Source-linked AI summary
Audio-Cogito: Towards Deep Audio Reasoning in Large Audio Language Models
Longhao Li, Hongjie Chen, Zehan Li, Qihan Hu, Jian Kang, Jie Li, Lei Xie, Yongxiang Li
TL;DR
Audio reasoning models remain limited and inconsistent, while available datasets often provide only brief labels or shallow reasoning examples. Audio-Cogito builds an open-source pipeline yielding 545k reasoning samples and applies self-distillation; on MMAR it reports top open-source performance and top-tier challenge results.
Problem
Audio reasoning capabilities remain limited and inconsistent, and common audio datasets provide brief labels or captions rather than deep reasoning traces.
Source: §1 · p. 1Method
Audio-Cogito uses an open-source four-stage data pipeline to produce 545k audio reasoning samples and applies self-distillation for fine-tuning.
Source: §1 · p. 1Results
Takeaways & Limitations
Audio-Cogito's MMAR results include the best Rubrics and CRS scores among LARMs, indicating reliable reasoning chains when answers are correct.
Source: §3.2 · p. 4
Abstract
from arXiv · showhide
Recent advances in reasoning models have driven significant progress in text and multimodal domains, yet audio reasoning remains relatively limited. Only a few Large Audio Language Models (LALMs) incorporate explicit Chain-of-Thought (CoT) reasoning, and their capabilities are often inconsistent and insufficient for complex tasks. To bridge this gap, we introduce Audio-Cogito, a fully open-source solution for deep audio reasoning. We develop Cogito-pipe for high-quality audio reasoning data curation, producing 545k reasoning samples. Based on this dataset, we adopt a self-distillation strategy for model fine-tuning. Experiments on the MMAR benchmark, the only audio benchmark evaluating the CoT process, show that our model achieves the best performance among open-source models and matches or surpasses certain closed-source models in specific metrics. Our approach also ranks among the top-tier systems in the Interspeech 2026 Audio Reasoning Challenge.
1. Introduction
Audio reasoning models remain limited and unstable, particularly in complex acoustic settings, while available datasets often lack the depth needed for complex reasoning. Audio-Cogito combines an open-source data pipeline and self-distillation, with reported strong MMAR and challenge results.
1. Introduction
- Current audio reasoning models can produce rigid traces lacking deep audio grounding and misinterpret subtle cues in complex acoustic environments.The paper attributes these limitations primarily to scarce high-quality reasoning data; public datasets often offer brief labels or captions, while existing reasoning datasets largely target shallow tasks.Source: §1 · p. 1
- Audio-Cogito introduces a fully open-source four-stage pipeline and self-distillation approach, releasing 545k high-quality samples across multiple audio domains.The pipeline comprises Data Collection, QA Construction, CoT Generation, and Quality Verification.Source: §1 · p. 1
- Audio-Cogito sets new state-of-the-art results among open-source models on MMAR and achieves top-tier performance in the Interspeech 2026 Audio Reasoning Challenge.The introduction reports that it surpasses several proprietary systems on MMAR.Source: §1 · p. 2
2. Audio-Cogito
Audio-Cogito constructs audio reasoning data through a four-stage pipeline and trains a model to generate explicit reasoning traces before final answers. Its pipeline uses diverse audio sources, generated question-answer pairs, self-distilled reasoning, and quality checks.
2.1. Cogito-Pipe
- Cogito-Pipe gathers multi-domain audio, constructs QA pairs, generates step-by-step reasoning traces, and verifies consistency while filtering hallucinated or low-quality samples.The four stages are Data Collection, QA Construction, CoT Construction, and Quality Verification.Source: §2.1 · p. 2
2.1.1. Data Collection
- The data collection stage covers sound events, speech, music, and mixed or interleaved audio, with metadata and about 500 expert-refined seed questions.The seed questions are generated by an LLM and reviewed, revised, and supplemented by domain experts.Source: §2.1.1 · p. 2
2.1.2. QA Construction
- QA Construction uses Qwen3-Omni-Instruct with 20 seed-question exemplars per instance and generates 1–3 QA pairs per audio clip.The prompts request confusing distractor options as hard negatives.Source: §2.1.2 · p. 2
2.1.3. CoT Generation
- CoT Generation uses Qwen3-Omni-Thinking for both reasoning-data generation and fine-tuning, with freeform traces and ground-truth answers withheld during generation.The authors report that rigid-template format misalignment degrades intrinsic reasoning capabilities; withholding answers is intended to keep reasoning grounded in acoustic cues.Source: §2.1.3 · p. 2
2.1.4. Quality Verification
- Quality Verification checks that CoT-derived answers match constructed QA answers, then uses an LLM judge to filter hallucinations and logical inconsistencies.The pipeline's models are interchangeable, allowing the target training model to generate its own self-distillation data.Source: §2.1.4 · p. 2
2.2. Model Training
- Model Training maximizes the joint likelihood of the reasoning trace and final response given audio and query, training the model to reason before answering.The sample representation contains audio input, query, structured reasoning trace, and final answer.Source: §2.2 · p. 3
3. Experiments
Experiments evaluate Audio-Cogito on MMAR for answer accuracy and reasoning quality, comparing it with audio-capable models and testing Cogito-Pipe components. Audio-Cogito leads the compared open-source models, while ablations report benefits from the pipeline components.
3.1.1. Training Details
- The model is built on 30-billion-parameter Qwen3-Omni-Thinking and fine-tuned for one epoch using LoRA on Cogito-Pipe data.The maximum learning rate is 1 × 10−5.Source: §3.1.1 · p. 3
3.1.2. Evaluation Metrics
- MMAR evaluates answer correctness and reasoning quality; incorrect answers receive a zero reasoning score, and CRS averages reasoning quality conditional on correct answers.The evaluation uses five rubric criteria per sample, GPT-4o as judge, and reports the mean of the middle three scores across five runs.Source: §3.1.2 · p. 3
3.1.3. Baseline Models
- Comparisons cover LALMs, OLMs, and LARMs, including representative open-source and proprietary models.LARMs incorporate explicit Chain-of-Thought reasoning mechanisms.Source: §3.1.3 · p. 3
3.2. Main Results
- Audio-Cogito achieves the best average accuracy among compared open-source models, surpassing Qwen3-Omni-Thinking by 5.44% in relative terms.Gains are especially notable on mixed-domain tasks.Source: §3.2 · pp. 3–4, Table 2 · p. 4
- Audio-Cogito surpasses the average accuracy of several closed-source systems and outperforms Gemini 2.5 Pro on Sound-Music-Speech, with comparable performance on Single Domain Sound.It also achieves the best Rubrics and CRS among LARMs.Source: §3.2 · p. 4
3.3. Ablation Study
- Removing any Cogito-Pipe component degrades performance; seed-question removal causes the largest drop, while removing quality verification increases hallucinations.Excluding metadata reduces QA accuracy by removing grounding cues.Source: §3.3 · p. 4, Table 3 · p. 4
4. Conclusion
Audio-Cogito combines open-source audio reasoning data curation with self-distillation and reports strong results on MMAR and the Interspeech 2026 Audio Reasoning Challenge. The authors also report high Rubrics and CRS scores as evidence of reliable Chain-of-Thought processes.
4. Conclusion
- Audio-Cogito uses Cogito-Pipe to curate and release a 545k-sample open-source audio reasoning dataset, then applies self-distillation to enhance complex reasoning.The authors describe the approach as an open-source solution for deep audio reasoning in LALMs.Source: §4 · p. 4
- On MMAR, Audio-Cogito achieves state-of-the-art performance among open-source models and narrows the gap with leading proprietary systems.Source: §4 · p. 4
- Audio-Cogito also performs at a top tier in the Interspeech 2026 Audio Reasoning Challenge, while high Rubrics and CRS scores indicate reliable Chain-of-Thought processes.Source: §4 · p. 4
5. Generative AI Use Disclosure
Generative AI tools were used only for linguistic refinement and editorial assistance, not for developing or evaluating the work. The authors retain responsibility for the manuscript.
5. Generative AI Use Disclosure
- Generative AI tools were used exclusively for linguistic refinement and editorial assistance.Source: §5 · p. 5
- The tools were not used to develop the methodology, conduct experiments, generate results, or draw conclusions.Source: §5 · p. 5
- The authors retain full responsibility and accountability for all aspects of the manuscript.Source: §5 · p. 5