Research map / Large language models / Language, safety and interpretability
AHAMask: Reliable Task Specification for Large Audio Language Models Without Instructions
Yiwei Guo, Bohan Li, Hankun Wang, Zhihan Li, Shuai Wang, Xie Chen, Kai Yu
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 |
|---|---|---|---|
| replace instruction with AHAMask | SALMONN · gender recognition | accuracy: 96.79 -> 98.05 | Improved |
| replace instruction with AHAMask | SALMONN · AudioCaps AAC | METEOR: 20.60 -> 24.15 | Improved |
| replace instruction with AHAMask | SALMONN · ASV, S2TT, ASR | accuracy/BLEU/WER: 93.49 -> 93.24 (ASV); 34.48 -> 33.90 (S2TT) | No clear effect |
| replace instruction with AHAMask | Qwen2Audio-Base · VoxCeleb1-O ASV | accuracy: 49.24 -> 85.75 | Improved |
| replace instruction with AHAMask | Qwen2Audio-Base · LibriSpeech test-clean ASR | WER: 1.77 -> 3.11 | Hurt |
| replace instruction with AHAMask | Qwen2Audio-Base · CoVoST2 en-zh S2TT | BLEU-4: 45.00 -> 39.27 | Hurt |
| use trained AHAMask instead of random mask with same head count (ablation) | SALMONN · LibriSpeech test-clean ASR | WER: 21.36 -> 2.10 | Improved |
| apply AHAMask trained on the Base model to Instruct model (swap) (ablation) | Qwen2Audio-Instruct · LibriSpeech test-clean ASR | WER: 2.99 -> 156.87 | Hurt |
| replace instruction with AHAMask for ASR|GR ordering | SALMONN · ASR|GR composite | IFR (%): 16.03 -> 97.63 | Improved |
| replace instruction with AHAMask | Qwen2Audio-Instruct · JSON composite | IFR (%): 0.00 -> 58.45 | Improved |
| activate heads gradually by importance | SALMONN · ASR, AAC, S2TT, OSR | task metrics | Improved |
| longer, more complex or alter-case instructions | SALMONN · ASR prompt variation | WER: to 12% | Hurt |
Closest papers by meaning
- Heads up! Large Language Models Can Perform Tasks Without Your Instruction via Selective Attention Head Masking (ICML 2025 · similarity 0.89)
- Listen, Think, and Understand (ICLR 2024 · similarity 0.87)
- MATS: An Audio Language Model under Text-only Supervision (ICML 2025 · similarity 0.87)
- Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models (ICLR 2026 · similarity 0.87)
- Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities (ICML 2025 · similarity 0.86)
- Masked Autoencoders with Multi-Window Local-Global Attention Are Better Audio Learners (ICLR 2024 · similarity 0.86)
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.
