README.md

---
license: cc-by-nc-4.0
---

# Audio-Cogito: Towards Deep Audio Reasoning in Large Audio Language Models

<p align="center">
  <a href="https://arxiv.org/abs/2604.12527">
    <img src="https://img.shields.io/badge/arXiv-2604.12527-b31b1b.svg" alt="arXiv">
  </a>
  <a href="https://github.com/llh666521/Audio-Cogito">
    <img src="https://img.shields.io/badge/GitHub-Audio--Cogito-black.svg" alt="GitHub">
  </a>
</p>

**Audio-Cogito** is a large-scale audio reasoning dataset introduced in the paper [Audio-Cogito: Towards Deep Audio Reasoning in Large Audio Language Models](https://arxiv.org/abs/2604.12527).

The released data contains **545k high-quality audio reasoning samples** spanning sound, speech, and music domains. Each sample includes label annotations, Chain-of-Thought (CoT) annotations, and final answers.

## Links

- Paper: [arXiv:2604.12527](https://arxiv.org/abs/2604.12527)
- GitHub: [llh666521/Audio-Cogito](https://github.com/llh666521/Audio-Cogito)
- Data file: [audio-cogito-data.jsonl](https://huggingface.co/datasets/lilonghao/Audio-Cogito/blob/main/audio-cogito-data.jsonl)

## Dataset Description

Audio-Cogito is designed to elicit and study deep audio reasoning capabilities in Large Audio Language Models (LALMs). The dataset is constructed with **Cogito-Pipe**, a four-stage pipeline for audio reasoning data construction:

- **Data Collection:** Gathering data from multi-domain audio sources spanning sound, speech, and music.
- **QA Construction:** Synthesizing diverse and challenging QA pairs based on the collected audio.
- **CoT Construction:** Producing detailed Chain-of-Thought reasoning annotations for each task.
- **Quality Verification:** Enforcing consistency between QA pairs and CoT rationales while filtering hallucinated or low-quality samples.

## Data Format

The dataset is provided as a JSONL file. Each line contains a conversation-style sample and an associated audio path identifier.

```json
{
  "messages": [
    {
      "role": "user",
      "content": "<audio>Question text ..."
    },
    {
      "role": "assistant",
      "content": "<think>CoT annotation ...</think>\n\nFinal answer"
    }
  ],
  "audios": [
    "audiocap/audios/audio_00000002.wav"
  ]
}
```

The `messages` field contains the user query and the annotated assistant response. The assistant response includes both CoT annotations and the final answer. The `audios` field stores the corresponding audio path identifier.

## Dataset Statistics

| Domain | Dataset Source | Main Skills Learning | Quantity | Ratio (%) |
| --- | --- | --- | --- | --- |
| Sound | AudioSet | General Audio Event | 179k | 32.53 |
| Sound | Clotho | Audio Captioning | 6k | 1.14 |
| Sound | AudioCaps | Audio Captioning | 40k | 7.20 |
| Sound | ComplexAudio | Complex Audio | 37k | 6.66 |
| Speech | MELD | Speech Emotion | 24k | 4.50 |
| Speech | CoVoST2 | Speech Translation | 56k | 10.10 |
| Speech | DailyTalk | Spoken Dialogue | 9k | 1.64 |
| Music | MusicBench | General Music | 88k | 16.04 |
| Music | FMA | Music Genre | 76k | 13.81 |
| Music | Medley-solos-DB | Instrument Analysis | 35k | 6.38 |

## Main Results

Audio-Cogito achieves top-tier performance in the Interspeech 2026 Audio Reasoning Challenge and sets new state-of-the-art results among open-source models on the MMAR benchmark.

| Model | Size | Sound | Music | Speech | S-M | S-S | M-S | S-M-S | Avg (%) | Rubrics (%) | CRS |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| Qwen3-Omni-Thinking | 30B | 64.24 | 50.00 | **79.25** | 54.55 | 72.48 | 69.51 | 70.83 | 68.00 | 57.97 | 0.85 |
| **Audio-Cogito** | 30B | **66.67** | **53.40** | **79.25** | **90.91** | **79.90** | **76.83** | **79.17** | **71.70** | **62.22** | **0.87** |

**Notes:** S-M: Sound-Music, S-S: Sound-Speech, M-S: Music-Sound, S-M-S: Sound-Music-Speech.

## Citation

If you find **Audio-Cogito** useful for your research, please cite our paper:

```bibtex
@misc{li2026audiocogitodeepaudioreasoning,
      title={Audio-Cogito: Towards Deep Audio Reasoning in Large Audio Language Models}, 
      author={Longhao Li and Hongjie Chen and Zehan Li and Qihan Hu and Jian Kang and Jie Li and Lei Xie and Yongxiang Li},
      year={2026},
      eprint={2604.12527},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2604.12527}, 
}
```