Research map / Large language models / Agents, code and math
AuTAgent: A Reinforcement Learning Framework for Tool-Augmented Audio Reasoning
Siqian Tong, Li Xuan, Yiwei Wang, Baolong Bi, Yujun Cai, Shenghua Liu, Yuchen He, HAO Chengpeng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| concatenate outputs of all six tools (ablation) | Qwen2-Audio-7B-Instruct · MMAU Test-mini | accuracy (%): 50.00 -> 40.90 | Hurt |
| use baseline-subtracted differential reward instead of binary reward (ablation) | AuTAgent · MMAU Test-mini | accuracy (%): 51.40 -> 54.20 | Improved |
| use baseline-subtracted differential reward instead of binary reward (ablation) | AuTAgent · MMAR | accuracy (%): 39.00 -> 41.30 | Improved |
| use differential reward instead of binary reward (ablation) | AuTAgent · MMAU Test-mini | avg tools invoked: 1.62 -> 1.30 | Improved |
| train tool-selection policy with GRPO instead of description-based zero-shot selection | Qwen2-Audio-7B-Instruct · MMAU Test-mini | accuracy (%): 51.20 -> 54.20 | Improved |
| random tool selection (ablation) | Qwen2-Audio-7B-Instruct · MMAU Test-mini | accuracy (%): 50.00 -> 49.10 | Hurt |
| random tool selection (ablation) | GPT-4o Audio · MMAU Test-mini | accuracy (%): 57.40 -> 53.60 | Hurt |
| add trained AuTAgent policy as plugin to GPT-4o Audio | GPT-4o Audio · MMAU Test-mini | accuracy (%): 57.40 -> 66.80 | Improved |
| plugin policy (trained on other backbone) vs self-trained policy (ablation) | AuTAgent · MMAU Test-mini / MMAR | accuracy (%): 67.20 -> 66.80 (MMAU, GPT-4o) | No clear effect |
| add tool library to GPT-4o Audio via AuTAgent | GPT-4o Audio · MMAR | accuracy (%): 55.00 -> 63.00 | Improved |
| invoke each single tool T2, T3, T5 individually (ablation) | Qwen2-Audio-7B-Instruct · MMAU Test-mini | accuracy (%) | Hurt |
Closest papers by meaning
- Incentivizing Consistent, Effective and Scalable Reasoning Capability in Audio LLMs via Reasoning Process Rewards (ICLR 2026 · similarity 0.92)
- Audio-Thinker: Guiding Large Audio Language Model When and How to Think via Reinforcement Learning (AAAI 2026 · similarity 0.90)
- Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities (ICML 2025 · similarity 0.89)
- AudSemThinker: Enhancing Audio-Language Models Through Reasoning over Semantics of Sound (NeurIPS 2025 · similarity 0.88)
- MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix (NeurIPS 2025 · similarity 0.88)
- MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
