various inputs, including audio. Recent LALMs, such as Qwen2-Audio (Chu et al., 2024), Qwen2.5Omni (Xu et al., 2025), DeSTA2.5-Audio (Lu et al., 2025), and MiMo- Audio (Xiaomi, 2025), have demonstrated notable effective- ness in understanding and processing acoustic information. Large Audio Reasoning ModelsRecent research has in- creasingly focused on enhancing LALMs with reasoning capability. While early efforts like Audio-CoT (Ma et al., 2025a) and Audio-Reasoner (Xie et al., 2025) introduced CoT via prompting or supervised fine-tuning (SFT), but achieving limited capability gains, failing to fully unlock the reasoning potential of audio models. Subsequently, rein- forcement learning has recently pushed boundaries. Meth- ods such as R1-AQA (Li et al., 2025) and SARI (Wen et al., 2025) utilize reinforcement learning to optimize structured CoT reasoning, whereas Omni-R1 (Rouditchenko et al., 2025) adopts a direct-answer strategy that surprisingly sur- passes CoT. However, results on the variants of R1-AQA and SARI reveal an unintuitive behavior in audio test-time scal- ing. Although other approaches like Audio-Thinker (Wu et al., 2025) explore a adaptive thinking mode and Step- Audio-R1 (Tian et al., 2025) apply modality-grounded frameworks to futher impove reasoning capability, yielding substantial improvements in reasoning performance, none have explicitly investigated this unintuitive phenomenon or proposed targeted solutions to address it. Adaptive Reasoning and Difficulty AwarenessThe suc- cess of reasoning models such as OpenAI o1 (Jaech et al., 2024) demonstrates the effectiveness of test-time scaling, while also highlighting the need to reduce computational redundancy. Prior work shows that LLMs implicitly encode problem difficulty in their hidden representations (Lee et al., 2025), enabling adaptive reasoning termination. Building on this insight, S-GRPO (Dai et al., 2025) employs reinforce- ment learning with decaying rewards to encourage early correct exits, while data-centric methods (Waheed et al., 2025) improve token efficiency by distilling difficulty-aware reasoning length through compressed CoT trajectories. 3. Reasoning-Time Audio Event Perception and Utilization Probing In this section, we define the task of reasoning-time audio event perception and utilization probing and introduce the CAFE framework. Subsequently, we present and analyze evaluation results across LALMs, Large Audio Reasoning Models (LARMs), and commercial models. 2 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Reasoning Process Caption Text Gemini-3-pro Audio Event Extraction Question&Answer Audio Audio Model Error-used Error-matched Matched Missed Neutral Figure 2. Overview of the CAFE framework evaluation pipeline. First, audio model generates a reasoning process based on the audio-QA input. Meanwhile, the audio is captioned by Gemini- 3-Pro. Finally, Gemini-3-Pro utilizes the inputs, the reasoning process, and the captions to extract different audio events. Task FormulationGiven ground-truth audio content and Question-Answer (QA) pairs, the reasoning-time audio event perception and utilization probing task aims to quan- tify the identification and utilization of audio events during model reasoning, employing metrics such as perception accuracy, utilization accuracy, and omission rate. We lever- age a powerful textual LLM to process the probing inputs. This approach enables robust generalization across semantic variations, while ensuring the precise extraction of audio events strictly required for the question-response context. Furthermore, it enables the quantification of audio fidelity throughout statistically analyzing the extracted audio events. 3.1. Overview To address this task, we introduce the CAFE framework. Formally, employing the probing textual modelP, the audio event extraction process can be formulated as: E =P(O,Q,A,C)(1) whereOrepresents the model’s reasoning output gener- ated by the model,QandAdenote the input question and the target answer, andCis the ground truth audio caption. The same abbreviations are used throughout the subsequent formulas. The resulting outputEconsists of categorized audio events with labelsE i , where aspect E i ∈ Matched, Hallucinated, Misused, Missed, Neutral, defined as follows: (1) Matched: existing audio events cor- rectly identified and utilized during the reasoning process to address the question. (2) Hallucinated: events that are hal- lucinatory or misidentified during reasoning; (3) Misused: existing events that are logically misused, leading to incor- rect conclusions; (4) Missed: existing events required for the answer but omitted during reasoning; and (5) Neutral: existing events mentioned but irrelevant to the reasoning out- come. Furthermore, we extract and count these categorized events to derive specific metrics below. Metric Formulation We define four distinct metrics based on the extracted event categories. LetN (·) denote the count of events for each category. We first define the prediction space N pred and the target space N tgt as follows: N pred = N mat + N hal + N misuse + N neu , N tgt = N mat + N miss (2) whereN pred donates the total number of audio events in model reasoning process, andN tgt represents the complete set of ground-truth events required to answer the question. Based on these, we derive the following metrics: •Perception Accuracy (Acc per ) is defined as the ratio of correctly identified events related to the QA task, formulated asAcc per = N mat N pred .Acc per is aimed at per- ception in reasoning rather than simple audio caption. •Perception Error Rate (Err per ) quantifies the rate of hallucinatory or misidentified events viaErr per = N hal N pred . •Usage Error Rate (Err use ) assesses events that are cor- rectly perceived but logically misused, which can be formulated as Err use = N misuse N pred . •Omission Error Rate (Err omit ) quantifies the events required for the answer that the model failed to capture, defined as Err omit = N miss N tgt . In particular,Acc per offers a holistic assessment of the audio model’s perceptual fidelity during reasoning. This metric serves as the quantitative basis for Figure 3, where we track perceptual performance as reasoning length increases, ulti- mately revealing the phenomenon of audio perception decay. Framework details Within CAFE, Gemini-3-pro gener- ates fine-grained acoustic descriptions for the benchmark’s audio and functions as the probing modelP, responsible for extracting the reasoning results. The overall workflow of the framework is shown in Figure 2. All prompts for captioning and extraction are provided in Appendix A. 3.2. Evaluation Experimental SetupWe evaluate a diverse set of models, including commercial models, open-source LALMs, and LARMs. For LALMs and commercial models, which lack explicit reasoning output, we employ designed prompts to 3 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 4080120160 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (a) Audio-Flamingo-3 224272320368416464512560608656 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (b) Audio Reasoner 20406080100120140160180200 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (c) DeSTA2.5-Audio 2060100140180220260300340380 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (d) MiMo-Audio 1262112162212262312362412462 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (e) Gemini-2.5-Flash 12376287112137162187212237 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (f) GPT-4o-Audio Figure 3. Reasoning Accuracy (accuracy on the evaluated benchmarks) and CAFE’s perception accuracy across reasoning token lengths. Reasoning accuracy aligns closely with perception accuracy. Specifically in the (a)-(d), both metrics decline as reasoning tokens increase, which we term Audio Perception Decay. The x-axis represents the mean of token length intervals, and the y-axis indicates the accuracy. get a reasoning process before the final answer. A unified prompt is used for the remaining models. All evaluation prompts and details are presented in Appendix B. 3.3. Evaluation Result To analyze the relationship between reasoning length and both perception and reasoning accuracy, we selected two rep- resentative models from each category: LALMs, LARMs, and commercial models, as shown in Figure 3. Table 1 presents the overall results, including the MPAR 2 method, which is detailed in Section 4. To further investigate, we conduct additional analysis and identify two key findings. Weak Perception during ReasoningAs shown in Table 1, most audio models perform poorly on CAFE for MMAR. Notably, open-source models rarely surpass the threshold of 40%Acc per , and even advanced systems like Gemini-2.5- Flash reach only 54.81%, with the majority falling below 30%. However, the implementation of curated RL strategies proves to significantly enhance capabilities. Specifically, Omni-R1 (51.21%) substantially outperforms its Qwen2.5- Omni baseline (31.74%), while Step-Audio-R1.1 (56.18%) and Step-Audio-R1 (53.08%) similarly achieve remarkable performance. For instance, Step-Audio-R1 series of mod- els achieves high accuracy by maintaining a well-balanced across all four metrics, effectively preventing any bottle- neck. In contrast, other models exhibit obvious variances, underscoring that reasoning capability is critically sensitive to even partial perceptual deficits. These findings highlight severe weakness in reasoning-time audio perception, along with degraded reasoning capability. Audio Perception Decay Across Extended Reasoning As illustrated in Figure 3 (a)-(d), open-source LARMs Table 1. Results of various audio models on the MMAR benchmark within the CAFE framework. † denotesMMAR acc evaluated by GPT-5 judge. All results are averaged over three inference runs. The best performance in each group is in bold and the second best one is underlined. The remaining table settings are the same. ModelAcc per ↑ Err per ↓ Err use ↓ Err omit ↓ MMAR acc ↑ Large Audio Language Models MiMo-Audio42.7927.7022.4048.8459.87 Qwen2.5-Omni-7B31.7429.6128.6957.3755.20 DeSTA2.5-Audio † 23.1933.9535.5769.3241.60 Phi-4-Multimodal20.4336.7530.3273.4139.80 Qwen2-Audio-Instruct8.2756.4628.5881.0129.90 Omni-R151.2128.7912.9042.6562.10 Large Audio Reasoning Models Step-Audio-R1.156.1823.7711.2521.4967.50 Step-Audio-R153.0824.8212.426.7767.40 Audio-Flamingo-3 † 41.3926.0323.1852.8756.40 Audio Reasoner27.4425.5532.5855.7236.80 Commercial Models Gemini-2.5-Flash54.8119.2715.9633.2466.30 GPT-4o-Audio48.6817.2021.7138.0563.80 MPAR 2 -7B63.5123.147.7430.5960.32 and LALMs exhibit a clear downward trend: both percep- tion and reasoning accuracy decline as reasoning length increases, eventually dropping to zero. Pearson (Benesty et al., 2009) analysis confirms a robust positive correlation between these two metrics across (a)-(d) groups, yielding coefficients ranging from 0.65 to 0.91, where the majority reach statistical significance withp < 0.05. In contrast, commercial models (e)-(f) maintain stable and superior per- formance without exhibiting decay trend. Consequently, the observed variations in these metrics validate our hypothesis that perception struggle to maintain a robust connection with the expanding reasoning chain. Instead, perception accu- racy degrades as reasoning length increases, which we term “Audio Perception Decay”, leading to an accumulation of errors that drives the degradation of reasoning capability. 4 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Cold Start Policy LALM Audio Query Reward Function Perception Reward Group of Sampled Reasoning Paths p 2 p N PerceptionSub- 퐀 퐀 퐀 Sub- 퐀 퐀 퐀 ... Review 1 Answer 1 p 1 Perception Review N Answer N Perception Sub- 퐀 퐀 퐀 Sub- 퐀 퐀 퐀 ... Review 2 Answer 2 ... Group Advantage A 1 ... A 2 A N Rewards r 1 r N r 2 ... Geometric Mean 퐀 퐀倀퐀 Score ... Sub-Question:What acts as the primary sound...? Answer:The primary source of... Score Sub-Question:What specific characteristics of ...? Answer:The sound includes many... Score Sub-Question:Now we can answer the question: Answer:Based... the answer is Parrot. Score <Perception> The audio begins with a sound... < > 퐀 倀䐀퐀᠀䐀倀퐀က퐀᠀ <Review> Evidence check: The reasoning accurately reflects the... Logic check: The logical flow is sound. The reasoning... </Review> <Answer> Answer </Answer> 퐀 퐀䐀퐀 =퐀∙ 퐀+퐀 ∙퐀 퐀䐀퐀က䐀尀 퐀 퐀䐀퐀 =퐀∙ 퐀+퐀∙퐀 퐀䐀퐀က䐀尀 Stepwise Perception-Reasoning Reward ( R spr ) Perception Reward Review-Enhanced Accuracy Reward ( R rea ) Final Reward 퐀 퐀ࠀ =퐀∙퐀 倀䐀퐀᠀䐀倀퐀က퐀᠀ +퐀∙퐀 퐀倀퐀 +᠀∙퐀 퐀䐀퐀 +퐀∙퐀 퐀䀀퐀 Review Enhenced Sub- 퐀 퐀 퐀 Sub- 퐀 퐀 퐀 ... Score GATE Chain Consistency Score Stepwise Validity Score SPR Reward REA Reward Format Reward Figure 4. An illustration of MPAR 2 training pipeline. The top of the figure outlines the overall GRPO training setup, whereas the lower section provides a detailed view of the reward design. The cold start policy model is derived from the Stage 1 SFT-trained model. 4. Methodology To address the perception decay of reasoning over extended sequences, we propose a novel reasoning paradigm, Multi- step Perception Aware Reasoning and Review strategy (MPAR 2 ), which is designed to structurally reinforce audio perception and reasoning while maintaining compatibility with inference-time scaling strategies. In this section, we present the MPAR 2 framework, formalize its design, and descirbe its two-stage training procedure. The overview of MPAR 2 is illustrated in Figure 4. 4.1. Cold Start for MPAR 2 Framework Prompt Design We begin by defining the reasoning pro- cess under the MPAR 2 paradigm. Each reasoning chain follows three stages: (1) Perception: Explicitly describe the audio events related to the QA task. Detail their inter- relationships, strictly following the time sequence; (2) Rea- soning: Dynamically decompose the original question into a structured sequence of sub-questions, adjusting the sub- question steps based on the question’s difficulty. Then, solve the sub-questions using audio-perception details and reason- ing, until the core inquiry is fully resolved. (3) Review: Double-check the fidelity of the audio events used during reasoning, and ensure the logical consistency of the entire reasoning process. Make necessary corrections to the final answer if errors are detected. Data Construction To teach LALMs this structured rea- soning pattern and prepare data for RL, we draw from the AVQA dataset (Yang et al., 2022), which is designed for video question answering. We first employed Gemini-2.5- Pro to generate detailed captions for the AVQA dataset. Based on these captions, Qwen3-32B 2 (Yang et al., 2025) synthesized reasoning-oriented QA pairs across four dimen- sions: timbre characteristics, temporal patterns, pitch con- tours, and rhythmic structures. Then, we filter the QA pairs for logic and consistency, and assess difficulty by running the initial base model 16 times, discarding samples that were completely correct or incorrect, resulting in 46,544 QA samples. From the filtered data pool, we curated a balanced set of 5,000 samples across four aspects and dif- ferent time durations, while the remaining 41,544 samples Data RL were reserved for Stage-2 RL. Finally, we prompt to use these 5,000 samples to generate CoT reasoning paths by Qwen3-32B. Following CoT quality filtering, 4,600 in- stances, denoted asData SF T , were selected for Stage-1 supervised fine-tuning (SFT). Detailed training processes, generation procedures, prompts, and four dimensions’ case studies are presented in Appendix E. 4.2. Reinforcement Learning with Generative Rewards To both refine perception quality and promote adaptive rea- soning budgets within the MPAR 2 paradigm, we leverage Group Relative Policy Optimization (GRPO) (Shao et al., 2024). This optimization process is steered by three key generative rewards and format reward, which jointly work to enhance both reasoning and perceptual fidelity. Specially, we utilize a powerful textual LLMSas a score evaluator to get process rewards, guided by carefully designed prompts. The prompt designs and other implementation details are 2 https://huggingface.co/Qwen/Qwen3-32B 5 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning provided in Appendix F. Perception Reward Given the severe perceptual limita- tions and audio perpception decay during reasoning, MPAR 2 begins the reasoning process with an explicit audio percep- tion step. It incorporates a perception reward to ensure high-precision estimation, as shown in Figure 4. Formally, using the scoring model S, the reward is defined as: R perception = S(O perception ,Q,A,C)(3) whereO perception represents the perception part of model’s reasoning output. Specifically, given the ground truth cap- tion with QA, we prompt Qwen3-32B as the modelSto assess the generated perception. The model outputs a score R perception ∈ [0, 1]as reward with a specific focus on the audio events essential for solving the problem, evaluating them based on faithfulness, completeness, and consistency. Stepwise Perception-Reasoning Reward Prior stud- ies (Marjanovi ́ c et al., 2025; Lee & Hockenmaier, 2025) demonstrate that structured reasoning trajectories, such as problem decomposition, significantly enhance complex problem-solving. Building on this, S-GRPO (Dai et al., 2025) further optimizes efficiency by employing RL with decaying rewards to incentivize early correct exits. Mean- while, existing approaches such as Audio-Thinker (Wu et al., 2025) primarily rewards the overall reasoning process in a coarse manner, but lacking fine-grained guidance over reasoning quality and explicit emphasis on audio percep- tion. Inspired by these findings, we introduce Stepwise Perception-Reasoning Reward (SPR Reward) to reinforce structured reasoning. SPR Reward using two key criteria: stepwise validity and chain consistency. Specifically, stepwise validity evaluates the event completeness and logical soundness of each sub- QA. It aims for high efficiency and direct resolution within the current step, ensuring that every question posed is both critical and perception-rich. We calculate these step scores using the geometric mean (McAlister, 1879), which is de- scribed in Eq. 5, a method chosen to enforce uniform quality by strictly penalizing any single step failure. Moving beyond individual steps, chain consistency evalu- ates the logic of the entire reasoning path. Crucially, this mechanism encourages the model to adaptively adjust its sub-question decomposition based on question difficulty, thereby effectively controlling the reasoning budget. It re- quires to construct a clear context via detailed audio de- scriptions and analysis, while enforcing a strict order where audio analysis precedes and supports final problem-solving. With these two criteria, we define SPR Reward as: R spr = θ· ̄ S sub-reason + (1− θ)· S all-reason (4) ̄ S sub-reason and S all-reason are calculated by: ̄ S sub-reason = n Y i=1 S O (i) sub-reason ,Q,A,C ! 1 n (5) S all-reason = S O all-reason ,Q,A,C (6) whereθ ∈ [0, 1]allocates the reward weights between each reasoning steps and entire reasoning path, which is set to 0.7. O (i) sub-reason denotes thei-th sub-reasoning step, andO all-reason is the overall reasoning content. ModelSis Qwen3-32B. Notably, for stepwise validity and chain consistency eval- uation, in addition to the reasoning content, we also input the audio captionC, the questionQ, and the ground-truth answer A, thereby improving the evaluation quality. Review-Enhanced Accuracy Reward Current explicit reasoning in audio models is inherently one-pass. Once the model generates reasoning steps, it continues without looking back to re-examine the alignment between the audio content and its text. Consequently, if an initial perceptual error occurs (e.g., misidentifying a “cat” as a “dog”), sub- sequent reasoning steps merely reinforce this error rather than correcting it. This lack of an “auditory feedback loop” necessitates a mechanism for backward verification. To address this, we introduce the Review-Enhanced Accuracy Reward (REA Reward), which integrates this double-check with a direct reward for the correctness of the final answer. Leveraging the Qwen3-32B as the modelS, we calculate it: R rea = R acc · (1 + μ· R review )(7) where we set μ to 0.5, and R review is defined as: R review = S(O review ,O all-reason ,O perception ,Q,A,C)(8) whereO review denotes the review part of model ouput.R acc is defined as a binary reward, taking the value of 1 if the generated answer aligns with the ground truth, and 0 oth- erwise. Notably, we treatR acc as a gating factor to prevent that review process is rewarded but the answer is still wrong, as show in Eq. 7. This ensures that the review reward is only granted when the final answer is correct. Format Reward We define a format rewardR format of 1 for outputs complying with the MPAR 2 structure illustrated in Figure 4, and 0 otherwise. Overall Reward By assigning specific weights to each component, the final reward R all is defined as follows: R all = α· R perception + β· R spr + γ· R rea + δ· R format (9) where we setα,β,γ,δto 1.5, 1.0, 1.5, and 0.1, respectively. 6 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Table 2. Accuracy (%) on MMAU (original) and MMAR benchmark. † denotes results evaluated by GPT-5. * means results are reproduced, while all other results are averaged over three inference runs. ‡ means the ablation setting with a simple accuracy reward. Model MMAU (Test-mini)MMAR Sound↑Music↑Speech↑Avg↑Sound↑Music↑Speech↑Avg↑ Large Audio Language Models MiMo-Audio80.7870.9663.0671.6058.7943.6960.5459.87 Qwen2.5-Omni-7B69.4066.8061.6065.9053.9442.2358.8455.20 DeSTA2.5-Audio † 58.2644.6158.2653.7053.7424.7629.7041.60 Phi-4-Multimodal57.9651.8048.3552.7032.1230.5845.5839.80 Qwen2-Audio-Instruct49.8552.4042.9448.4033.3324.2732.3129.90 Omni-R173.6074.3066.1071.3067.3051.5064.3062.10 Kimi-Audio61.6873.2760.6665.0048.4829.1342.5240.60 Large Audio Reasoning Models Audio-Flamingo-3 † 79.58 72.4660.3670.8058.2644.6158.2656.40 Audio Reasoner ∗ 60.0664.3060.7061.7143.6433.5032.9936.80 SARI ∗ 72.7567.2261.2667.08---- R1-AQA ∗ 68.7764.3763.6665.60---- Audio-Thinker ∗ 77.4870.3673.3773.7068.3253.8864.2965.30 Step-Audio-R172.9761.6874.47 69.7057.5845.6377.8967.40 Step-Audio-R1.172.6772.4677.1874.1060.0045.1579.2567.50 Commercial Models Gemini-2.5-Flash67.5661.4863.7664.2753.3350.9775.8566.30 GPT-4o-Audio60.3657.1967.2761.6054.5548.5469.3963.80 MPAR 2 -7B77.9871.3074.4774.5962.4044.3064.2660.32 w/o R perception 75.8472.9071.3473.3660.1042.6063.8258.84 w/o R spr 77.7470.6873.7874.0861.9043.8063.1659.62 w/o R spr and R perception 77.1970.6671.4373.1159.6042.4062.6058.20 w/o R spr , R perception and R rea ‡ 77.5467.3273.2772.8158.9041.8061.8657.52 5. Experiment 5.1. Model Setup In our experiments, we adopt the Qwen2.5-Omni with 3B and 7B parameters (Xu et al., 2025) as the base models. During the SFT phase, base model was trained for 3 epochs on theData SF T with a total batch size of 16 and an initial learning rate of 5e-5. In the GRPO phase, we use a node with 4 H100 GPUs. The batch size per GPU is 1 with gradient accumulation steps of 2 for a total effective batch size of 8. We use a learning rate of 1e-5, a temperature of 1.0, 8 responses generated per sample at each GRPO optimization step. The reward model is applied via VLLM. 5.2. Main Results We primarily evaluate model performance on the MMAU and MMAR benchmarks, and further conduct reasoning analysis under the CAFE framework, with benchmark de- tails provided in Appendix C. The MMAU-v05.15.25 and 3B model’s results in Appendix D. From the experimental results in Figure 5, Tables 1 and 2, we conclude: MPAR 2 Stably Improves Audio Reasoning Performance As shown in Table 2, compared to the Qwen2.5Omni-7B baseline, MPAR 2 achieves substantial performance gains. Specifically, MPAR 2 boosts accuracy on MMAU from 30507090110130150170190210 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy Qwen2.5-Omni-7B (Base) 224272320368416464512560608656 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy MPAR 2 -7B Figure 5. Accuracy of MPAR 2 -7B vs. Base across token lengths. 65.90% to 74.59% and on MMAR from 55.20% to 60.32%. Notably, stable improvements are observed across distinct audio domains, including Sound, Music, and Speech. Such consistent improvements across diverse domains not only underscore the robustness of our approach but also support our core insight: enhanced audio perception and multi-step reasoning are pivotal for boosting performance. MPAR 2 Facilitates Effective Perception RefinementAs demonstrated in Table 1 and Figure 5, MPAR 2 exhibits ro- bust perceptual capabilities and effectively addresses the challenge of perception decay. Specifically, in Figure 5, MPAR 2 successfully mitigates the trend of perception decay associated with increasing reasoning length, maintaining both perception and reasoning accuracy at consistently high and stable levels. In Table 1, MPAR 2 achieves SOTA perfor- mance among the evaluated models in terms of perception 7 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 20406080100120 0.00 0.02 0.04 0.06 0.08 0.10 Audio Attention Ratio Higher Attention Attention Wave Qwen2.5-Omni-7B (Base) MPAR 2 -7B Figure 6. Audio attention ratio across reasoning token length. accuracyAcc per and event utilization errorErr per , highlight- ing the potential of our method. 6. Analysis Ablation Study We show the impact of different reward ablation in Table 2, demonstrating that each mechanism contributes positively to MPAR 2 . First, relying solely on format and simple accuracy rewards already yields substan- tial gains of 6.91% on MMAU and 2.32% on MMAR over the baseline, validating the effectiveness of the MPAR 2 pattern. Second, incorporating the review-enhanced accu- racy mechanism leads to further improvements across all metrics, underscoring the robustness and efficacy of the re- thinking process. Finally, the integration of perception and multi-step reasoning rewards specifically boosts perception capabilities, elevating MMAU performance from 73.36% and 74.08% to 74.59%. Notably, the perception reward con- tributes a relatively larger gain, highlighting the critical role of perceptual grounding in audio modal reasoning. MPAR 2 Strengthens Audio-Focused Attention during Reasoning We further investigate how MPAR 2 training reshapes the model’s attention distribution over audio to- kens. Specifically, we calculate the attention ratios between generated tokens and input audio. As illustrated in Figure 6, we visualize the attention dynamics by plotting the attention ratios of the base model and MPAR 2 across reasoning token length. The results indicate that MPAR 2 maintains a consis- tently higher attention level on audio compared to the vanilla Qwen2.5-Omni. Notably, attention peaks during the early reasoning stage, driven by the initial perceptual process- ing. Following this peak, rather than decaying, the attention exhibits wave motions, reflecting the model’s continuous retrieval with audio context to address perception-related sub-questions. This pattern suggests that MPAR 2 enhances low-level grounding at the input stage while also reinforcing audio-conditioned reasoning in the later decoding process. Together, these results indicate that MPAR 2 encourages the model to more effectively integrate acoustic evidence throughout the perception–reasoning structure. MMAU (Test-mini) MMAR 400 440 480 520 560 600 640 Reasoning Length Gap ∆ = 55 ∆ = 53 ∆ = 65 ∆ = 61 EasyInformation ExtractionAcoustic Mixtures MediumReasoningPerception HardSingle-SourceHigh-level Cognitive Figure 7. Reasoning token length of MPAR 2 -7B across different task categories on the MMAU and MMAR benchmarks. The y- axis represents the number of tokens. Adaptive Reasoning Budget across Task Complexities As illustrated in Figure 7, we calculated the reasoning token lengths of MPAR 2 across different aspects on the MMAR and MMAU benchmarks. Notably, since MMAR focuses on deep reasoning problems, the model exhibits a longer overall output length on this benchmark compared to MMAU. Be- yond this dataset-level difference, we observe that straight- forward inputs, such as single-source audio, easy-level ques- tions, or basic information extraction tasks, result in signifi- cantly shorter reasoning paths. In contrast, complex acoustic mixtures, hard-difficulty scenarios, and high-level reasoning tasks result in obviously longer responses. This behavior suggests that our MPAR 2 training strategy successfully em- powers the model to dynamically adjust its reasoning budget rather than following a fixed long CoT pattern. By dynami- cally condensing the inference path for simple tasks while extending it for complex scenarios, the model effectively balances computational efficiency with reasoning quality. 7. Conclusion In this work, we investigate the unintuitive phenomenon where models RL for structured reasoning trajectories re- sults in marginal or even negative gains compared to RL for direct answering. Through the proposed CAFE framework, we quantify this issue and identify the critical bottleneck: as reasoning scales in LALMs, reasoning capability suffer from audio perception decay. To address this limitation, we propose MPAR 2 . Experimental results demonstrate MPAR 2 ’s consistent performance improvements across di- verse benchmarks, as well as the positive contributions of distinct reward components during training. Furthermore, in-depth analysis confirms that MPAR 2 strengthens audio attention during reasoning and capable of dynamically ad- justing reasoning budget based on task complexity. 8 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 8. Acknowledgements This work was supported in part by the National Science Foundation of China (Nos.62276056 and U24A20334), the Yunnan Fundamental Research Projects (No.202401BC070021), the Yunnan Science and Technol- ogy Major Project (No. 202502AD080014), the Funda- mental Research Funds for the Central Universities (Nos. N25BSS054 and N25BSS094), and the Program of In- troducing Talents of Discipline to Universities, Plan 111 (No.B16009). 9. Impact Statement This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here. References Abouelenin, A., Ashfaq, A., Atkinson, A., Awadalla, H., Bach, N., Bao, J., Benhaim, A., Cai, M., Chaudhary, V., Chen, C., et al. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture- of-loras. arXiv preprint arXiv:2503.01743, 2025. Benesty, J., Chen, J., Huang, Y., and Cohen, I. Pearson correlation coefficient. In Noise reduction in speech pro- cessing, p. 1–4. Springer, 2009. Brown, B., Juravsky, J., Ehrlich, R., Clark, R., Le, Q. V., R ́ e, C., and Mirhoseini, A. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787, 2024. Chang, K., Xu, S., Wang, C., Luo, Y., Liu, X., Xiao, T., and Zhu, J. Efficient prompting methods for large language models: A survey. arXiv preprint arXiv:2404.01077, 2024. Chu, Y., Xu, J., Yang, Q., Wei, H., Wei, X., Guo, Z., Leng, Y., Lv, Y., He, J., Lin, J., et al. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759, 2024. Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. Dai, M., Yang, C., and Si, Q. S-grpo: Early exit via rein- forcement learning in reasoning models. arXiv preprint arXiv:2505.07686, 2025. Goel, A., Ghosh, S., Kim, J., Kumar, S., Kong, Z., Lee, S.-g., Yang, C.-H. H., Duraiswami, R., Manocha, D., Valle, R., et al. Audio flamingo 3: Advancing audio intelligence with fully open large audio language models. arXiv preprint arXiv:2507.08128, 2025. Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. Hurst, A., Lerer, A., Goucher, A. P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Car- ney, A., et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. Lee, J. and Hockenmaier, J. Evaluating step-by-step reason- ing traces: A survey. arXiv preprint arXiv:2502.12289, 2025. Lee, S., Yin, Q., Leong, C. T., Zhang, J., Gong, Y., Ni, S., Yang, M., and Shen, X. Probing the difficulty perception mechanism of large language models. arXiv preprint arXiv:2510.05969, 2025. Li, G., Liu, J., Dinkel, H., Niu, Y., Zhang, J., and Luan, J. Reinforcement learning outperforms supervised fine- tuning: A case study on audio question answering. arXiv preprint arXiv:2503.11197, 2025. Lu, K.-H., Chen, Z., Fu, S.-W., Yang, C.-H. H., Huang, S.-F., Yang, C.-K., Yu, C.-E., Chen, C.-W., Chen, W.-C., Huang, C.-y., et al. Desta2. 5-audio: Toward general- purpose large audio language model with self-generated cross-modal alignment. arXiv preprint arXiv:2507.02768, 2025. Ma, Z., Chen, Z., Wang, Y., Chng, E. S., and Chen, X. Audio-cot: Exploring chain-of-thought reasoning in large audio language model. arXiv preprint arXiv:2501.07246, 2025a. Ma, Z., Ma, Y., Zhu, Y., Yang, C., Chao, Y.-W., Xu, R., Chen, W., Chen, Y., Chen, Z., Cong, J., et al. Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix. arXiv preprint arXiv:2505.13032, 2025b. Marjanovi ́ c, S. V., Patel, A., Adlakha, V., Aghajohari, M., BehnamGhader, P., Bhatia, M., Khandelwal, A., Kraft, A., Krojer, B., L ` u, X. H., et al. Deepseek-r1 thoughtol- ogy: Let’s think about llm reasoning. arXiv preprint arXiv:2504.07128, 2025. 9 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning McAlister, D. Xiii. the law of the geometric mean. Pro- ceedings of the Royal Society of London, 29(196-199): 367–376, 1879. Rouditchenko, A., Bhati, S., Araujo, E., Thomas, S., Kuehne, H., Feris, R., and Glass, J. Omni-r1: Do you really need audio to fine-tune your audio llm? arXiv preprint arXiv:2505.09439, 2025. Sakshi, S., Tyagi, U., Kumar, S., Seth, A., Selvakumar, R., Nieto, O., Duraiswami, R., Ghosh, S., and Manocha, D. Mmau: A massive multi-task audio understanding and reasoning benchmark. arXiv preprint arXiv:2410.19168, 2024. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al. Deepseekmath: Push- ing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm test- time compute optimally can be more effective than scal- ing model parameters. arXiv preprint arXiv:2408.03314, 2024. Tian, F., Zhang, X. T., Zhang, Y., Zhang, H., Li, Y., Liu, D., Deng, Y., Wu, D., Chen, J., Zhao, L., et al. Step-audio-r1 technical report. arXiv preprint arXiv:2511.15848, 2025. Waheed, A., Mitra, C., Wang, L. Z., Ramanan, D., and Raj, B. Less is more tokens: Efficient math reasoning via difficulty-aware chain-of-thought distillation. arXiv preprint arXiv:2509.05226, 2025. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. Wen, C., Guo, T., Zhao, S., Zou, W., and Li, X. Sari: Structured audio reasoning via curriculum-guided rein- forcement learning. arXiv preprint arXiv:2504.15900, 2025. Wu, S., Li, C., Wang, W., Zhang, H., Wang, H., Yu, M., and Yu, D. Audio-thinker: Guiding audio language model when and how to think via reinforcement learning. arXiv preprint arXiv:2508.08039, 2025. Xiaomi, L.-C.-T. Mimo-audio: Audio language models are few-shot learners, 2025. URLhttps://github. com/XiaomiMiMo/MiMo-Audio. Xie, Z., Lin, M., Liu, Z., Wu, P., Yan, S., and Miao, C. Audio-reasoner: Improving reasoning capability in large audio language models. arXiv preprint arXiv:2503.02318, 2025. Xu, J., Guo, Z., He, J., Hu, H., He, T., Bai, S., Chen, K., Wang, J., Fan, Y., Dang, K., et al. Qwen2. 5-omni techni- cal report. arXiv preprint arXiv:2503.20215, 2025. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., and Zhu, W. Avqa: A dataset for audio-visual question answering on videos. In Proceedings of the 30th ACM international conference on multimedia, p. 3480–3491, 2022. Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Ad- vances in neural information processing systems, 36: 46595–46623, 2023. 10 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning A. CAFE Framework Prompts In terms of prompt design, we have referred to efficient prompting methods (Chang et al., 2024). The following are the Gemini-3-pro caption and audio event extraction prompts used by the CAFE framework. A.1. Gemini-3-pro Caption prompt Please generate a detailed chronological description of the following audio clip. Listen carefully to the sequence of events and describe the audio flow from beginning to end. Your caption should: - Explicitly state the order of sounds using transitional phrases (e.g., The audio begins with..., Followed by..., Simultaneously..., As the sound fades...). - Capture the subtle details of each sound event, including its duration and intensity changes. - Distinguish between foreground events (main actions) and background noise (ambience). Output the result as a cohesive narrative text without line breaks or bullet points. A.2. Audio Event Extraction prompt Role You are an expert Audio Logic Consistency Evaluator. Your task is to evaluate a Model Reasoning Path against the Question, Correct Answer, and Ground Truth Caption. Step 1: Analyze Requirements 1. Determine ‘requiredevents‘: The essential sounds from the [Ground Truth Audio Caption] needed to answer the [Question] correctly. 2. Identify ‘all captionevents‘: All sounds actually present in the audio. Step 2: Categorize Model Events For every audio event mentioned in the [Model Reasoning Path], categorize it into ONE of the following 4 lists based on its Validity (Is it real?) and Usage (How did the model use it?): 1. matched events (Correct & Necessary): - The event is in ‘requiredevents‘. - The model used it effectively to derive the answer. 2. errormatched (Fabrication / Misidentification): - The event is NOT in ‘allcaptionevents‘ (Hallucination). - OR The model heard sound A (real) but identified it as sound B (fake/wrong), and B is not in the caption. - These are False Positives regarding perception. 3. error use (Distraction / Wrong Reasoning): - The event exists in ‘all captionevents‘ but is NOT in ‘requiredevents‘ (Irrelevant). - CRITICAL CONDITION: The model activley used this irrelevant sound to support a conclusion, or the model was confused by it. - Example: I hear a bird (irrelevant), so the answer must be ’Forest’. (When the answer is actually ’Park’ due to other sounds). 4. neutralevents (Harmless Mention / Valid Filtering): - The event exists in ’allcaptionevents’ but is NOT in ‘requiredevents‘. - The model mentioned it only to describe the scene or explicitly stated it was not relevant. - Example: I hear wind in the background, but the main sound is the car engine. (Here, ’wind’ is a neutral mention, not an error). - Action: Do NOT count these as errors. 5. missedevents (Omission): - Events in ‘required events‘ that are NOT found in ‘matchedevents‘. 11 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Input Data [Question]: QUESTION [Correct Answer]: CORRECTANSWER [Ground Truth Audio Caption]: GROUND TRUTHCAPTION [Model Reasoning Path]: MODELREASONING Output Format (JSON Only) all reasoningevents: [list of all events mentioned by model], matchedevents: [list], error matched: [list], erroruse: [list], neutralevents: [list], missedevents: [list] B. Inference prompt In Section 3, the open-source LALMs evaluated include MiMo-Audio (Xiaomi, 2025), Qwen2.5-Omni-7B (Xu et al., 2025), DeSTA2.5-Audio (Lu et al., 2025), Phi-4-Multimodal (Abouelenin et al., 2025), Qwen2-Audio-Instruct (Chu et al., 2024), as well as Omni-R1 (Rouditchenko et al., 2025) without Chain-of-Thought (CoT) RL training. Regarding LARMs, we evaluate Audio-Flamingo-3 (Goel et al., 2025), Audio Reasoner (Xie et al., 2025), Step-Audio-R1.1 and Step-Audio- R1 (Tian et al., 2025). Additionally, we incorporate the commercial models Gemini-2.5-Flash (Comanici et al., 2025) and GPT-4o-Audio (Hurst et al., 2024). Finally, we evaluate the model trained using our MPAR 2 method. B.1. LALMs and Commercial Models The following prompt is used to evaluate LALMs and commercial models. [Quesiton] Output the selected option first. Then write Reasoning: and explain your choice, mentioning the audio events that support it. [Choices] B.2. LARMs The prompt below is used for evaluating LARMs and MPAR 2 trained model. [Quesiton] Select one option from the provided choices. [Choices] B.3. Implicit and Explicit Unstructured Reasoning Prompts for RL Implicit Prompt: Output the final answer in <ANSWER ></ANSWER> Explicit Unstructured Reasoning Prompt:Output the thinking process in<THINK></THINK>and final answer in <ANSWER></ANSWER>. C. Benchmark and Framework Details We primarily evaluate model reasoning performance using accuracy on multiple-choice questions, and further conduct reasoning analysis under the CAFE framework. The main evaluations and metrics used are: 12 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Table 3. Performance Comparison on MMAU-v05.15.25 (Test-mini). Results for other methods are sourced from the MMAU Leaderboard: MMAU-v05.15.25. The best results highlighted in bold, and the second-best scores are underlined. ModelSoundMusicSpeechAvg Large Audio Language Models Omni-R181.7073.4076.0077.00 MiMo-Audio81.6874.2568.1774.70 Qwen2.5-Omni-7B78.1065.9070.6071.50 Kimi-Audio75.6866.7762.1668.20 Qwen2-Audio-Instruct67.2756.2955.2659.60 Large Audio Reasoning Models Audio-Thinker81.98 74.2576.8877.70 Step-Audio 284.0473.5675.1577.58 Audio Flamingo 379.5873.9566.3773.30 Audio-Reasoner67.8769.1666.0767.70 Commercial Models Gemini 2.5 Flash73.2765.5776.5871.80 GPT-4o Audio64.5656.2966.6762.50 MPAR 2 -7B79.2072.8075.1075.70 MMAU Benchmark(Sakshi et al., 2024) We evaluate the model on the MMAU test-mini split, which features demanding audio QA tasks requiring advanced reasoning. Model performance is reported as multiple-choice accuracy. The MMAU- v05.15.25 release revises approximately 25% of the questions and answers from the original MMAU benchmark to improve clarity, accuracy, and overall quality. In addition, around 5% of the audio files are refined to enhance acoustic consistency and signal fidelity. The official MMAU benchmark had not been updated during our initial experiments, particularly during the development of the MPAR 2 framework with different reward functions. As a result, all experiments were conducted using the origin version of MMAU. To ensure experimental consistency, all results reported in the main paper are based on this previous version. After the release of MMAU-v05.15.25, we further evaluated the trained models on the updated benchmark to assess their performance under the latest setting and to enable direct comparison with current state-of-the-art models. MMAR Benchmark(Ma et al., 2025b) The benchmark is designed to evaluate the reasoning capabilities of Audio Language Models in realistic acoustic environments. It consists of 1,000 audio–question–answer samples collected from open-domain videos and curated through expert annotation and quality control. Unlike prior benchmarks that focus on isolated audio categories, MMAR includes mixed and overlapping audio compositions. For fair comparison with the MMAU dataset, which reports results only on isolated categories, Tables 2 and 4 present MMAR results under isolated audio settings. Queries are further organized into four hierarchical reasoning levels, ranging from low-level acoustic analysis to high-level semantic and culturally grounded inference. CAFE Framework As detailed in Section 3, CAFE quantifies the perception and utilization of audio events during reasoning. Leveraging ground-truth audio and QA pairs, it evaluates performance via metrics covering perception accuracy, reasoning accuracy, etc. D. Supplementary Results D.1. MMAU (v05.15.25) Results As shown in Table 3, when evaluated on MMAU-v05.15.25, the MPAR 2 model achieves a substantial improvement over the baseline Qwen2.5-Omni-7B, which is consistent with the performance trends observed on the original MMAU benchmark. In addition, compared with previous state-of-the-art models, MPAR 2 delivers highly competitive results. 13 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Table 4. Accuracy (%) on MMAU (original) and MMAR benchmark. Model MMAU (Test-mini)MMAR Sound↑Music↑Speech↑Avg↑Sound↑Music↑Speech↑Avg↑ Qwen2.5-Omni-3B70.2760.4859.1663.3053.9446.1253.7453.80 Qwen2.5-Omni-7B69.4066.8061.6065.9053.9442.2358.8455.20 MPAR 2 -3B76.4265.1169.0470.1753.6841.5658.4555.62 MPAR 2 -7B77.9871.3074.4774.5962.4044.3064.2660.32 2535455565758595105115 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (a) Qwen2.5-Omni-3B (Base) 224272320368416464512560608656 0 0.2 0.4 0.6 0.8 1 Perception Accuracy Reasoning Accuracy (b) MPAR 2 -3B MMAU (Test-mini) MMAR 400 440 480 520 560 600 640 398 445 486 382 443 545 623 552 633 EasyInformation ExtractionAcoustic Mixtures MediumReasoningPerception HardSingle-SourceHigh-level Cognitive (c) Reasoning Token Length of MPAR 2 -3B on Benchmarks Figure 8. (a)-(b) is the accuracy of MPAR 2 -3B vs. Base across token lengths. (c) is the Reasoning token length of MPAR 2 -3B across different task categories on the MMAU and MMAR benchmark. D.2. MPAR 2 -3B Results Performance Enhancement Trajectory Remains Consistent at the 3B Scale As detailed in Table 4, MPAR 2 -3B significantly outperforms its baseline, elevating accuracy on MMAU from 63.30% to 70.17% and on MMAR from 53.80% to 55.62%. Mirroring the 7B results, widespread gains are evident across Sound, Music, and Speech domains. These findings confirm that the efficacy of MPAR 2 is agnostic to model size, effectively empowering smaller models with robust audio perception and reasoning capabilities. Similar Mitigation of Perception Decay is also Observed at the 3B ScaleAs illustrated in Figure 8 (a)-(b), the MPAR 2 - 3B model successfully mitigates the trend of perception decay, maintaining robust accuracy even as reasoning chains expand. This consistency confirms that our framework effectively stabilizes perceptual fidelity irrespective of model size. Adaptive Reasoning Length Persists at the 3B Scale As illustrated in Figure 8 (c), MPAR 2 -3B effectively aligns its reasoning length with task complexity, generating notably longer chains for the reasoning-intensive MMAR benchmark compared to MMAU. Specifically, the model produces concise paths for simple tasks while reserving expanded trajectories for complex acoustic scenarios. This demonstrates a learned ability to dynamically scale reasoning depth, thereby optimizing the balance between computational efficiency and quality. Table 5. Results of various audio models on the MMAR benchmark within the CAFE framework. † denotesMMAR acc evaluated by GPT-5 judge. All results are averaged over three inference runs. ModelAcc per ↑ Err per ↓ Err use ↓ Err omit ↓ MMAR acc ↑ Qwen2.5-Omni-7B31.7429.6128.6957.3755.20 Qwen2.5-Omni-3B28.5439.3530.952.1453.80 MPAR 2 -7B63.5123.147.7430.5960.32 MPAR 2 -3B61.0423.3918.7432.4655.62 Robust Perceptual Performance in 3B Scale Table 5 reveals that this improvement is scale-invariant: MPAR 2 -3B achieves a perception accuracy (Acc per ) of 61.04%, doubling the performance of its baseline (28.54%) and delivering results comparable to the 7B variant. This highlights the method’s potential to significantly enhance perception even in smaller-scale models.” 14 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning E. Training Details The overview of the data construction is in Figure 9. AVQA DatasetGemini 2.5-pro AVQA Dataset with Detailed Captions Qwen3-32BAudio QA Qwen3-32B QA Quality filtering Base Model simples 16 time Filter out All Correct and All Incorrect U n i f o r m l y s a m p l e 5 0 0 0 R e m a i n i n g QA Data for RL QA Data for SFT Qwen3-32BAudio QA with CoT Qwen3-32B CoT Quality filtering Data for SFT Stage1: Q-A Synthesis Stage 2: CoT Construction Figure 9. The overview of the data construction pipline. E.1. Gemini-2.5-Pro Caption Prompt You are an expert audio event analyst. Your task is to carefully analyze the entire audio and extract all meaningful events as can as possible, including source entities, time intervals and captions. For each extracted event, you must: 1. Identify the possible sound sources within the event and make a brief descriptions for it. 2. Provide the accurate time interval [start time,end time] for each event when it occurs within the audio. 3. Provide a detailed, precise, and contextually rich caption of what happens in that event interval. Do not infer or imagine information that cannot be clearly heard. The max number of extracted events is 5. Please separate different parts using <event1> < description></description><caption>...</caption></event1><event2><description>...</descrip- tion><caption>... </caption></event2>...... Here is a sample.Please strictly follow the format in the sample.<event1><description>dogs barking </description><caption>the audio captures a brief, intense episode of multiple dogs barking aggressively in an outdoor setting.</caption></event1><event2><description>...</description><caption>...</cap- tion></event2><event3><description>... </description><caption>... </caption></event3>...... Please extract events from the follow audio: 15 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning E.2. QA Generate Prompt E.2.1. EVENT COUNTING QA GENERATE PROMPT We are constructing training data to enhance audio perception and reasoning in audio large language models (LLMs). Your job is to create a single, high-quality multiple-choice question that tests whether a model can perform numerical reasoning and quantitative analysis by listening to a complete audio clip. Input Audio Events (Chronological Order): eventsdescription Task Description: You are provided with the ground truth sequence of sound events above. Your goal is to generate ONE multiple-choice question (MCQ) focusing on Event Counting & Numerical Reasoning. Step 1: Suitability Check (CRITICAL) First, analyze the event list provided above. Ask yourself: 1. Are there distinct, countable discrete events (e.g., ’bark’, ’gunshot’, ’footstep’) rather than just continuous ambience (e.g., ’wind’, ’static’, ’silence’)? 2. Is there enough variety or repetition to form a valid numerical question (e.g., counting total occurrences, comparing counts of two different sources)? If the content consists mainly of continuous noise, ambiguous sounds, or a single non-repeatable event that makes counting trivial or impossible, you must output exactly: Not suitable for this hallucination type. Step 2: Question Generation (If Suitable) If the audio events support numerical reasoning, generate ONE MCQ. The question must target the auditory experience. Do NOT ask about the text descriptions directly. The question must be answerable by listening to the audio and counting/analyzing the sounds. Focus Areas: - Total Count: ’How many times is the [specific sound] heard?’ - Source Comparison: ’Did the dog bark more times than the cat meowed?’ - Sequence Logic: ’After the first door slam, how many footsteps follow?’ (Note: Since you do not have exact timestamps, focus on the count and order of events, not the speed or rate per minute.) STRICT CONSTRAINT: The question must NOT contain phrases like ’According to the list’, ’In the description’, or ’text’. It must sound like a natural question asked to someone who has just listened to the recording. Output Format (if suitable): Question: <question about the sound itself> A. <option A> B. <option B> C. <option C> D. <option D> Correct answer: <the correct option letter> Case Study of Event Counting Question: How many distinct cawing events are clearly heard before the audio becomes chaotic with overlapping caws? Choice: A. One B. Two C. Three D. Four Correct answer: C 16 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning E.2.2. PITCH CONTOUR We are constructing training data to enhance audio perception and reasoning in audio large language models (LLMs). Your job is to create a single, high-quality multiple-choice question that tests whether a model can perform deep inference based on the dynamic pitch contour (frequency modulation) by listening to a complete audio clip. Input Audio Events (Chronological Order): eventsdescription Task Description: You are provided with the ground truth sequence of sound events above. Your goal is to generate ONE multiple-choice question (MCQ) focusing on Pitch Contour & Semantic Inference. Step 1: Suitability Check (CRITICAL) Analyze the event list provided above. To support a question about pitch contour, the audio must contain sources with detectable tonal properties or modulation. Ask yourself: 1. Does the audio contain speech (where intonation conveys meaning like sarcasm or questions)? 2. Does it contain tonal machinery or vehicles (where speed/movement affects pitch, e.g., Doppler effect, accelera- tion)? 3. Does it contain musical or biological sounds (e.g., bird calls, singing) where frequency changes involve meaning? If the audio consists ONLY of unpitched sounds (e.g., ’rain’, ’static noise’, ’footsteps’, ’door slam’, ’wind’) where pitch analysis is irrelevant or impossible, you must output exactly: Not suitable for this hallucination type. Step 2: Question Generation (If Suitable) If the audio contains tonal or pitch-varying events, generate ONE MCQ. The question must require the model to hear the curve of the sound (rising, falling, wavering) and deduce the underlying cause, intent, or movement. Focus Areas (Audio-Centric): - Intonation & Intent: ’The speaker’s pitch rises sharply at the end. What does this suggest about their certainty?’ (Reasoning: Statement vs. Question) - Doppler Effect (Physics): ’As the [vehicle] sound gets louder and then fades, the pitch drops significantly. What does this confirm about its movement?’ (Reasoning: Passing the listener) - Mechanical State: ’The pitch of the engine continuously increases without dropping. What does this imply about the machine’s operation?’ (Reasoning: Acceleration/Revving) - Emotional State: ’The tremor or wavering pitch in the voice suggests what emotion?’ (Reasoning: Fear/Excitement STRICT CONSTRAINT: The question must NOT contain phrases like ’According to the list’, ’In the description’, or ’text’. It must sound like a natural question asked to someone who has just closed their eyes and listened to the recording. Output Format (if suitable): Question: <question about the pitch/tone/intonation> A. <option A> B. <option B> C. <option C> D. <option D> Correct answer: <the correct option letter> Case Study of Pitch Contour Question: A siren is heard starting low, rising quickly to a high pitch, and then holding that pitch for a while before stopping. What is this pattern most likely indicating? Choice: A. A fire truck arriving at the scene B. An emergency evacuation alert C. A vehicle accelerating away D. A 17 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning malfunctioning car alarm Correct answer: B E.2.3. RHYTHMIC STRUCTURE We are constructing training data to enhance audio perception and reasoning in audio large language models (LLMs). Your job is to create a single, high-quality multiple-choice question that tests whether a model can perform deep inference based on the rhythmic structure and temporal regularity by listening to a complete audio clip. Input Audio Events (Chronological Order): eventsdescription Task Description: You are provided with the ground truth sequence of sound events above. Your goal is to generate ONE multiple-choice question (MCQ) focusing on Rhythmic Structure & Behavioral Inference. Step 1: Suitability Check (CRITICAL) Analyze the event list provided above. To support a question about rhythm, the audio must contain sounds that repeat or form a pattern. Ask yourself: 1. Does the audio contain repetitive impulsive sounds (e.g., ’footsteps’, ’typing’, ’heartbeat’, ’clapping’, ’knocking’)? 2. Does it contain rhythmic machinery or engines (e.g., ’train wheels’, ’clock ticking’, ’idling engine’)? 3. Does it contain music or percussion? If the audio consists ONLY of continuous amorphous noise (e.g., ’wind’, ’water flow’, ’static’) OR single isolated events (e.g., ’one gunshot’, ’a single drop’, ’one scream’) where no rhythm exists, you must output exactly: Not suitable for this hallucination type. Step 2: Question Generation (If Suitable) If the audio contains rhythmic patterns, generate ONE MCQ. The question must require the model to hear the pattern of the sound (regularity, tempo, acceleration, chaos) and deduce the underlying behavior or state. Focus Areas (Audio-Centric): - Activity State (Tempo): ’The footsteps transition from a slow, steady walking pace to a rapid, frantic rhythm. What does this suggest about the subject?’ (Reasoning: Leisure vs. Fleeing) - Stability (Regularity): ’The engine sound changes from a steady hum to an irregular, sputtering pattern. What does this indicate?’ (Reasoning: Mechanical failure) - Coordination (Synchronization): ’The clapping sounds represent a large group hitting the beat in perfect unison. What context does this imply?’ (Reasoning: Organized audience/Performance vs. Chaotic crowd) - Environmental Context: ’The rhythm of the typing is sporadic and hesitant, with long pauses. What does this suggest about the typist?’ (Reasoning: Thinking/Uncertainty vs. Professional transcription) STRICT CONSTRAINT: The question must NOT contain phrases like ’According to the list’, ’In the description’, or ’text’. It must sound like a natural question asked to someone who has just closed their eyes and listened to the recording. Output Format (if suitable): Question: <question about the pitch/tone/intonation> A. <option A> B. <option B> C. <option C> D. <option D> Correct answer: <the correct option letter> Case Study of Rhythmic Structure 18 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Question: The pig’s vocalizations occur in a tightly spaced, repetitive sequence at the start of the audio. What does this rhythmic pattern most likely suggest about the pig’s emotional or behavioral state? A. It is calmly and slowly exploring its environment. Choice: A. It is calmly and slowly exploring its environment. B. It is experiencing distress or agitation. C. It is communicating with a distant group of pigs. D. It is playing or engaging in a harmless interaction with another animal. Correct answer: B E.2.4. ETEMPORAL LOGIC We are constructing training data to enhance audio perception and reasoning in audio large language models (LLMs). Your job is to create a single, high-quality multiple-choice question that tests whether a model can perform Temporal Logic and Causal Reasoning by listening to a complete audio clip. Input Audio Events (Chronological Order): eventsdescription Task Description: You are provided with the ground truth sequence of sound events above. Your goal is to generate ONE multiple-choice question (MCQ) focusing on Sequence, Causality, or Temporal Relationship. Step 1: Suitability Check (CRITICAL) Analyze the event list provided above. To support temporal or causal reasoning, the audio must contain a sequence of multiple distinct events. Ask yourself: 1. Are there at least two distinct events happening in succession (e.g., ’thunder’ then ’rain’, ’footsteps’ then ’door open’)? 2. Is there a logical link (cause-and-effect) or a clear chronological order to test? If the audio consists of a SINGLE event (e.g., just ’dog barking’) OR only continuous background noise (e.g., ’static’, ’city traffic’) where no sequence exists, you must output exactly: Not suitable for this hallucination type. Step 2: Question Generation (If Suitable) If the audio contains a valid sequence of events, generate ONE MCQ. The question must require the model to track the order of sounds or deduce the cause of a sound based on what happened before/after. (Note: Do NOT ask about specific duration in seconds, as you do not have exact timestamps. Focus on relative order and logic.) Focus Areas (Audio-Centric): - Chronological Order: ’Which sound occurred immediately after the glass shattered?’ (Reasoning: Tracking sequence). - Causality/Reaction: ’The sudden braking sound was immediately followed by a crash and shouting. What does this sequence imply?’ (Reasoning: Cause and Effect). - Interruption: ’The music was abruptly cut off by which specific sound?’ (Reasoning: Identifying the disruptor). - Reverse Inference: ’The sound of the audience clapping suggests that what event likely just finished?’ (Reasoning: Inferring the preceding context). STRICT CONSTRAINT: The question must NOT contain phrases like ’According to the list’, ’In the description’, or ’text’. It must sound like a natural question asked to someone who has just closed their eyes and listened to the recording. Output Format (if suitable): Question: <question about the pitch/tone/intonation> A. <option A> B. <option B> 19 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning C. <option C> D. <option D> Correct answer: <the correct option letter> Case Study of Etemporal logic Question: A single dog begins barking, and soon after, multiple dogs join in. Which of the following sounds is most likely to have triggered the initial barking? Choice: A. The low-frequency electronic tone B. The person sighing or groaning C. The sound of footsteps D. The rustling of clothing Correct answer: A E.2.5. TIMBRE CHARACTERISTICS We are constructing training data to enhance audio perception and reasoning in audio large language models (LLMs). Your job is to create a single, high-quality multiple-choice question that tests whether a model can perform Temporal Logic and Causal Reasoning by listening to a complete audio clip. Input Audio Events (Chronological Order): events description Task Description: You are provided with the ground truth sequence of sound events above. Your goal is to generate ONE multiple-choice question (MCQ) focusing on Sequence, Causality, or Temporal Relationship. Step 1: Suitability Check (CRITICAL) Analyze the event list provided above. To support temporal or causal reasoning, the audio must contain a sequence of multiple distinct events. Ask yourself: 1. Are there at least two distinct events happening in succession (e.g., ’thunder’ then ’rain’, ’footsteps’ then ’door open’)? 2. Is there a logical link (cause-and-effect) or a clear chronological order to test? If the audio consists of a SINGLE event (e.g., just ’dog barking’) OR only continuous background noise (e.g., ’static’, ’city traffic’) where no sequence exists, you must output exactly: Not suitable for this hallucination type. Step 2: Question Generation (If Suitable) If the audio contains a valid sequence of events, generate ONE MCQ. The question must require the model to track the order of sounds or deduce the cause of a sound based on what happened before/after. (Note: Do NOT ask about specific duration in seconds, as you do not have exact timestamps. Focus on relative order and logic.) Focus Areas (Audio-Centric): - Chronological Order: ’Which sound occurred immediately after the glass shattered?’ (Reasoning: Tracking sequence). - Causality/Reaction: ’The sudden braking sound was immediately followed by a crash and shouting. What does this sequence imply?’ (Reasoning: Cause and Effect). - Interruption: ’The music was abruptly cut off by which specific sound?’ (Reasoning: Identifying the disruptor). - Reverse Inference: ’The sound of the audience clapping suggests that what event likely just finished?’ (Reasoning: Inferring the preceding context). STRICT CONSTRAINT: The question must NOT contain phrases like ’According to the list’, ’In the description’, or ’text’. It must sound like a natural question asked to someone who has just closed their eyes and listened to the recording. 20 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Output Format (if suitable): Question: <question about the pitch/tone/intonation> A. <option A> B. <option B> C. <option C> D. <option D> Correct answer: <the correct option letter> Case Study of Timbre Characteristics Question: Based on the sound characteristics of the buzzing in the background, what is the most likely physical source producing this continuous, low-frequency noise? Choice: A. A swarm of bees inside a wooden beehive B. A malfunctioning electrical device emitting a hum C. A distant, low-flying airplane D. A person humming a tune consistently and softly Correct answer: A E.3. QA Fliter Prompt Role You are an expert Data Quality Evaluator for Audio-Text Reasoning Datasets. Your goal is to strictly filter out low-quality QA pairs and keep only those that require genuine logical reasoning. Objective Evaluate the following Audio QA pair. You must determine if the Question requires logical deduction, causal inference, or temporal analysis of the events described in the Caption. Input Data - Audio Caption: caption - Question and Answer: question Evaluation Criteria (The ReasoningLitmus Test) 1. POSITIVE INDICATORS (High Score / KEEP) The question requires the model to: Infer Cause/Effect: Why did the sound stop?(Requires understanding the preceding event). Analyze Sequence (Temporal): What happened immediately after the explosion? Deduce State/Intent: Based on the footsteps and breathing, is the person running or walking? Synthesize Multiple Clues: Combining background noise + specific actions to determine the location. 2. NEGATIVE INDICATORS (Low Score / DISCARD) Simple Pattern Matching: The answer is just a word lifted directly from the caption (e.g., Caption says a red car, Question asks what color is the car?). Common Sense / General Knowledge: Can be answered without the audio caption (e.g., Do birds fly?). Summarization: Describe the audio(This is generation, not reasoning). Unsolvable/Hallucination: The answer assumes facts not present in the caption. Task 1. Analyze the relationship between the Caption, Question, and Answer. 2. Assign a Score (1-5) based on reasoning depth. - 5: Complex reasoning (multi-hop, causal, or temporal). - 4: Clear deduction required. - 3: Simple inference. - 2: Direct text retrieval / keyword matching. - 1: Common sense or irrelevant. 21 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 3. Make a final Decision: KEEP(Score >= 4) or DISCARD(Score <4). Output Format Return ONLY a valid JSON object: analysis: Brief explanation of the reasoning logic required., score: <int>, decision: KEEPor DISCARD E.4. MPAR 2 CoT Generate Prompt Role: You are an advanced AI assistant specializing in Audio Reasoning and Chain-of-Thought (CoT) synthesis. You also act as a strict auditor to ensure data quality. Task: Your task is to synthesize the provided input data into a structured CoT format using specific XML tags, and then perform a critical validation (Review) of the reasoning process. You will be given: 1.Question: The main query about the audio. 2.Final Answer: The ground truth answer 3.Caption: A time-interval detailed description of the audio events (Ground Truth) 4.Model Outputs: A list of step-by-step sub-questions and answers generated logically in previous steps. Output Format Requirements: You must output the content strictly inside the following XML structure: <thinking> <perception> 1. [start time, endtime]: Description of event A. 2. [start time, endtime]: Description of event B. ... (List ALL events from the Caption, chronologically) </perception> <reasoning> 1. Sub-question: [First step from Model Outputs] Answer: [Answer to first step] ... (Include all steps from the Model Outputs) </reasoning> <review> 1. Evidence Check: [Simulate a Re-listeningprocess. Verify if the audio events cited in the ’Reasoning’ are factually supported by the events listed in ’Perception’. Check for hallucinations, misinterpretations, or missing details. Note: Treat the ’Perception’ content as the audio itself; do not refer to ’captions’, ’text’, or ’provided descriptions’.] 2. Logic Check: [Evaluate the logical validity. Does the conclusion naturally follow from the evidence? Is the overall chain coherent?] </review> </thinking> <answer> [The Final Answer provided in the input] </answer> Directives: 1.Perception Section (Full Extraction): Translate the ENTIRE provided ‘Caption‘ into a structured, chronological list of time intervals and event descriptions. Do not filter, summarize, or omit any events, even if they seem irrelevant to the specific Question. Format: ‘[t1, t2]: content‘. 22 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 2.Reasoning Section: Directly utilize the provided ‘Model Outputs‘. Organize them into a numbered list of Sub- questionand Answerpairs. Ensure the reasoning logic flows smoothly. 3.Review Section (Critical Audit): Evidence Check: Perform a Re-Perception check. Context: Imagine you are re-checking the audio stream directly. Validity: Did the reasoning cite sounds that actually exist in the ‘Perception‘ list? Accuracy: Did the reasoning interpret the sound properties correctly? Constraint: Strictly avoid phrases like according to the captionor the text says. Instead, use phrases like The audio contains..., I hear..., or The event at [timestamp] shows.... Logic Check: Verify the soundness of the deductive process. Ensure there are no logical leaps or contradictions. 4.Answer Section: State the final answer clearly. Input Data: Question: ORIGINAL QUESTION Final Answer: FINALANSWER Caption: caption wtime Model Outputs (Reasoning Chain): subquestionlistgenerated E.5. CoT Data Fliter Prompt Role You are an expert evaluator for Audio-Language Models. Your task is to audit a Chain-of-Thought(CoT) process generated by an AI model. You will assess how well the model hearsthe audio (Perception), thinksabout it (Reasoning), and auditsits own conclusion (Review). Input Data 1. Original Question: The user’s query. 2. Final Answer: The ground truth answer. 3. Caption: Ground truth events with precise time ranges. 4. CoT: The step-by-step reasoning process containing <perception>, <reasoning>, and <review>tags. Evaluation Dimensions Dimension 1: Perception Evaluation (The Ear) Analyze specific claims about audio events/timestamps in the <perception>and <reasoning>sections. Evaluate based on: 1.Accuracy: Does the event described match the Caption? 2.Hallucination/Omission: Are there invented sounds or missed critical sounds? Dimension 2: Reasoning Evaluation (The Brain) Analyze the logical flow in the <reasoning>section. Evaluate each step on: 1.Utility: Is this step necessary for solving the Original Question? 2.Factuality: Is the statement factually true based on the audio content? 3.Logical Validity: Does the conclusion naturally follow from the cited evidence? Dimension 3: Review Evaluation (The Auditor) Analyze the <review>section, specifically the Evidence Checkand Logic Check. Evaluate based on: 1.Evidence Re-verification: Did the model correctly re-examine and re-perceive the audio events cited in the reasoning? Did it accurately confirm whether the events exist in the perception data? Critical: Did it successfully identify valid evidence versus hallucinated evidence? 2.Rationality Check: Did the model correctly assess the logical coherence of the entire chain? Did it ensure that the Final Answer is the only logical conclusion derived from the evidence? 23 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning Scoring & Output Format Output only two numbers (0-10) strictly in accordance with the following format: ReasoningScore/ReviewScore Scoring Criteria: ReasoningScore (num1): Rate the quality of Dimension 1 (Perception) and Dimension 2 (Reasoning). 10 = Perfect audio detection and flawless logic. ReviewScore (num2): Rate the quality of Dimension 3 (Review). 10 = The model performed a rigorous, accurate self-audit that correctly validated the evidence and logic. 0 = The review was superficial, inaccurate, or failed to catch obvious errors. Input Data: Question: ORIGINAL QUESTION Final Answer: FINALANSWER Caption: caption wtime Model Outputs (Reasoning Chain): sub questionlistgenerated F. Reward prompt F.1. Perception Score Prompt You are an expert audio perception evaluator. I will give you a record containing: 1. A Detailed Audio Caption (Ground Truth): A comprehensive, factual text description of the audio events. 2. A Question and its Correct Answer: To determine which audio events are ’critical’ for the task. 3. A Model Perception Output: The content within the <perception>tags generated by the model, describing events with timestamps. Your task is to evaluate the fidelity, precision, and completeness of the Model Perception against the Ground Truth, and output a single numeric score from 0, 0.1, 0.2, ..., 1.0. You must output only the score with no explanation or extra text. Evaluate based on the following CRITICAL principles: 1. Audio Hallucination (Strict) — The model must NOT report events that do not exist in the ’Detailed Audio Caption’. Reporting a sound that is completely absent (e.g., hearing a siren when the description only mentions birds) is a fatal failure. 2. Content Accuracy & Sequential Logic — While evaluating the model’s generated timestamps, focus on: - Event Identity: Does the model correctly identify the sound sources described in the Ground Truth? (e.g., distinguish- ing ’footsteps’ from ’knocking’). - Chronological Flow: Does the sequence of events in the model’s output match the narrative order of the Ground Truth? (e.g., if the description says ’a door opens then slams’, the model must not place the slam before the opening). 3. Critical Event Coverage (Relevance) — The model must capture all ’Key Events’ necessary to answer the provided ’Question’. Compare with the Question/Answer pairs: if the answer depends on a specific sound cue, omitting this specific event in the <perception>phase is a critical failure. 4. Consistency & Identity — The model should describe the same audio source consistently across different timestamps (unless the sound evolves). Avoid contradictory descriptions for the same ongoing event. 5. Redundancy & Conciseness — The perception output should be dense and informative. Penalize distinct ’loops’ (repeating the exact same phrase for adjacent timestamps) or extreme verbosity that adds no new details. Scoring guideline 1.0 = Flawless. Perfectly matches the Ground Truth description. Events are correctly identified and listed in the correct logical order. Concise. 24 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 0.8-0.9 = Excellent. Accurate detection of all key events described. The sequence is logical. Maybe minor verbosity. 0.5-0.7 = Mediocre. The KEY event was detected, but the description is vague, or information irrelevant to the question has been omitted. 0.2-0.4 = Poor. Misses a KEY event needed for the Answer, or misidentifies a sound source. Sequence is disorderly compared to the description. 0.0-0.1 = Severe. HALLUCINATION (inventing sounds not in the description), or total failure to identify the main audio event. Penalty guideline (Apply these cumulatively to reduce the score): [CRITICAL PENALTY] (Set Score to 0.0∼ 0.2) : - Hallucinating an event not present in the Ground Truth description. - Misidentifying the main sound source (e.g., ’gunshot’ vs ’drum’). - Missing the specific audio cue required to answer the Question. [MODERATE PENALTY] (-0.3∼ -0.5) : - Sequential Logic Error (Events are listed in an order contradicting the description). - Significant omission of details mentioned in the description. [MINOR PENALTY] (-0.1∼ -0.2) : - Excessive wordiness or repetitive phrasing without new information. - Vague descriptions (e.g., ’noise’ instead of ’dog barking’) if the Ground Truth is specific. Operational rule: Always output only one score (0-1 in 0.1 increments). Now evaluate the following record and output only the score. The Detailed Audio Caption (Ground Truth) is:captiontext The Question is:questiontext The Correct Answer is:answertext The Model perception to evaluate is:cottext F.2. Step-Level Reasoning Score Prompt You are an expert logic and reasoning evaluator for Audio-LLMs. I will give you a record containing: 1. A Detailed Audio Caption (Model Perception): A comprehensive text description of the audio events. 2. A User Question and Context: The goal of the reasoning. 3. A Reasoning History: The steps taken so far. 4. The CURRENT STEP: The specific sub-question or reasoning step to evaluate now. Your task is to evaluate the validity, necessity, and audio-grounding of the CURRENT STEP only, and output a single numeric score from 0, 0.1, 0.2, ..., 1.0. You must output only the score with no explanation or extra text Evaluate based on the following Micro-Level dimensions: — CRITERIA: Local Quality Check — 1. Usefulness: Is this specific reasoning step useful for answering the main question? 2. Evidence-Based Conclusion: Is the content of this step supported by the provided ’Audio Caption’? - Every claim must align with specific events, sound sources, or acoustic details described in the caption. 3. Criticality & Efficiency: Is this step a logical next move based on the [Reasoning History]? - Penalize ’tangential reasoning’ (analyzing irrelevant noise) or redundant repetition of previous steps. Scoring guideline: 1.0 = Perfect. The step is firmly grounded in the caption, necessary, and logically follows the history. 0.8-0.9 = Strong. Good step, but maybe slightly inefficient or the evidence citation is slightly vague. 0.5-0.7 = Mediocre. Relevant, but weak grounding (making assumptions not explicitly in the caption). Logic holds but is messy. 0.2-0.4 = Weak. The step makes a claim not supported by the caption, or merely repeats previous steps without adding value. 25 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 0.0-0.1 = Failed. Completely incoherent, visual hallucination, or factual contradiction with the caption (e.g., claiming a sound exists when caption implies silence). Penalty guideline (Apply these cumulatively to reduce score): [CRITICAL PENALTY] (Set Score to 0.0∼ 0.1): - ’Factual Contradiction’: The step claims a specific sound or event occurs which is explicitly absent or contradicted by the Caption. Operational rule: Always output only one score (0-1). If [Current Step] is empty, return 0.0. The Detailed Audio Caption is:caption text The User Question is:question text The Reasoning History is:historytext The CURRENT STEP to evaluate is:currentsteptext F.3. Holistic Level Reasoning Score Prompt You are an expert logic and reasoning evaluator for Audio-LLMs. I will give you a record containing: 1. A Detailed Audio Caption (Model Perception): A comprehensive text description of the audio events. 2. A Question and its Correct Answer. 3. The COMPLETE Model Reasoning: The entire chain of thought generated by the model. Your task is to evaluate the logical architecture, coherence, efficiency, and final derivability of the entire process, and output a single numeric score from 0, 0.1, 0.2, ..., 1.0. You must output only the score with no explanation or extra text. Evaluate based on the following Macro-Level dimensions: — CRITERIA: Holistic Logical Architecture — 1. Goal-Orientation: Is the reasoning path linear and directed towards the [Correct Answer]? - Penalize circular logic. 2. Causal Dependency: Does Step B legitimately follow Step A? - Penalize ’Logic Jumps’ where a conclusion appears out of nowhere without a preceding premise defined in the audio caption. 3. Error Propagation Check: Does an early error render the rest of the chain invalid? - If Step 1 is wrong (e.g., misidentifying a gender or sound source compared to the caption), and subsequent steps rely on it, the whole chain collapses. 4. Final Derivability: Does the reasoning naturally flow to the [Correct Answer]? - The conclusion must be the inevitable result of the reasoning steps, not a sudden guess. 5. Efficiency & Conciseness: Is the length of the reasoning proportional to the complexity of the question? - Penalize ’Over-Analysis’: If the question is simple (e.g., ’Is there a dog?’), the reasoning should be short. Writing a 500-word essay for a simple question is a failure. - Penalize Repetition: Check if the model repeats the same analysis in different words just to make the chain longer. Scoring guideline: 1.0 = Perfect. Every step is necessary, concise, and the logic flows flawlessly from the caption evidence to the correct conclusion. 0.8-0.9 = Strong. Good logic, but maybe slightly verbose or includes one unnecessary step, yet the path is valid. 0.5-0.7 = Mediocre. Logic holds but is bloated or unfocused. Contains repetitive analysis or over-explains simple facts found in the caption. 0.2-0.4 = Weak. Major Logic Jumps, or the reasoning is excessively long and tedious without adding value (Filibuster- ing). 26 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning 0.0-0.1 = Failed. The reasoning contradicts the final answer, relies on ’Fatal Error Propagation’, or is complete nonsense. Penalty guideline (Apply these cumulatively): [CRITICAL PENALTY] (Set Score to 0.0∼ 0.2): - ’Fatal Error Propagation’: Early false premise (contradicting the Audio Caption) corrupts the entire remaining chain. - ’Contradiction’: The reasoning concludes something different from the actual Correct Answer provided. [MODERATE PENALTY] (-0.2∼ -0.4) : - ’Bloated Reasoning’: The reasoning is too long for the problem’s difficulty (e.g., 10 steps for a Yes/No question). - ’Irrelevance’: Wasting steps on analyzing audio events that are present in the caption but do not help answer the specific Question. Operational rule: Always output only one score (0-1). The Detailed Audio Caption is:caption text The Question is:question text The Correct Answer is:answertext The COMPLETE Model Reasoning is:fullreasoning F.4. Review Score Prompt You are a critical meta-evaluator for the ”Self-Correction” (Review) phase of an Audio-LLM. Your task is to judge whether the [Review Content] effectively audits, verifies, and corrects the [Model Reasoning]. Input Data: 1. [Detailed Audio Caption] (Model Perception): captiontext 2. [Ground Truth Annotations] (Fact Reference): ground truthtext 3. [User Question]: question text 4. [Correct Answer] (Ground Truth): answer text 5. [Model Reasoning] (Target to Audit): reasoningtext 6. [Review Content] (The Audit Output): reviewtext Task: Output a single score 0.0, 0.1, ..., 1.0 for the [Review Content]. Evaluation Criteria (Review Quality): 1. Evidence Verification (Content Alignment): - Does the Review explicitly verify that every event cited in the [Model Reasoning] actually exists in the [Detailed Audio Caption]? - Did it catch ”Hallucinations” where the Reasoning cites a sound (e.g., ”dog barking”) that is completely absent from the Caption? - CRITERIA: The Review must confirm that the evidence used in reasoning is physically present in the text description. 2. Temporal & Causal Logic Audit: - Does the Review check the narrative sequence? (e.g., if Reasoning says ”A causes B”, did the Review check if the Caption describes A happening before or leading into B?) - Did it catch chronological errors where the Reasoning flips the order of events described in the text? 3. Logical Integrity Check: - Does the Review ensure the conclusion is strictly derived from the perceived evidence? - Did it flag any ”Logic Jumps” (conclusions without premises) in the Reasoning? 4. Genuine Error Correction & Answer Verification (Anti-Rubber-Stamping): - ANSWER CHECK (CRITICAL): Compare the conclusion/answer in [Model Reasoning] with the [Correct Answer]. - If Reasoning leads to a WRONG answer, the Review MUST detect and flag this failure. - If Reasoning leads to a CORRECT answer, the Review must validate the logic flow. - HALLUCINATION CHECK: If [Model Reasoning] contains factual errors (contradictions with Caption), the Review MUST point them out. 27 When Scaling Fails: Mitigating Audio Perception Decay of LALMs via Multi-Step Perception-Aware Reasoning - PENALTY RULES: - Score 0.0 IMMEDIATELY if the [Model Reasoning] concludes with a wrong answer (mismatch with [Correct Answer]) but the Review states ”The reasoning is correct” or ”The answer is valid”. - Score 0.0 IMMEDIATELY if the Review approves reasoning that contradicts the Audio Caption (e.g., approving a sound that isn’t in the description). Scoring Guide: - 1.0: Perfect Audit. The review rigorously checked evidence presence, sequential logic, AND accurately validated the final answer against Ground Truth. - 0.8-0.9: Strong. Good check, caught major issues, but maybe missed a minor detail in the description. - 0.5-0.7: Generic Validation. Says ”Logic is good” without citing specific text evidence. (Rubber-stamping). - 0.2-0.4: Weak. Fails to catch obvious hallucinations or logic jumps. - 0.0-0.1: [FATAL] The Review acts as a ”Yes-Man” (Rubber-Stamp) for an INCORRECT Answer. It approves a reasoning path that leads to a result different from the [Correct Answer]. Output Rule: Return ONLY the numeric score. 28

Instant research discovery