Source-linked AI summary

Reinforcement Learning Outperforms Supervised Fine-Tuning: A Case Study on Audio Question Answering

Gang Li, Jizhong Liu, Heinrich Dinkel, Yadong Niu, Junbo Zhang, Jian Luan

arXiv:2503.11197v4cs.SDcs.AIcs.CLeess.AS

TL;DR

Audio reinforcement learning has been underexplored despite progress in language and visual reasoning, motivating a study of AQA with LALMs. The paper applies GRPO to Qwen2-Audio-7B-Instruct and achieves state-of-the-art MMAU Test-mini accuracy of 64.5%.

  • Problem

    Reinforcement learning has been largely overlooked for audio understanding, while more complex reasoning tasks are needed to explore LALMs beyond descriptive audio tasks.

  • Method

    The study applies GRPO to Qwen2-Audio-7B-Instruct using customized instruction prompts and answer-based reinforcement-learning optimization.

  • Results

    64.5% accuracy on MMAU Test-mini represents state-of-the-art performance, with reinforcement learning outperforming SFT using 38k post-training samples.

  • Takeaways & Limitations

    GRPO can be effectively applied to LALMs and AQA without large datasets, including a model with 8.2 billion parameters.

  • Takeaways & Limitations

    Explicit reasoning has not shown significant benefits for AQA, and efficiently leveraging deep thinking remains an open question; LALMs still lag far behind humans.

Abstract

from arXiv · show

Recently, reinforcement learning (RL) has been shown to greatly enhance the reasoning capabilities of large language models (LLMs), and RL-based approaches have been progressively applied to visual multimodal tasks. However, the audio modality has largely been overlooked in these developments. Thus, we conduct a series of RL explorations in audio understanding and reasoning, specifically focusing on the audio question answering (AQA) task. We leverage the group relative policy optimization (GRPO) algorithm to Qwen2-Audio-7B-Instruct, and our experiments demonstrated state-of-the-art performance on the MMAU Test-mini benchmark, achieving an accuracy rate of 64.5%. The main findings in this technical report are as follows: 1) The GRPO algorithm can be effectively applied to large audio language models (LALMs), even when the model has only 8.2B parameters; 2) With only 38k post-training samples, RL significantly outperforms supervised fine-tuning (SFT), indicating that RL-based approaches can be effective without large datasets; 3) The explicit reasoning process has not shown significant benefits for AQA tasks, and how to efficiently utilize deep thinking remains an open question for further research; 4) LALMs still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further exploration. Our project is available at https://github.com/xiaomi-research/r1-aqa and https://huggingface.co/mispeech/r1-aqa.

1 Introduction

Audio question answering extends audio understanding with linguistic reasoning, offering a complex setting for testing reinforcement learning. This report explores GRPO on LALMs and reports state-of-the-art AQA performance with limited-scale training.

  • Audio modality research has largely overlooked reinforcement-learning approaches despite progress in LLM reasoning and visual multimodal tasks.
  • AQA requires extracting insights from audio, inferring implicit relationships, and generating contextually relevant answers.Its combination of auditory and linguistic modalities makes it suitable for evaluating complex logical reasoning.
  • The report applies GRPO to Qwen2-Audio-7B-Instruct to explore reinforcement learning and deep thinking for audio question answering.
  • 64.5% accuracy on MMAU Test-mini establishes state-of-the-art performance for the reported AQA experiments.
  • With 38k post-training samples, reinforcement learning outperforms supervised fine-tuning, while explicit reasoning shows no significant AQA benefit.The report also states that LALMs remain far behind human auditory-language reasoning.

2 Related Works

Related work positions AQA as a reasoning-intensive audio task and reviews emerging multimodal reasoning methods alongside predominantly SFT-based audio language models.

  • Audio Question Answering: AQA builds on audio captioning but additionally requires comprehension of complex acoustic patterns, temporal relationships, and contextual information.
  • Audio Question Answering: The study focuses on single-audio, selection-based questions, where generating an answer is difficult but verifying options is straightforward.
  • Multimodal Reasoning: Recent multimodal studies have renewed interest in reinforcement learning and chain-of-thought reasoning for multi-step tasks.
  • Large Audio Language Models: Representative audio understanding models, including Qwen2-Audio, Audio Flamingo 2, and SALMONN, are described as SFT-based.Whether reinforcement learning can unlock their potential remains studied in this report.

3 Method

The method directly applies GRPO to Qwen2-Audio-7B-Instruct with customized prompts, comparing explicit and implicit reasoning formats. GRPO optimizes sampled responses using group-relative rewards for answer correctness and output format.

  • The training method applies GRPO directly to the Qwen2-Audio-7B-Instruct large audio language model.
  • For each question, the model generates an answer-formatted response that is optimized with a reinforcement-learning objective.
  • Prompts compare direct answer generation with an explicit chain-of-thought format requiring reasoning inside <think> tags.
  • GRPO samples a group of responses and uses their average reward as the advantage baseline, avoiding an additional value-function approximation model.
  • The policy objective includes a Kullback-Leibler term, with current and former policies controlled by PPO hyperparameters.
  • Responses receive rule-based rewards for correctness and, when applicable, enclosing reasoning and answers in the required tags.

4 Experiments

Experiments compare full fine-tuning, LoRA, SFT, and GRPO-based RL using approximately 38k AVQA audio-text samples and out-of-distribution MMAU Test-mini evaluation. GRPO achieves state-of-the-art average accuracy, while RL-based methods outperform SFT and explicit chain-of-thought does not consistently help AQA.

  • Experimental Setup: Approximately 38k AVQA audio-text training samples were evaluated on the MMAU Test-mini out-of-distribution test set.The training and test data came from different sources.
  • Main Results: GRPO with Prompt <2> achieved state-of-the-art average accuracy on the MMAU Test-mini benchmark.The comparison included direct inference, SFT, RL methods, leaderboard baselines, and recent RL studies.
  • Main Results: Deep-thinking methods overall outperformed classic SFT, with the top four methods using RL or chain-of-thought approaches.The reported top four were GRPO + Prompt <2>, GRPO + Prompt <3>, Audio-Reasoner, and Audio-CoT.
  • Main Results: Using only 38k AVQA samples, RL achieved state-of-the-art performance on out-of-distribution tests, supporting strong generalization with limited training data.The authors relate this result to the generation-verification structure of AQA and small-sample structured-thinking findings.
  • Main Results: Explicit CoT templates and <think> prompting did not outperform directly prompting the model to generate <answer> responses for AQA.The contribution of deep thinking and step-by-step reasoning remains unresolved for this task.
  • Main Results: The accuracy of LoRA + Prompt <1> was 8.1% lower than that of GRPO + Prompt <2>.LoRA partially addresses small-data fine-tuning challenges, but SFT still performed worse than RL.

5 Conclusion

Direct application of GRPO to Qwen2-Audio-7B-Instruct achieves state-of-the-art MMAU Test-mini performance, supporting RL for audio-language models and AQA. However, explicit reasoning has not significantly benefited AQA, and these models still lag behind human auditory-language reasoning.

  • 64.5% accuracy on MMAU Test-mini establishes state-of-the-art performance for the reported AQA system.The system directly applies GRPO to Qwen2-Audio-7B-Instruct.
  • Reinforcement learning can be effectively applied to LALMs and audio multimodal tasks, particularly AQA.The report identifies AQA's generation-verification gap as especially relevant to this application.
  • Explicit reasoning has not shown significant benefits for AQA, leaving efficient use of deep or step-by-step reasoning unresolved.The report identifies integrating chain-of-thought into audio as a future research focus.
  • LALMs still lag far behind human auditory-language reasoning, motivating further exploration of reinforcement learning.
Loading 2503.11197v4…