Update README.md (#7)

- Update README.md (2e5d1a49e0b26482a6313139524b43b3c3ffcf6a)


Co-authored-by: Gang Li <[email protected]>

README.md

@@ -11,7 +11,7 @@ pipeline_tag: audio-text-to-text
 
 ## Introduction
 
-R1-AQA is a audio question answering (AQA) model based on `Qwen2-Audio-7B-Instruct`, optimized through reinforcement learning using the group relative policy optimization (GRPO) algorithm. 
+R1-AQA is a audio question answering (AQA) model based on `Qwen2-Audio-7B-Instruct`, optimized through reinforcement learning using the group relative policy optimization (GRPO) algorithm.
 This implementation has achieved state-of-the-art performance on MMAU *Test-mini* benchmark with only 38k post-training samples.
 For more details, please refer to our [Github](https://github.com/xiaomi-research/r1-aqa) and [Technical Report](https://arxiv.org/abs/2503.11197).
 
@@ -23,10 +23,12 @@ Our main findings are as follows:
 - Large audio language models (LALMs) still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further explorations.
 
 Additional Notes:  
-The AVQA training set originally consists of approximately 40k samples. However, we use only about 38k samples because some data sources have become invalid. Other datasets using YouTube sources face a similar issue, such as AudioSet. We believe that the missing 2k samples do not have a significant impact on the training results.
 
+- The AVQA training set originally consists of approximately 40k samples. However, we use only about 38k samples because some data sources have become invalid. Other datasets using YouTube sources face a similar issue, such as AudioSet. We believe that the missing 2k samples do not have a significant impact on the training results.
+- The statement about the 8.2B parameters is based on the *Qwen2-Audio Technical Report*.
 
 ### Table: Accuracies (%) on MMAU Test-mini benchmark
+
 | Model                                      | Method                  | Sound  | Music  | Speech | Average |
 |--------------------------------------------|-------------------------|--------|--------|--------|---------|
 | \                                          | Human\*                 | 86.31  | 78.22  | 82.17  | 82.23   |
@@ -39,18 +41,18 @@ The AVQA training set originally consists of approximately 40k samples. However,
 | GPT4o + Weak Cap.                          | Direct Inference\*      | 39.33  | 41.90  | 58.25  | 45.70   |
 | Llama-3-8B-Instruct + Weak Cap.            | Direct Inference\*      | 34.23  | 38.02  | 54.05  | 42.10   |
 | SALMONN                                    | Direct Inference\*      | 41.00  | 34.80  | 25.50  | 33.70   |
-| Qwen2-Audio-7B-Instruct                    | CoTA \[1\]            | 60.06  | 64.30  | 60.70  | 61.71   |
-| Qwen2-Audio-7B-Instruct                    | Zero-Shot-CoT \[2\]   | 61.86  | 56.29  | 55.26  | 57.80   |
+| Qwen2-Audio-7B-Instruct                    | CoTA \[1\]              | 60.06  | 64.30  | 60.70  | 61.71   |
+| Qwen2-Audio-7B-Instruct                    | Zero-Shot-CoT \[2\]     | 61.86  | 56.29  | 55.26  | 57.80   |
 | **Qwen2-Audio-7B-Instruct**                | **GRPO (Ours)**         | **69.37** | 66.77  | 57.36  | **64.50** |
 
-#### Notes:
+#### Notes
+
 \* The data are sourced from the MMAU official website: [https://sakshi113.github.io/mmau_homepage/](https://sakshi113.github.io/mmau_homepage/)  
 \[1\] Xie, Zhifei, et al. "Audio-Reasoner: Improving Reasoning Capability in Large Audio Language Models." arXiv preprint arXiv:2503.02318 (2025).  
 \[2\] Ma, Ziyang, et al. "Audio-CoT: Exploring Chain-of-Thought Reasoning in Large Audio Language Model." arXiv preprint arXiv:2501.07246 (2025).  
 
-
-
 ## Inference
+
 ```python
 import torch
 import torchaudio
@@ -87,4 +89,16 @@ generated_ids = generated_ids[:, inputs.input_ids.size(1):]
 response = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
 
 print(response)
-```
+```
+
+## Citation
+
+```bib
+@article{li2025reinforcement,
+  title={Reinforcement Learning Outperforms Supervised Fine-Tuning: A Case Study on Audio Question Answering},
+  author={Li, Gang and Liu, Jizhong and Dinkel, Heinrich and Niu, Yadong and Zhang, Junbo and Luan, Jian},
+  journal={arXiv preprint arXiv:2503.11197},
+  year={2025},
+  url={https://github.com/xiaomi-research/r1-aqa; https://huggingface.co/mispeech/r1-aqa}
+}
+```