Research map / Large language models / Language, safety and interpretability
Echo: Towards Advanced Audio Comprehension via Audio-Interleaved Reasoning
Daiqing Wu, Xuan Zhang, Dongbao Yang, Jiashu Yao, Longfei Chen, Qingsong Liu, Sicheng Zhao, Can Ma, Yangyang Kang, Yu ZHOU
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 |
|---|---|---|---|
| add SFT on EAQA-SFT to the base model (audio-grounded format) (ablation) | Qwen2.5-Omni-7B · MMAR | accuracy: 51.80 -> 56.77 | Improved |
| switch inference of the SFT cold-start model to audio-interleaved format without RL (ablation) | Cold-start model · MMAR | accuracy: 56.77 -> 52.26 | Hurt |
| add RL after interleaved adaptation (ablation) | Echo · MMAR | accuracy: 52.26 -> 69.99 | Improved |
| use audio-interleaved format instead of audio-grounded text format under the same RL (ablation) | Echo · MMAR | accuracy: 64.63 -> 69.99 | Improved |
| add SFT before RL (audio-grounded) vs direct RL on base format (ablation) | Qwen2.5-Omni-7B · MMAR | accuracy: 63.15 -> 64.63 | Improved |
| audio-interleaved reasoning adds inference cost (ablation) | Echo · MMAR | reasoning latency: 2.06 s -> 2.12 s (direct RL vs Echo) | No clear effect |
| use EAQA-RL (21.9k) instead of AVQA (40.4k) as RL data (ablation) | Echo · MMAR | accuracy: 67.58 -> 69.99 | Improved |
| use EAQA-RL instead of AVQA for direct RL (ablation) | Direct RL model · MMAR | accuracy: 59.51 -> 63.15 | Improved |
| filter SFT data with re-evaluation (75,862 high quality vs 111,868 unfiltered) (ablation) | Echo · MMAR | accuracy: 64.67 -> 69.99 | Improved |
| filter RL data with re-evaluation (ablation) | Echo · MMAR | accuracy: 67.03 -> 69.99 | Improved |
| use stronger audio extractor Qwen2.5-Omni instead of Qwen2-Audio for data generation (ablation) | Echo data pipeline · MMAR | accuracy: 63.91 -> 69.99 | Improved |
| use Qwen2.5-Omni instead of Kimi-Audio as audio extractor (ablation) | Echo data pipeline · MMAR | accuracy: 68.82 -> 69.99 | Improved |
Closest papers by meaning
- Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities (ICML 2025 · similarity 0.92)
- Incentivizing Consistent, Effective and Scalable Reasoning Capability in Audio LLMs via Reasoning Process Rewards (ICLR 2026 · similarity 0.91)
- Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models (NeurIPS 2025 · similarity 0.91)
- MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix (NeurIPS 2025 · similarity 0.90)
- Mellow: a small audio language model for reasoning (NeurIPS 2025 · similarity 0.90)
- Listen, Think, and Understand (ICLR 2024 · similarity 0.90)
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.
